https://tokencost.app/blog LLM API pricing news, model launch cost breakdowns, and practical guides for developers. en Mon, 26 Oct 2026 00:00:00 GMT https://tokencost.app/blog/openai-12-hour-prompt-cache-pricing https://tokencost.app/blog/openai-12-hour-prompt-cache-pricing Mon, 26 Oct 2026 00:00:00 GMT Guide OpenAI added a 12-hour cache lifetime on October 20, with writes billed at 2.5x input. We work out where it beats the 30-minute default and Claude's 1-hour cache, why a user who comes back every 90 minutes now costs less than half as much on GPT-6.1 Sol as on Sonnet 5.5, and how to set it per breakpoint so it doesn't tax your hot loop. https://tokencost.app/blog/llm-api-price-changes-october-2026 https://tokencost.app/blog/llm-api-price-changes-october-2026 Sun, 25 Oct 2026 00:00:00 GMT Industry We list every LLM API price change in October 2026 and price one 500M-in, 100M-out month before and after each, then lay out the dated changes through January 14. Most cuts need you to switch models. Most increases arrive on their own, and OpenAI's December 11 shutdowns are the biggest of them. https://tokencost.app/blog/kimi-k3-1-pro-pricing https://tokencost.app/blog/kimi-k3-1-pro-pricing Sat, 24 Oct 2026 00:00:00 GMT Model Release Moonshot cut a third off K3's price for K3.1 Pro. We work out where its cheaper cache writes and dearer cache hits cross over against Sonnet 5.5 and Sol, why Artificial Analysis puts it at $2.11 a task to Sol's $0.72 on the same card, and whether Highspeed at Opus prices is worth it. https://tokencost.app/blog/recursive-flint-local-vs-api-cost https://tokencost.app/blog/recursive-flint-local-vs-api-cost Fri, 23 Oct 2026 00:00:00 GMT Model Release Recursive's 40B-A2.5B open coder shipped October 21 with a free OpenRouter tier and no paid host. We price a developer's agent month on Haiku 5.5, DeepSeek, MiniMax and Sonnet 5.5, work out what $1,000 of local hardware buys instead, and read the vendor chart line by line. https://tokencost.app/blog/minimax-m3-1-flash-pricing https://tokencost.app/blog/minimax-m3-1-flash-pricing Thu, 22 Oct 2026 00:00:00 GMT Model Release MiniMax's new Flash model matches DeepSeek V4.1 Flash on input and output but charges 5x on cache hits and never doubles at peak. We price a cached agent month across the $0.15 tier, find where the two cross over, and reprice Artificial Analysis' cost per task at DeepSeek's off-peak rate. https://tokencost.app/blog/grok-voice-3-pricing-per-minute https://tokencost.app/blog/grok-voice-3-pricing-per-minute Wed, 21 Oct 2026 00:00:00 GMT Model Release xAI repointed grok-voice-latest on October 14 and, unlike in August, the price didn't move. We price 40,000 minutes a month on six voice models, divide by the Speech to Speech Index, and list the tool, keypress and concurrency lines the $0.08 leaves out. https://tokencost.app/blog/gemini-4-argon-ga-price-november-30 https://tokencost.app/blog/gemini-4-argon-ga-price-november-30 Tue, 20 Oct 2026 00:00:00 GMT Industry Google made Gemini 4 Argon generally available on October 13 and printed the date its price doubles. We price a 400K-token document job and a short agent call on Argon, GPT-6.1 Sol, Opus 5.5 and Gemini 3.1 Pro, before and after December 1, and look at the caching and batch lines the launch post left out. https://tokencost.app/blog/x-subscription-grok-cursor-pricing https://tokencost.app/blog/x-subscription-grok-cursor-pricing Mon, 19 Oct 2026 00:00:00 GMT Industry xAI's October 12 bundle sells Grok, Cursor and X from $8 to $200 with a weekly pool that covers only Grok and Composer. We price a Cursor agent session on five models and four kinds of user: Grok-first coders and SuperGrok subscribers save $20 to $40, Claude users on Pro pay $30 more at any volume. https://tokencost.app/blog/deepseek-search-fetch-api-pricing https://tokencost.app/blog/deepseek-search-fetch-api-pricing Sun, 18 Oct 2026 00:00:00 GMT Comparison DeepSeek's October 17 release opened standalone Search and Fetch APIs at $0.75 per 1,000 calls and kept V4.1 Flash's token price on a bigger 1017 build. We price a 10,000-question research agent on DeepSeek, Haiku 5.5, GPT-6 Luna and Gemini 3.8 Flash, and on Haiku with DeepSeek's tools: $72.50 against $335. https://tokencost.app/blog/muse-spark-open-pricing https://tokencost.app/blog/muse-spark-open-pricing Sat, 17 Oct 2026 00:00:00 GMT Model Release Meta released Muse Spark Open's 816B weights on October 15 and kept it off its own API. Together, Fireworks, Baseten, Modal and Nebius all sell it at $1.25 in and $4.25 out, the closed 1.3's rate to the cent. We price it against GLM-5.3, Kimi K3 and Mistral Large 4, and work out when a 4x B300 box beats renting the tokens. https://tokencost.app/blog/gpt-4o-gpt-4-1-retirement-migration-cost https://tokencost.app/blog/gpt-4o-gpt-4-1-retirement-migration-cost Fri, 16 Oct 2026 00:00:00 GMT Industry OpenAI's October 14 notice shuts down five GPT-4o and GPT-4.1 snapshots on January 14, 2027 and points them all at GPT-6 Luna. A $700 GPT-4o month drops to $30, but 4o mini's saving vanishes once Luna's default reasoning passes 1.2x your output, and GPT-4.1 nano, gone October 23, has no cheaper OpenAI successor that is not also retiring. https://tokencost.app/blog/gpt-6-1-sol-ultrafast-pricing https://tokencost.app/blog/gpt-6-1-sol-ultrafast-pricing Thu, 15 Oct 2026 00:00:00 GMT Comparison OpenAI's Ultrafast tier for GPT-6.1 Sol went live on October 8 at exactly 6x Standard on every line, cache reads and the long-context step included. We price a 12-turn coding loop on Ultrafast, Fast, Astra, Opus 5.5 fast mode and Argon Priority, and work out what an hour of your waiting has to be worth before it pays. https://tokencost.app/blog/openai-decisions-api-vs-microsoft-decision-1-pricing https://tokencost.app/blog/openai-decisions-api-vs-microsoft-decision-1-pricing Wed, 14 Oct 2026 00:00:00 GMT Comparison GPT-6 Luna Decisions went into beta on October 6 at $0.10 per million input tokens with no output or cache charges. Microsoft-Decision-1 followed on October 9 at $0.042. We price one support ticket across nine cards, show why a cached 2,000-token rubric costs $61.50 per million calls on the Responses API and $230 on Decisions, and check the first independent accuracy tests. https://tokencost.app/blog/meta-muse-wonder-pricing https://tokencost.app/blog/meta-muse-wonder-pricing Tue, 13 Oct 2026 00:00:00 GMT Model Release Muse Wonder bills $4 input, $0.20 cached and $20 output on the Meta Model API, the same three numbers as Opus 5.5, with no batch tier and no cache-write fee. We price an 80M-in, 8M-out agent month against Opus, Sonnet 5.5, GPT-6.1 Sol, Grok 4.7 and Muse Spark 1.3, line up Artificial Analysis cost per task, and work out what the $50 Muse Code plan buys you that the API doesn't. https://tokencost.app/blog/nano-banana-2-1-pricing https://tokencost.app/blog/nano-banana-2-1-pricing Mon, 12 Oct 2026 00:00:00 GMT Model Release Google's new image model bills $0.0336 at 1K, $0.0504 at 2K and $0.113 at 4K, against Nano Banana 2's $0.067, $0.101 and $0.151. Input went from $0.50 to $1.50 and thinking from $3 to $7.50. We work out how much prompt and thinking it takes to erase the saving, why the $0.0756 4K price going around is wrong, and where 2.1 sits next to Lite, Pro and GPT Image 2.5. https://tokencost.app/blog/gemini-2-5-shutdown-october-20-migration-cost https://tokencost.app/blog/gemini-2-5-shutdown-october-20-migration-cost Sun, 11 Oct 2026 00:00:00 GMT Industry Google's two consoles now disagree about the same three models. Vertex retires 2.5 Pro, Flash and Flash-Lite on October 20; the Gemini API calls them not deprecated but locks out new projects. We price a 100M-in, 10M-out month on every named replacement, show why thinking defaults make those numbers a floor, and work out what losing 1,500 free grounded searches a day costs. https://tokencost.app/blog/mistral-large-4-pricing https://tokencost.app/blog/mistral-large-4-pricing Sat, 10 Oct 2026 00:00:00 GMT Model Release Mistral's docs show Large 4 at $0.68 input and $2.09 output with the $1.36/$4.18 list price struck through and no end date. We price a 50M-in, 10M-out month against V4-Pro, Luna, GLM-5.3 and Large 3, work out cost per Artificial Analysis task (it used 200M output tokens against an 82M median), and explain why the EU endpoint gets half the 1M context. https://tokencost.app/blog/claude-haiku-5-5-pricing https://tokencost.app/blog/claude-haiku-5-5-pricing Fri, 09 Oct 2026 00:00:00 GMT Model Release Anthropic cut Haiku 90% on October 7 and landed exactly on Luna's rate card, cache read included. Below 100K tokens the two bill identically. Above it Haiku jumps to $0.50/$2.50 on every token, so a 200K prompt costs $0.105 against Luna's $0.021. Also inside: Anthropic's benchmark table, the 3x output-token habit, why 77K Haiku 4.5 tokens are now 100K, and the Haiku 4.5 retirement date. https://tokencost.app/blog/decision-models-pricing-jev-clef-d1 https://tokencost.app/blog/decision-models-pricing-jev-clef-d1 Mon, 05 Oct 2026 00:00:00 GMT Comparison Jev, Clef, d1, Perplexity's Decider and seven more return probabilities instead of text and charge input only, from $0.04 to $0.24 per million. On the same 356-token ticket that is $0.014 per 1,000 on d1, $0.047 on GPT-6 Luna and $0.085 on Clef. Also inside: latency from 39 ms to 1.5 s, the only independent accuracy board, Clef's 2K truncation and OpenAI's unpriced Decisions API. https://tokencost.app/blog/grok-4-7-vs-mimo-v2-6-pro-pricing https://tokencost.app/blog/grok-4-7-vs-mimo-v2-6-pro-pricing Sun, 04 Oct 2026 00:00:00 GMT Comparison Grok 4.7 kept Grok 4.6's $2/$6 card. Xiaomi's open-weight MiMo-V2.6-Pro lists at $0.435/$0.87 with cache hits at $0.0036. Both score 46 on the Artificial Analysis Intelligence Index. At xAI's default effort, Grok cost $3,881 to run it and MiMo $207. Also inside: why Grok 4.7 costs 65% more per run than Grok 4.6 at the same price, the 200K line, the 257 tokens-a-second MiMo host, and the Sonnet 5.5 row that sits above both. https://tokencost.app/blog/qwen3-8-max-prime-glm-5-3-prime-pricing https://tokencost.app/blog/qwen3-8-max-prime-glm-5-3-prime-pricing Sat, 03 Oct 2026 00:00:00 GMT Comparison Alibaba's new Prime mode sells Qwen3.8 Max at $4/$12 and GLM-5.3 at $2.80/$8.80, twice the standard cards, for a promised 1.5x to 2x throughput. OpenRouter's endpoint stats show Qwen Prime at 58 to 70 tokens a second against 36, with first token under a second. GLM-5.3 Prime measured 64, below Alibaba's own standard GLM endpoint at 75 and far below Decart, Mistral and Baseten. Also inside: how OpenAI's Fast and new 6x Ultrafast tiers measure up, and why GLM-5.3 Prime isn't a Z.ai product. https://tokencost.app/blog/gpt-6-luna-pricing https://tokencost.app/blog/gpt-6-luna-pricing Fri, 02 Oct 2026 00:00:00 GMT Comparison GPT-6 Luna lists at $0.10 in, $0.01 cached and $0.50 out, half of GPT-5.6 Luna and 1/20th of GPT-6.1 Sol. Per task the picture changes: Luna at max scores 37 for $0.07, Sol at low scores 42 for $0.13. Luna's real case is its default medium setting at two cents a task. Also inside: Xiaomi's MiMo-V2.6-Flash beating Luna on both indexes for less, the 272K line that doubles a request, Bedrock's 10% regional uplift and the September 25 image fix. https://tokencost.app/blog/gemini-4-argon-pricing https://tokencost.app/blog/gemini-4-argon-pricing Thu, 01 Oct 2026 00:00:00 GMT Model Release Google priced Gemini 4 Argon at $2 in, $0.10 cached and $10 out for an introductory period with no end date, then $4, $0.20 and $20. The first card is GPT-6.1 Sol's to the cent, the second is Opus 5.5's. On the same prices Argon spends far more tokens: Artificial Analysis has it at 53 for $1.99 a task, 2.76x Sol's max run for one point, and Opus 5.5 on high scores 54 for $1.82. Vals points the other way, with Argon first of 41 at about half Opus 5.5's cost per test. Also inside: who can actually call it today, the 1M output claim, and what the model doesn't have yet, starting with a model ID. https://tokencost.app/blog/gpt-6-1-sol-pricing https://tokencost.app/blog/gpt-6-1-sol-pricing Wed, 30 Sep 2026 00:00:00 GMT Model Release OpenAI shipped GPT-6.1 Sol at DevDay on September 29 at $2 in and $10 out, with cached input halved to $0.10 on every tier. On agent loops at 90% to 95% cache hits that is 16% to 24% off the bill. Artificial Analysis puts it at 52 for $0.72 a task on max, against GPT-6 Sol's 48 for $1.05, and at 48 for $0.21 on medium. Vals has it 57% cheaper per test. Sonnet 5.5 shares every other line of the card, scores higher, costs 10.5x as much per task on AA at max, and stays cheaper past 272K tokens. Also inside: the effort setting that errors on 6.1, Astra at 4.5x for one point, and Azure's 20% EU uplift. https://tokencost.app/blog/claude-sonnet-5-5-pricing https://tokencost.app/blog/claude-sonnet-5-5-pricing Tue, 29 Sep 2026 00:00:00 GMT Model Release Sonnet 5.5 shipped September 28 on Sonnet 5's card to the cent: $2 in, $0.20 cache hit, $10 out, which is also GPT-6 Sol's. Anthropic says it costs up to 30% less per task. On Artificial Analysis's index it is 42% and 44% cheaper than Sonnet 5 on medium and high and 13 to 15 points better, and on medium it beats Sonnet 5's best for a tenth of the money. On max it costs 28% more, after spending about 193,000 output tokens per task. Put all fifteen settings of Sonnet 5.5, Opus 5.5 and GPT-6 Sol on one ladder and ten survive: Sol holds every score up to 44, Opus 5.5 everything from 51, and the only Sonnet 5.5 setting left standing is high at 46.7 for $1,176, which is the API default. On max, Sonnet 5.5 costs more than Opus 5.5 on max and scores lower, and Opus on xhigh matches its score for 2.2x less. Also inside: Anthropic's own cost-per-task charts, the Terminal-Bench exception, and the breaking changes that will hit old code first. https://tokencost.app/blog/claude-opus-5-5-vs-gpt-6-sol-pricing https://tokencost.app/blog/claude-opus-5-5-vs-gpt-6-sol-pricing Mon, 28 Sep 2026 00:00:00 GMT Comparison Anthropic and OpenAI shipped their new mid-price flagships about 90 minutes apart on September 22. Opus 5.5 is $4 in, $0.20 cache hit, $20 out; GPT-6 Sol is $2, $0.20 and $10. Cache reads are the one line where they match. Artificial Analysis ran both at every effort level: at max, Opus 5.5 scored 58 for $8,708 to run the index and Sol scored 48 for $1,550, a 5.6x bill. But Opus at medium, its new default, scores 51 for $1,627, which beats Sol at max by three points for $77 more, and finishes each task faster. Past 272,000 input tokens Sol's whole request doubles on input and cache and goes 1.5x on output, so at 90% cache hits a 400K-token agent turn costs $0.272 on Opus and $0.334 on Sol. Also inside: why Opus 5.5 at max costs more per task than Opus 5 despite a card 20% lower, what Anthropic's 40% claim is actually scoped to, and the Batch card that makes Opus 5.5 cost exactly what Sol does. https://tokencost.app/blog/gpt-rosalind-pricing-billing-october-5 https://tokencost.app/blog/gpt-rosalind-pricing-billing-october-5 Tue, 15 Sep 2026 00:00:00 GMT Model Release GPT-Rosalind ran for 172 days without a rate card. On September 8 the API changelog called it generally available through the trusted-access program, on September 11 OpenAI added an update to both Rosalind posts, and a row appeared under 'Specialized models' on the pricing page: gpt-rosalind-research, $5.00 input, $0.50 cached, $25.00 output, with a note that billing begins October 5, 2026 and that cache-write pricing does not apply. The card is GPT-5.5's on input and cache and $5 under it on output, 0.833x on the line that dominates a reasoning bill, which makes Rosalind the only OpenAI specialist priced below its base: GPT-5.6 Cyber sits at 3.125x Sol. It is also Claude Opus 5's $5 / $0.50 / $25 to the cent, the difference being the $6.25 Anthropic charges to write a cache and OpenAI does not. On OpenAI's own June 3 numbers, Rosalind uses 31% fewer tokens than GPT-5.5 on GeneBench, 7.2% fewer on MedChemBench and 5.3% fewer on LabWorkBench, so a task costs between 5.3% and 42.5% less depending on which tokens were saved, a split OpenAI has not published. Also inside: no Batch, Flex, Fast or long-context row; no model page at all, so no context window, max output, cutoff or rate limits for a model sold on long reasoning over omics data; API use restricted to internal research and barred from customer-facing products; a ChatGPT credit card that bills 20 credits a message against Sol's 10 while the token rate is 1.25x Sol's; a setup guide that files the newly GA model under 'Legacy models'; and Anthropic's answer to the same market, which is the standard Claude line at the standard price with a published window. https://tokencost.app/blog/kimi-claude-opus-routing-distillation-cost https://tokencost.app/blog/kimi-claude-opus-routing-distillation-cost Mon, 14 Sep 2026 00:00:00 GMT Industry Anthropic's September 10 threat report gives five per-lab counts of 'exchanges observed' against Claude, over 151 million for Alibaba, 23 million for Moonshot, 12.1 million for DeepSeek, 3.4 million for Zhipu and 400,000 for Xiaomi, and never adds them up, prices them or says how long an exchange was. We did. Every Opus from 4.5 to 5 shares one card, $5 input, $0.50 cached, $25 output, so the version the report leaves vague does not change the arithmetic, only the shape of an exchange does: $0.03 at 1,000 tokens each way, $0.125 at the mean Opus 5 request OpenRouter measured last week, $0.21 at a 2,000-in, 8,000-out reasoning dump. That makes the campaign $5.7 million, $23.8 million or $39.9 million at list, which is 46 minutes, 3.2 hours or 5.4 hours of Anthropic's $65 billion run rate, and none of it was billed, because the report says the accounts ran on stolen cards. Then the relay. Moonshot sent almost 300,000 Kimi customer requests to Opus over ten days through 5,380 accounts, and Kimi K3's card is Opus's times 0.6 on input, cache and output alike, so a relayed request costs $0.1253 and bills $0.0752 for any token mix; on the K2.6 card the customers were more likely on, since K3 did not exist until July 16, it bills $0.0267, 21 cents on the dollar. Also inside: Alibaba's 3-million-a-day peak against OpenRouter's 2.4 million daily Opus 5 requests; the Xiaomi free trial the report says was a transcript-collection device; Cognition's SWE-2, post-trained on K3 and sold as 64% cheaper than Fable 5.1; the countermeasures that changed the product without moving a price; and seven things we could not find, starting with a token count. https://tokencost.app/blog/openai-agents-api-hosted-sandbox-pricing https://tokencost.app/blog/openai-agents-api-hosted-sandbox-pricing Sun, 13 Sep 2026 00:00:00 GMT Industry The Agents API went into public beta on September 10 with one pricing sentence: 'you simply pay for the tokens and tools your agents use.' The sandbox is a tool. OpenAI-hosted sandboxes bill at the container card Code Interpreter has used all year, $0.03 for 1 GB, $0.12 for 4 GB, $0.48 for 16 GB and $1.92 for 64 GB per 20-minute session, by the minute with a 5-minute minimum, and the Wayback Machine shows the built-in tools table byte-identical across the launch. Two things are missing. The Create Session API's openai_hosted object has no memory_limit, cpu or gpu field, so nothing tells you which of four rates spanning 64x you are on. And the pricing page's 'eligible container sessions' footnote never defines eligible or says whether the up-to-one-hour idle period before a sandbox is deleted bills, a question Simon Willison asked on a 345-point Hacker News thread that nobody from OpenAI has answered. Claude Managed Agents charges $0.08 per session-hour for up to 8 GB, metered only while running, with idle time free in writing. So OpenAI's box is 1.125x Anthropic's per hour if it is the 1 GB tier, 4.5x at 4 GB and 72x at 64 GB. Ported to Anthropic's own worked example, an hour of 50,000 input and 15,000 output tokens is $0.705 on Opus 5 with the runtime at 11.3% of the bill, $1.34 on GPT-6 Astra with a 1 GB sandbox at 6.7%, and $0.118 on GPT-5.6 Luna with the sandbox at 76.3%; on the 4 GB tier the Luna sandbox is 12.9x the model. Also inside: the rate card's two restructurings this year, from flat per container to 20-minute blocks on March 31 to per-minute in June, with no rate moving; the November 2025 'we no longer charge for unused containers' fix and why it does not cover a sandbox held open between turns; the 20 minutes of a 2 vCPU, 4 GB box priced at $0.0552 on E2B, Daytona and Blaxel against OpenAI's $0.12; Google's 'not billed during the preview period'; and the six things OpenAI has not published, starting with which row you are paying. https://tokencost.app/blog/sakana-fugu-max-pricing-orchestration-tokens https://tokencost.app/blog/sakana-fugu-max-pricing-orchestration-tokens Sat, 12 Sep 2026 00:00:00 GMT Model Release Fugu Max went on sale September 11 at $2.00 input, $6.00 output and $0.25 cached input per million tokens, flat to 1M context, as a new line on a Sakana card whose Ultra tier has not moved since June 22. The output claim is arithmetic and it checks: $6 is 40% under Sonnet 5, 50% under GPT-5.6 Terra and 60% under Kimi K3. The input line is identical to the first two. Then the paragraph on Sakana's pricing page that the launch post does not mention: orchestration tokens, the ones the router spends calling models in its pool and copies of itself, 'represent real token usage outside of the input and output tokens and will be counted in the final price of the request,' at the same $2 and $6, and unlike OpenAI's reasoning_tokens they are additive to input_tokens and output_tokens rather than inside them. Sakana publishes the fields and not a single typical value. So we computed the overhead at which the saving disappears: on a 30,000-in, 2,000-out coding-agent turn, Fugu Max costs the same as Sonnet 5 once orchestration adds 11.1% to the token count, Terra at 16.7%, Kimi K3 at 66.7%; on a 1,000-in, 5,000-out answer, 62.5%, 93.8% and 143.8%. Also inside: why Qwen3.8 Max, a single model on Sakana's own chart at the identical $2, $6 and $0.25, makes any orchestration token at all a loss; the launch chart's Gemini 3.8 Flash dot drawn at $7.50 output, Google's January 2027 price, when the price in force for the next 110 days is $3.75, and the recount that turns 'seven of ten' Pareto wins into six, three of them by half a point or less; the DeepSeek V4 Pro dot plotted at its seven-hours-a-weekday peak rate, on the same day DeepSeek quietly cancelled the September 14 routing that would have retired the model, a reversal we found in the Wayback Machine and have corrected in our own catalogue; Sonnet 5 and Terra batch cards that undercut Fugu Max's standard card on input; a pool the text calls open-weights and the diagram labels closed and open; the 72 tokens per second OpenRouter reads against Gemini's 274.9; the Artificial Analysis benchmark on a chart for a model Artificial Analysis does not list; and a Hacker News thread with 5 points where June's had 247. https://tokencost.app/blog/gpt-live-1-api-pricing-per-minute https://tokencost.app/blog/gpt-live-1-api-pricing-per-minute Fri, 11 Sep 2026 00:00:00 GMT Model Release GPT-Live-1 went generally available in the API on September 10 at $0.05 per minute of session, metered per second with no rounding, on the only endpoint it supports. Three things sit under that number. Active session time includes silence, so three minutes of hold is $0.15, the opposite of the Realtime API, where voice activity detection makes quiet free. The model doing the thinking is not the one you pay $0.05 for: GPT-Live-1 delegates to a Responses model of your choice, which bills its own tokens on top, and we priced a 10-minute call at $0.51 on GPT-5.6 Luna, $0.58 on Terra, $0.64 on Sol and $0.86 on GPT-6 Astra, with the backend at 1.5% of the bill at one end and 42% at the other. And the same 10-minute call on gpt-realtime-2.1 comes out between $0.51 and $3.22, because Realtime resends the whole conversation on every turn at $0.40 per million cached or $32 uncached, a 6.3x spread that GPT-Live-1 simply deletes. Also inside: the 15 seconds of duration that WebRTC session creation bills before anyone speaks; the $0.055 rate we infer for data residency endpoints from a footnote OpenAI has not applied to this model in print; why $0.05 is the exact number Vapi charges for hosting and Retell charges $0.055 for, with no model included; the Tau3 chart that puts GPT-Live-1 first with GPT-6 Astra at medium effort behind it, which is not the $0.05 product; concurrency limits that cap a Tier 1 account at $75 an hour; the unpublished session duration limit; and the Hacker News thread that has 10 points where July's had 750. https://tokencost.app/blog/deepseek-v4-1-flash-pricing-v4-pro-routing https://tokencost.app/blog/deepseek-v4-1-flash-pricing-v4-pro-routing Thu, 10 Sep 2026 00:00:00 GMT Model Release Correction, September 12: DeepSeek withdrew the September 14 routing on September 11, 'in response to user demand', and deepseek-v4-pro keeps its card with no end date; the Pro routing described below is the plan as announced. DeepSeek V4.1 Flash went live at 04:00 UTC on September 10 under the model name deepseek-flash at $0.15 input, $0.003 cache hit and $0.60 output off-peak, against $0.22, $0.007 and $0.66 for the V4 Flash it retires. Three lines, three different cuts: 31.8%, 57.1% and 9.1%, and the 60/33/11 you may have read are the yuan card's percentages, because DeepSeek set the new dollar lines at a different exchange rate from every other line on the same page. Then the footnote that matters. From 12:00 Beijing time on September 14, every request to deepseek-v4-pro is routed to V4.1 Flash and billed at its price: a 77.3% input cut, 69.7% on output and 86.4% on cache hits for anyone who does nothing, delivered by a 552B-parameter model with 16B active replacing a 1.6T one with 49B active. The original plan, posted to a user group on September 9, was to make the switch at launch; a 399-point Hacker News thread later, DeepSeek added four days, during which it is charging a 4.40x premium for a model it has stated in writing is worse. Also inside: the two of sixteen DeepSeek-published benchmarks on which Pro still wins, so 'comprehensively surpassed' is fourteen of sixteen; the fact that Artificial Analysis has not measured V4.1 Flash and that the 427 tokens per second going round comes from nobody's lab; the $0.15 input price DeepSeek now shares with GLM-5.3-Flash and Qwen3.8-Flash, and why the cheapest of the three depends on your cache-hit rate rather than the card; the peak row that puts V4.1 Flash above GPT-5.6 Luna for seven hours a weekday; and the silent edit we found in the Wayback Machine, which made weekends off-peak in late August with no changelog entry, and which means every blended DeepSeek figure we and everyone else published since then was about 7% too high. https://tokencost.app/blog/gpt-image-2-5-pricing-cost-per-image https://tokencost.app/blog/gpt-image-2-5-pricing-cost-per-image Wed, 09 Sep 2026 00:00:00 GMT Model Release GPT Image 2.5 arrived on September 8 with five price lines identical to GPT Image 2, to the cent, on both the Sunburst and Flare snapshots. The rate card is not where the money is. Image generation bills by the token, and the token counts behind each quality setting were rewritten: a 1024x1024 image at high fell from 7,024 output tokens to 1,756, so the same call costs $0.0527 instead of $0.2107. The ladder shifted down exactly one rung and the integers are exact rather than approximate, with 2.5's high spending what GPT Image 2's medium spent and 2.5's max spending what its high spent. So the widely reported story, that two new premium tiers mean a dearer model, is backwards: the most expensive image either model will sell you is $0.2107, and both still default to auto. The 2x price rise that led coverage was a batch column read against a standard one, a mistake OpenAI's own tabbed pricing page makes easy, and we show exactly how. Also inside: why the token counts in this post had to be reconstructed from the calculator OpenAI ships on its docs, since it publishes a static tokens-per-image table only for models older than these two; the cell where two reconstructions both survive the published figures, and what we left out because of it; OpenAI stating that Sunburst and Flare can cost different amounts per image and then shipping one estimator covering both; the streaming surcharge that can triple a low-quality call; why gpt-image-1 dies six weeks before the three models every roundup groups it with; and the reason our own pricing table carried no OpenAI image model until today. https://tokencost.app/blog/artificial-analysis-index-v4-3-cost-per-point https://tokencost.app/blog/artificial-analysis-index-v4-3-cost-per-point Tue, 08 Sep 2026 00:00:00 GMT Research Artificial Analysis shipped Intelligence Index v4.2 on September 4 and v4.3 on September 7. Claude Fable 5.1 scored 65.65 before them and 53.37 after, Claude Opus 5 went 63.05 to 50.70, and every model that was on the board beforehand lost between 12.3 and 18.1 points. No lab shipped weights in that window and not one rate card moved a cent. The version in the middle was current for three days: Artificial Analysis records its own active window for it as September 2026 to September 2026. Then the part that belongs on a pricing site. The last step alone made running the index 25% to 103% dearer per model on rate cards that did not move, because the suite got longer and models now emit half again to twice the output tokens per task, so Fable 5.1 went from $3.69 per task to $7.63 in five days while listing at the same $10.00 and $50.00 throughout. Divide the two moving numbers and cost per intelligence point rose between 154% and 290% across the models measured through both rewrites, which means there is no scalar that converts an old figure to a new one and the ordering genuinely reshuffled: rank eight models by cost per point and the entire middle four swap places with nothing but the methodology moving. Also inside: why the tie at 53 that made this week's headlines is a v4.3 artifact of a 0.560-point gap that was 2.101 three days earlier, and why GPT-6 Astra lost less than any other model on the board in that step; the Terminal-Bench slot where the task set, the agent harness and the step budget all changed at once, and the 5.9-point disagreement between Artificial Analysis and Snorkel on the same benchmark version that exceeds Snorkel's own error bar; the four places Artificial Analysis writes 'not directly comparable' about constituent benchmarks and the one place it has never written it; its own July headline claiming six labs field a model above 50 when the answer is now two labs; the live page still answering that four categories contribute 25% each under a v4.3 header; and our own rankings, which carry 104 models on an index fetched July 27, three revisions behind, with no version stamp on it at all. https://tokencost.app/blog/mercury-2-5-pricing-discount-expiry https://tokencost.app/blog/mercury-2-5-pricing-discount-expiry Mon, 07 Sep 2026 00:00:00 GMT Industry Mercury 2.5 has been selling since August 31 at $0.04 and $0.15 per million tokens. On September 8 it becomes $0.20 and $0.75. The discount is a flat 80% struck against every line at once, so the reversion is exactly 5.00x for everybody: no prompt shape, no cache-hit rate and no reasoning setting moves it a cent. That uniformity is rarer than it sounds, and it means there is nothing to optimise around. The price that lands is also not really a new price. Inception has charged $0.25 and $0.75 for Mercury 2 since March, and Mercury 2.5's list card is $0.20 and $0.75, identical on output to the cent across a version bump that doubled the context window. Then the part that made us write it: the countdown is published on OpenRouter and nowhere Inception controls. Its own site shows the list price with no discount and no date, its docs do not list the model, its API does not serve it, and its blog never announced it. Also inside: why the cheapest model in our basket changes hands twice in 33 hours without a single model changing, as Z.AI's GLM-5.3-Flash discount lapses a day and a half after this one; why an hour of Mercury 2.5 costs twice an hour of Mercury 2 at an identical output price, and what that says about buying speed on a per-token meter; the 1,107 tokens per second Inception claims against the 471 OpenRouter actually serves; the complete absence of any benchmark for this model from any source, vendor or independent; the four incompatible conventions six vendors use to tell you when a price expires; and a correction to our own May post, whose 788 tokens per second no longer matches the page it came from. https://tokencost.app/blog/document-parsing-api-cost-per-page https://tokencost.app/blog/document-parsing-api-cost-per-page Sun, 06 Sep 2026 00:00:00 GMT Research Cohere shipped Parse 5 on August 27 at $1.50 per 1,000 pages. AWS Textract charges $1.50 per 1,000 pages. Azure charges $1.50 and has since March 1, 2022. Google charges $1.50 past its free tier. The newest arrival on that list matched a four-and-a-half-year-old price to the cent rather than undercutting it, in a market where the average million tokens fell from $2.04 to $1.17 over a single summer. That $1.50 buys the characters and nothing else. The same physical page costs $4.00 for standalone layout at AWS, $10.00 for layout inside Azure's prebuilt bundle, $15.00 for tables, $50.00 for form fields and $70.00 for a loan packet, a 46.67x range across meters that all bill per page, and most of that is decided by which verb you picked rather than which vendor. Then the comparison everybody actually runs: model input price times tokens per page against $1.50. It is the wrong sum. A parser's page price includes the Markdown it hands back and an LLM's input price does not, so Gemini 3.8 Flash looks 7.75x cheaper at $0.19 per 1,000 pages and lands at $3.01 once output is counted, about twice the parser rate. The breakeven is about 348 output tokens a page, less than half of what a dense page of Markdown runs to. Also inside: why Claude Haiku 4.5 costs exactly $1.50 per 1,000 pages and we did not engineer that; why GPT-6 Astra cannot be costed per page from anything OpenAI publishes; the two live Google pages that say a PDF page is 258 tokens and 560 tokens without referencing each other; the $10.00 Cohere puts on AWS Textract in its own price-performance chart against the $1.50 on Textract's rate card; the fact that Cohere's canonical pricing documentation never mentions Parse or pages at all; why only Google will tell you that a spreadsheet tab is a page; and the twenty OCR rows in the industry's shared price file that have nowhere to put a number, including the one for Parse 5 that has kept it off this site since launch. https://tokencost.app/blog/mythos-5-1-daybreak-cyber-model-pricing https://tokencost.app/blog/mythos-5-1-daybreak-cyber-model-pricing Sat, 05 Sep 2026 00:00:00 GMT Industry Claude Mythos 5.1 landed September 1, Gemini 3.8 Flash Cyber September 2, and OpenAI's $1 billion Daybreak commitment September 3. Three approval-gated models in three days, and three incompatible answers to what one costs. Anthropic printed seven exact figures and they are Fable 5.1's figures to the cent, $10.00 input, $50.00 output, $12.50 and $20.00 cache writes, $0.25 cache read and $5.00 and $25.00 on Batch, including the 0.025x cache-read multiplier that broke a formula Anthropic had honoured on every model it ever sold. Then read the Cyber Verification Program page: it carries Opus- and Sonnet-class models today and will include Mythos-class models in the near future, so the complete price belongs to a model the cyber programme cannot presently sell a cyberdefender. Google went the other way and published nothing. No rate on Vertex, none on the Gemini API pricing page, none on the DeepMind model page, no cost statement in the Fairwind terms, and no API model ID anywhere, which stops looking like an oversight once you read the clause saying access cannot be shared, redistributed or sold. OpenAI published neither: gpt-daybreak-blue-latest and gpt-daybreak-red-latest are aliases onto gpt-5.6-sol and gpt-5.6-cyber, the two model pages carry rate limits and no rate card, and OpenAI states the pointers will be repointed later with pricing adjusted to match. That last one has a bill attached. The same $1 billion converts to 125 trillion tokens at Blue's $8.00 blended and 35.6 trillion at Red's $28.125, a 3.52x swing decided by nothing but which alias an organisation is approved for. Also inside: why Red is the only gated model of the three that also loses its discount tiers, turning a 1.41x rate gap against Mythos 5.1 into 2.81x on batchable work because Responses-API-only means no Batch endpoint at all; one 1,000-turn triage run priced on every card that exists, $116.40 to $401.25 inside a single programme; the 431-model OpenRouter catalogue with zero rows matching mythos, daybreak, cyber or glasswing while five Fable rows and two Gemini 3.8 rows sit there; why "same model" is a flat identity claim at Anthropic, a shared base plus extra training at Google, and not an identity claim at all at OpenAI; the 70% figure about to be misquoted this week and the better-sourced 86.2% and more sobering 47.2% beside it; and the two different US-only restrictions on Mythos 5.1 that do not cancel each other out. https://tokencost.app/blog/gpt-6-astra-pricing https://tokencost.app/blog/gpt-6-astra-pricing Fri, 04 Sep 2026 00:00:00 GMT Model Release OpenAI shipped GPT-6 Astra on September 3 at $10.00 per million input tokens, $1.00 cached, $12.50 for a cache write and $50.00 output. Claude Fable 5 has carried those four numbers since launch, and on September 1 Anthropic replaced it with Fable 5.1, which changed exactly one of them: the cache read, $1.00 to $0.25. So the two live flagships now agree on input, output, the five-minute cache write, both batch legs and the 128,000-token output ceiling, and disagree on one cell, 4x wide, in OpenAI's favour on nothing. Astra sits at the 0.100x cache-read multiplier that eight models across both price lists share; Fable 5.1 and Mythos 5.1 are the only two at 0.025x, and they are three days old. What that 4x is worth turns out to be bounded: on a 1,000-turn agent loop with 60,000-token prompts, Astra costs 4.4% more at a 50% hit rate and 20.2% more at 99%, because output at $50.00 anchors both bills. The gap that actually bites is not on the card at all. Any Astra request over 272,000 input tokens reprices the entire request at 2x input and 1.5x output, and Fable 5.1 bills its full million-token window flat, so one token takes a thousand-turn job from $839.20 to $1,578.40 and the same job at 400,000-token prompts runs $2,180.00 against $885.00. Also inside: why every cost claim in OpenAI's launch post is a token claim in disguise, back-solved to implied token ratios that range from 1.45x to 7.14x against the same competitor in the same post; the exact 2.50x that separates Astra from GPT-5.6 Sol on all ten published lines, including Fast, Batch and both long-context bands; why 1,050,000 is a sum of two ceilings rather than an addressable window, and why 70% of the input range you can fill bills at double; the Tier 1 account whose one maximum-length prompt is 84% larger than its entire per-minute token budget; Anthropic's own footnote conceding its tokenizer emits about 30% more tokens for the same text; and the fact that three venues were named on launch day and only one of them has published a price. https://tokencost.app/blog/gemini-3-8-flash-introductory-price-expiry https://tokencost.app/blog/gemini-3-8-flash-introductory-price-expiry Thu, 03 Sep 2026 00:00:00 GMT Model Release Google shipped its third Flash model in six weeks on September 2, at $0.75 per million input tokens and $3.75 per million output. Gemini 3.7 Flash and Gemini 3.6 Flash already carry those numbers, and not just those two: all thirteen priced rows across Standard, Batch, Flex and Priority are identical on all three models, twenty-six figures counting both sides of the expiry, with no cell where they differ. The word Google uses is introductory, but the date is December 31, 2026 for every one of them regardless of when they launched, so the length of your introduction depends on how late you were born. 3.7 Flash gets 141 days of the discount. 3.8 Flash gets 121. A model shipped in mid-December on the same three-week cadence would get sixteen. Every one of those thirteen rows doubles by exactly 2.0x on January 1, with no rounding residue anywhere, which makes the reversion trivial to forecast and impossible to route around: no tier, no cache hit rate and no prompt shape changes a uniform multiplier, and if you are not batching today, batching in January buys back exactly your December bill and nothing more. Also inside: the two archived captures 38 hours apart proving 3.6 Flash spent 23 days at full price before Google swept it onto a promotion built for a newer model; why Gemini 3.5 Flash is not the expensive one but the only current Flash whose price is a price, and how its 2.00x input gap becomes 1.00x and its 2.40x output gap becomes 1.20x on New Year's Day; the tell that 3.5 Flash's cache storage is already sitting on the post-reversion number; a thousand-agent-turn job priced on eight cards either side of the date, where 3.8 Flash on Priority undercuts 3.5 Flash on Standard by 18.2% today and costs 1.64x in January; why Vertex calls the same discount 50% credits back on net spend while AI Studio calls it a lower price; the sentence in Google's own launch post telling you to stay on 3.7 Flash because the new model might use more tokens, and the 3.4x token allowance that shrinks to 1.2x once the price doubles; a Pareto frontier claimed for the gated Cyber model on an axis with no published number; and the fact that OpenRouter carries every figure correctly and has nowhere to put the date. https://tokencost.app/blog/claude-fable-5-1-cache-read-price-cut https://tokencost.app/blog/claude-fable-5-1-cache-read-price-cut Wed, 02 Sep 2026 00:00:00 GMT Model Release Anthropic shipped Fable 5.1 and Mythos 5.1 on September 1 and left almost the whole card alone: $10.00 input, $50.00 output, $12.50 and $20.00 cache writes, $5.00 and $25.00 on Batch, a 1,000,000-token window billed flat. The cache read went from $1.00 to $0.25, and it is the only line that moved. That matters more than it sounds, because Anthropic has priced a cache hit at exactly 0.1x base input on every model it has ever sold, from Haiku 3.5's $0.08 to the retired Opus 4.1's $1.50, and these two models are the first exceptions at 0.025x. The company acknowledges it in a footnote whose word 'standard' is carrying the whole change. Since one line moved, your saving is forced: 75% of whatever share of your Fable 5 bill was cache reads, ceiling 75%. Anthropic's quoted 25% and 45% decode exactly to cache-read shares of 33.3% and 60%, so the two headline figures are one claim about two workloads, and your own invoice answers it better than either. Also inside: why the break-even did not move at all, one read on the five-minute cache and two on the one-hour before and after, because the write premium sets that threshold and writes did not change; why a cache miss now costs 50 reads instead of 12.5, quadrupling what a badly-ordered prompt costs you relative to a correct one; how the untouched one-hour write went from 1.86x to 3.56x cheaper than the five-minute one across a sustained hour; a 40-turn coding session where Fable 5.1 on the default TTL bills exactly what Fable 5 on the one-hour cache billed, $13.68 both; why the flagship now reads cached tokens for half of what Opus 5 does and only 1.25x what Sonnet 5 does on a base price 5x higher; what batch and US-only inference do to $0.25 in both directions; and the fact that the second model on the new multiplier is one you cannot buy. https://tokencost.app/blog/grok-4-6-long-context-cloud-pricing https://tokencost.app/blog/grok-4-6-long-context-cloud-pricing Sun, 30 Aug 2026 00:00:00 GMT Comparison xAI, Google, Amazon and Microsoft all list Grok 4.6 at $2.00 per million input tokens, $0.50 cached and $6.00 output. Copy those three numbers into a spreadsheet and every venue ties, which is why every comparison we found stops there. The rate card is not where these four differ. xAI doubles all three lines once a prompt reaches 200,000 tokens and bills the doubled rate across the entire request, output included, and Google's Vertex table copies that structure exactly. AWS publishes a flat card for the full 500,000-token window with no tier on it at all. Microsoft publishes a flat card too, and the reason is the one thing nobody wrote down: Azure documents grok-4.6 with a 200,000-token context window, not 500,000. There is no long-context surcharge on Azure because there is no long context on Azure. The ceiling is the threshold, and it sells 40% of the model at 100% of the headline price. Priced on one 260,000-token agent turn at an 85% cache hit rate, the same model and the same request runs $473.00 per thousand on xAI and Vertex, $260.15 on Bedrock's In-Region card, $236.50 on Bedrock Global, $118.25 on Bedrock Flex, and does not run at all on Azure. Also inside: why Bedrock's most expensive tier still undercuts xAI's ordinary long-context rate by 12.5%; the single token that costs 43 cents on xAI and two millionths of a cent on Bedrock; why xAI's caching docs say cached tokens push you over the line rather than under it; the one-token-wide disagreement between Google's 'longer than 200K' and xAI's 'at or above 200k' that sits on the same threshold; why OpenRouter's Bedrock row contradicts AWS's own model card and which one we think is wrong; and the fact that xAI's launch post quotes $2 and $6 and never mentions the tier at all. https://tokencost.app/blog/tencent-hy4-preview-pricing https://tokencost.app/blog/tencent-hy4-preview-pricing Sat, 29 Aug 2026 00:00:00 GMT Model Release Hy4 preview landed on Thursday at 06:09 UTC: 770B parameters with 49B active, a 1,048,576-token window, Apache 2.0 weights, and one seller. Tencent published the same announcement in two languages with two currencies, ¥6, ¥18 and ¥0.30 per million tokens on the Chinese page and $0.834, $2.501 and $0.042 on the English one, and every writeup we found repeated the dollar figures without saying what they are. Divide any line into the other and the three of them pin one exchange rate to a window 0.040% wide, between 7.1957 and 7.1986. Friday's spot was about 6.73, so the international card is frozen roughly 7% above the market and the dollar price is 6.49% below what the yuan price converts to today. Hy3 is frozen too, at a different rate, somewhere between 7.5686 and 7.5829, and that is where the interesting number comes from: the dollar cards make Hy4 look 6.3182 times its predecessor on input when Tencent moved ¥1 to ¥6, which is 6.0000. The 5.30% nobody is reporting is a currency market, not a pricing decision. Hy3 preview supplies a third rate, between 6.6116 and 6.7227, and the three windows do not overlap, so this is not one stale number left to drift but a fresh one taken at each listing from something that is not the spot market. Also inside: why the rounding residue in $2.501 is the evidence, and why Hy3's card cannot be tested the same way; how a $0.042 cache line that looks like a 1.27x increase is really Tencent cutting the cache rate from a quarter of input to a twentieth; the evening discount OpenRouter applies to Hy3 that does not follow to Hy4 and that Tencent does not document anywhere; what a coding agent replaying one repository costs on eleven cards; and the 6.95% price rise Tencent could ship on Hy4 tomorrow by announcing nothing at all. https://tokencost.app/blog/llm-api-price-increases-2026 https://tokencost.app/blog/llm-api-price-increases-2026 Fri, 28 Aug 2026 00:00:00 GMT Guide Introductory pricing has become the normal way to launch a model, so a growing share of the rates in your spreadsheet are offers with a clock on them rather than prices. We looked for every one carrying a provider-published expiry date. Eight clear that bar and they split down the middle: four price tokens, four price video. On the token side, GLM-5.3-Flash doubles on September 9 and got a week of coverage, while Upstage's Solar Pro 4 goes up tenfold the following evening, from $0.03/$0.12 to $0.30/$1.20, and we found nobody writing about it. On the video side, ByteDance is running three Seedance promotions at once and Alibaba's Wan 3.0 offer ends September 23. Excluding GLM-5.3-Flash, we could not find a single dated promotion on a text model from any Chinese vendor; every one is video generation, which is the finding rather than a footnote. Priced on one ordinary stack, these dates take a $605 month to $1,010 with no change to what you run, and 92.6% of that increase is the Gemini 3.6 and 3.7 Flash doubling on January 1 that nobody mentions, against 7.4% for the GLM expiry that dominated the week. The ninth entry is the interesting one: GPT-5.6 Sol's promotional $4/$20 is listed everywhere as expiring on November 21, but OpenAI's wording is 'available at least through' that date, which is a floor and not a deadline. Our own catalogue stored it as a cliff and would have raised the calculator's Sol estimate 40% on OpenAI's behalf for a change OpenAI never announced; this post removes it. Also inside: why no Gemini discount lever helps, since caching, Batch and Vertex Priority all come to exactly 2.00x across the date; why GLM-5.3-Flash stays 11.5x cheaper than Gemini Flash even after losing its promotion; why only 3 of 15 sellers charge GLM's promotional rate today; how Vercel ships two model IDs so you can choose whether a free tier ending is a bill or an outage; and the one increase in this window that a provider called off. https://tokencost.app/blog/glm-5-3-flash-vs-qwen3-8-flash-pricing https://tokencost.app/blog/glm-5-3-flash-vs-qwen3-8-flash-pricing Thu, 27 Aug 2026 00:00:00 GMT Comparison Z.ai listed GLM-5.3-Flash at 13:59 UTC on August 26 and Alibaba listed Qwen3.8-Flash at 19:37. Every tracker shows the first at $0.075 per million input tokens against the second's $0.16, which reads as a rout and is a fortnight-long one. OpenRouter's endpoint API returns a discount field of 0.5 on Z.ai's own endpoint, and Z.ai's pricing page confirms what that means: the list card is $0.15, $0.03 cached and $0.50, and the halved rates expire on September 9 at 24:00 UTC+8. Six of the ten sellers already charge list, so the same MIT-licensed weights trade at a clean 2.00x spread this morning and OpenRouter's catalogue page reports only the bottom of it. To price the two cards against each other we recovered a real workload instead of inventing one: Artificial Analysis published a $138.02 total and 150M output tokens for GLM-5.3-Flash at list rates, which backs out to 420.13M input tokens. On those measured tokens the list card bills $138.02 and Qwen3.8-Flash bills $137.72, a gap of thirty cents, or 0.22%. Also inside: why GLM-5.3-Flash is cheaper on fresh input and dearer on both output and cached input, why Alibaba's automatic 10x cache discount beats Z.ai's manual 5x on any agent workload and flips the ranking outright, why a cache write costs Alibaba's customers 1.25x a fresh token and still pays off after two reads, why GLM-5.3-Flash cannot switch reasoning off while Qwen can cap it, why the flagship GLM-5.3 generates 1.70x faster than the model named for speed, and the scoring of our own August 24 estimate for this model, which came in 8.93x too high. https://tokencost.app/blog/vision-api-cost-per-image https://tokencost.app/blog/vision-api-cost-per-image Wed, 26 Aug 2026 00:00:00 GMT Comparison Every pricing page quotes dollars per million tokens and almost none of them say how many tokens a picture is. We collected the counting rules from seven labs, implemented them, and checked them against the vendors' own published examples. Anthropic's held up on all twelve figures of its table: 28-pixel patches, a resize that satisfies an edge limit and a token budget at once, and a tier split at Claude 4.7 that makes an identical 4K screenshot cost 1,560 visual tokens on Haiku 4.5 and 4,784 on Sonnet 5. OpenAI is the outlier and the story. GPT-5.6 Sol, Terra and Luna have no patch-budget limit on the detail settings original and auto, and auto is what you get by writing nothing, so a 4K frame bills 9,792 tokens and one thousand of them cost $39.17 against Opus 5's $23.92, despite Sol charging less per token. Setting detail to high takes 71.8% off. Alibaba ships the same capability with the flag defaulting the other way, and Qwen3-VL Plus lands at $0.4962 where GPT-5.6 Luna, at the identical $0.20 per million, lands at $1.9584. Also inside: why a 1080p, a 1440p and a 4K screenshot all cost the same 1,560 tokens on a standard-tier Claude model, why 1456x819 is the one resolution worth memorising and what pre-resizing to it saves on Opus 5, why the tool that generates screenshots fastest is the one where automatic downscaling is switched off, the one-pixel error in Anthropic's own table, Google's two documentation pages that cannot both be right, and why xAI is the only lab here we could not put a row against, because it still bills image tokens and deleted the rule that produces them. https://tokencost.app/blog/hy-mt2-pricing-vs-google-translate https://tokencost.app/blog/hy-mt2-pricing-vs-google-translate Tue, 25 Aug 2026 00:00:00 GMT Comparison Tencent put its Hy-MT2 translation family on a metered API last week, three months after releasing the weights under Apache 2.0. We counted the actual tokens rather than estimating them: 3,000 segments of the WMT22 general test sets through Hy-MT2's own tokenizer, through OpenAI's o200k_base, and through the arithmetic on Google's pricing page. Localizing one million English characters into Chinese, German, Russian and Czech costs $0.4800 on the flagship card, $0.2877 on the 1.8B, $1.5145 on GPT-5.6 Luna and $80.00 on Google's default NMT card. The 166x headline is the least useful number in the table. Two findings matter more. Hy-MT2 publishes no cached-input rate on any of its three models, so the terminology and style control its own model card leads with is billed fresh on every request; a 400-token glossary prefix raises the English to German bill 244.5%, Luna absorbs the same prefix for 21.9% at its $0.02 cached rate, and a 3.02x price advantage collapses to 1.07x. And the token itself is not a fixed unit: Hy-MT2 encodes Chinese in 0.7875 of the tokens o200k_base needs and Russian in 1.5241 of them, so two labs can publish the same rate and bill 52% apart on identical text. Also inside: why the 7B is strictly dominated by the flagship at a byte-identical price, why Google's own NMT and Translation LLM cards swing 3.7x depending only on which direction you are translating, why the 4,096-token completion cap binds long before the 8,192-token context does, and the 23 output tokens per second nobody mentions. https://tokencost.app/blog/ox-alpha-stealth-model-pricing https://tokencost.app/blog/ox-alpha-stealth-model-pricing Mon, 24 Aug 2026 00:00:00 GMT Model Release An unattributed model appeared on OpenRouter on August 20 at 20:04 UTC priced at nothing, and the weekend went to guessing which lab owns it. The listing answers a more useful question. Ox Alpha's record matches Z.ai's GLM-5.3 on eight of nine descriptive fields OpenRouter publishes, including a default temperature of 1 with top_p 0.95, a 2098-12-31 expiration placeholder, and a reasoning block offering exactly max, high and low. The ninth field breaks the popular conclusion: Ox Alpha takes image and video input and GLM-5.3 takes text only, so this is a superset rather than a rebadge, and multimodal siblings do not usually arrive on the same card. The expensive detail is the reasoning block. Reasoning is mandatory and defaults to max, a combination only 3 of 422 listings carry, and the other two are GLM-5.3 and its alias. GLM-5.3 is the worked example of what that costs: three generations on one unchanged $1.40 and $4.40 card, and cost per Artificial Analysis task still rose from about $0.44 to about $0.68, because output went from 140M tokens to 170M. We priced one month of agent work, held identical, on every card Ox Alpha could land on: $0.00 today, $493.20 at the cheapest of GLM-5.2's 33 open-weight endpoints, $1,067.04 on GLM-5.3's single-seller card, $1,777.25 at the dearest. Output is 59.4% of that middle figure. Also inside: why the zero has no list price behind it, why Ox Alpha is the only free reasoning model on the platform with no companion row to read, why nothing in this family caches implicitly and what that does to the bill, the $0.5767 a single response can reach, why the seller with the lowest input rate finishes eighth on the actual bill, and where the August 27 end date everyone is repeating actually comes from, which is a six-day promotion by a coding tool rather than anything OpenRouter has published. https://tokencost.app/blog/gpt-5-6-sol-price-cut https://tokencost.app/blog/gpt-5-6-sol-price-cut Sun, 23 Aug 2026 00:00:00 GMT Industry OpenAI dropped GPT-5.6 Sol on August 21 under a headline of "over 20%", which undersells it on output and oversells it as permanent. Input fell 20% to $4.00, cached input to $0.40, output fell 33.3% to $20.00, and every derived cell followed exactly: cache reads stay 90% off, cache writes 1.25x input, batch and Flex half, Fast a flat 2x, and the above-272K rule still 2x input and 1.5x output. Twelve cells, two degrees of freedom. Put the new card beside Claude Opus 5 and every comparable ratio is 0.800000, on input, cached input, output and both batch legs, which makes the comparison mix-independent: we priced four very different workloads and Sol comes out at 0.8000 times Opus 5 on all four. Before Thursday it matched on input and was 1.2x on output, so Opus 5 won anything that generated a token. The catch is a date. OpenAI says the rate holds at least through November 21, and if it lapses Sol goes back to costing between 1.033x and 1.191x Opus 5 depending on how output-heavy the job is, so the uniform ratio is uniform only until it is not. Then there is the channel lag: Bedrock passed the cut through in two days and preserved its 1.1000x In-Region multiplier to the cent, while every Azure meter still carries a July 1 effective date and reads $5.00 and $30.00, turning parity into 1.25x input and 1.50x output. Eight channels, eight prices, 2.75x on input and 3.30x on output for one model on one day. Also inside: why our own August 17 Terminal-Bench post went stale in five days and what Sol's run reprices to, why the Daybreak Blue approval penalty grew from 10% to 37.5% and 65% without anything about the approval changing, why Terra is the tier that quietly lost, and the 62.5x that now separates the cheapest and dearest output rates in one family. https://tokencost.app/blog/glm-5-3-pricing https://tokencost.app/blog/glm-5-3-pricing Sat, 22 Aug 2026 00:00:00 GMT Model Release Z.ai put a per-token price on GLM-5.3 on August 18, four days after the model arrived reachable only through a subscription, and the price is interesting precisely because it is not new: $1.40 input, $0.26 cached, $4.40 output, identical to GLM-5.2 and to GLM-5.1 before it. Three generations, one card. Then Artificial Analysis published what its Intelligence Index actually costs to run, and that went from about $0.44 a task on GLM-5.2 to about $0.68 on GLM-5.3. Same rates, 55% more money, because the newer model spends 170M output tokens across the suite against 140M and 2.4 times the median for its class. It ties Kimi K3 at 60 on the index while costing $0.90 per million blended against K3's $2.31, but that single number hides a spread: K3 is 3.41x on output, 2.14x on fresh input and only 1.15x on cached reads, which is exactly the shape a coding agent runs at. The bigger omission in most coverage is that GLM-5.3 has one seller. GLM-5.2's MIT weights put it on 21 providers at a flat 76% below Z.ai's own card, so upgrading from a third-party GLM-5.2 costs 4.1667x more on both lines before the verbosity penalty. Also inside: why the weights are still not out on day eight of a two-week promise, why there is no batch lane at all, and why the much-quoted 50% off-peak discount never touches the metered API. https://tokencost.app/blog/claude-browser-use-tool-token-cost https://tokencost.app/blog/claude-browser-use-tool-token-cost Fri, 21 Aug 2026 00:00:00 GMT Guide Anthropic made computer use generally available on August 19 and launched a browser tool beside it, publishing the token overhead for both, which almost no vendor does. Declaring browser_toolset_20260801 costs about 6,610 input tokens on Fable 5, Mythos 5, Opus 5 and Opus 4.8, and about 6,670 on Sonnet 5; the computer toolset costs about 4,520. Those are fixed and they land on every request, so on Fable 5 the browser figure is 6.6 cents before Claude does anything at all. The computer-use beta charged 1,201 to 1,234 tokens for the same job, though part of that gap is the newer tokenizer rather than the toolset, which puts the like-for-like rise nearer 2.8x to 4.2x. The fix is one line: Anthropic explicitly supports cache_control on the toolset entry, refuses it inside a member's configs because the members load as one definition, and bills cache reads at a tenth of input. A 50-turn loop on Opus 5 then pays $0.20 in definitions instead of $1.65, a ratio of 8.13x that holds for every model at every price because both the token count and the rate cancel out. Also inside: why changing a tool definition clears the whole cache, the July beta that exempts four models from exactly that, why batch actions quietly eat your four breakpoints, why a 1920x1080 screenshot is 2,691 visual tokens rather than 1,560 and costs more per turn than the cached toolset, why an oversized screenshot is a validation error rather than an expensive request, why Bedrock sells only the legacy tool and browser use not at all, and the tool-use baseline Anthropic publishes for thirteen models but not for the two dearest ones on its own API.