Grok 4.6 costs exactly what Grok 4.5 costs on input, on output, and at the cliff. The only line that moved is cached input, up 67%, and that is the line an agent loop actually lives on.

Four of the five lines are the same number twice

xAI publishes both models in one table, which makes this easy to check rather than infer. Grok 4.5 has not been repriced, retired, or moved. It is still listed, still sold, still carries the same 150 RPS and 50M TPM limits as its successor, and its model page has no deprecation banner. The two models are being sold side by side at prices that differ in exactly one place.

One more thing worth checking before you assume you have already upgraded. Grok 4.6 ships with no published aliases, and grok-build-latest still points at Grok 4.5. If your coding setup pins the alias rather than the version, you are on the older model and the cheaper cache, and nothing about this post costs you anything yet.

The better your cache works, the more the upgrade costs

A cache discount only matters in proportion to how much of your input is cached. That makes the headline 67% misleading in both directions: it overstates the impact on one-shot chat traffic and understates it on the agent loops xAI built this model for. The honest way to express it is a blended input rate. Mix $2.00 misses and cached reads at whatever ratio your workload actually hits, and compare.

The shape of that column is the point. Cache optimisation is the main lever teams pull to make agent work affordable, and this change takes back a share of every gain that lever produced. A team that did nothing about caching pays 9% more. A team that spent a quarter getting to 95% pays 49% more. We wrote up the mechanics of what cached tokens cost across providers in our caching pricing guide, and the ratios there are worth rereading before you assume the discount is a constant.

Above the 200K cliff the arithmetic is identical, because both tiers double together. The premium at any given hit rate is the same whether you are paying the cheap tier or the expensive one.

Cached tokens count toward the 200K cliff

xAI's threshold rule is harsher than the version most write-ups describe, in two separate ways. First, crossing the line rebills the whole request, not the overage. The pricing page footnote says requests whose prompt reaches the listed token threshold are billed at the higher rate for all tokens in the request. A 201,000 token prompt costs roughly twice a 199,000 token one, and there is no gentle slope between them.

Second, and this is the part that catches caching-heavy setups, cached tokens count toward the trigger. xAI's caching documentation states that long context pricing applies when total prompt tokens including cached tokens exceed the threshold, and that both cached and non-cached tokens then use their respective long-context rates. So a large cached prefix can push you over the line by itself, and the reward for crossing it is that your cache reads reprice from $0.50 to $1.00. Caching lowers the rate you pay per token. It does not keep you below the cliff.

The threshold sits at 200,000 tokens on a 500,000 token window, which means the penalty tier starts at 40% of stated capacity. That is a tighter ratio than OpenAI's, where the GPT-5.6 cliff arrives at 272K. Cliffs are not unique to xAI, though: Google prices Gemini 3.1 Pro Preview at $2.00 below 200K and $4.00 above it, the same threshold. What is worth knowing is which of your options are flat. Current Claude models bill the full window at standard rates, and so do Gemini 3.7 Flash and DeepSeek V4 Pro, at any prompt length.

One number xAI does not publish anywhere is the cache TTL. The FAQ says only that cache entries can be evicted at any time due to server load or restarts. Routing is the other half of it: xAI's documented way to force a cache miss is to change or omit the x-grok-conv-id header, which routes the request to a potentially different server where no cache exists for your prompt. Read that in reverse and it is a warning: keeping the header stable is what keeps you on a warm server. There is no cache write fee and no storage rent, which is a real advantage we come back to below, but there is also no guaranteed lifetime, so the hit rates in the table above are something you measure rather than something you are promised.

A month of agent traffic, held constant and then not

Take a coding agent that pushes 400M input tokens and 20M output tokens a month at a 90% cache hit rate, all of it under the 200K line. Holding the workload identical isolates the rate change, which is the only fair first comparison.

Row two is the rate change alone: $72 a month, or 23.4%, on a card whose headline numbers did not move. Row three adds the second effect, which is that Grok 4.6 is more verbose. Artificial Analysis measured 21,735 output tokens per task against Grok 4.5's 14,795 across the same evaluation suite, a 46.9% increase, and output tokens are billed at $6.00 whether they were worth generating or not. Apply that to the same agent and the month lands at $436.28, which is 41.6% above where it started.

Two honest caveats on row three. The verbosity figure comes from AA's benchmark mix at high reasoning effort, not from your workload, so treat it as a direction rather than a coefficient. And longer responses are sometimes the reason the model scores better, so this is not pure waste. It is still money. Run your own numbers in the calculator with your real hit rate before you assume either row applies.

Artificial Analysis paid 132% more per task for 9% more intelligence

The cleanest independent read on all of this comes from someone who actually ran both models end to end on the same suite and published the invoice. Artificial Analysis spent $0.3601 per task on Grok 4.5 and $0.8367 on Grok 4.6, an increase of 132.4%, against an index score that rose from 55.76 to 60.92, an increase of 9.3%. Neither the input price nor the output price changed between those two runs.

AA's own cost decomposition explains where it went. These next figures are totals across the whole run rather than per task, where the bill moved from $579.21 to $1,068.47. Non-cached input spend was essentially flat, moving from $24.01 to $23.96. Cache-read spend went from $84.70 to $357.94 and cache-write from $111.33 to $253.63. The bill grew almost entirely on the cache side, which is exactly what a 67% cache read increase plus a larger replayed context produces.

The genuinely good rows are the hallucination rate, which fell almost twenty points, and the agentic index, which is where xAI aimed. The flat rows are worth naming too: Humanity's Last Exam moved 0.23 points, SciCode went slightly backwards, and median time to first token got 3.16x worse. If you are paying by the task rather than by the token, that last one is a cost as well.

A caution on benchmark shopping. xAI's launch table reports Terminal-Bench v3.0, where Grok 4.6 scores 26% and Grok 4.5 scores 15.7%. AA's index uses Terminal-Bench v2.1, where the same two models score 88.39% and 81.65%. Both pairs are real; they are different benchmarks with similar names, and putting 26% and 88% into the same column is how bad comparison tables get made. xAI also did not publish SWE-bench Pro for Grok 4.6 at all, having switched to DeepSWE v1.1 and APEX-SWE, so there is no like-for-like successor to the 64.7 that Grok 4.5 posted.

Grok 4.6 now has the shallowest cache discount of any frontier model

Cache read rates cluster hard across the industry. Anthropic, OpenAI, and Google all price a cached read at exactly 10% of their input rate, which has become the de facto standard. Grok 4.5 was already off that standard at 15%. Grok 4.6 moves to 25%, and as far as we can tell that is the thinnest cache discount on any frontier model currently sold.

The percentage column is the one to read, because it is what survives when input prices differ. Claude Sonnet 5 has the same $2.00 input rate as Grok 4.6 and charges $0.20 for a cached read against Grok's $0.50, so cached context costs 60% less on the line that dominates agent bills, though its output is $10.00 against $6.00. Which one wins depends entirely on your output-to-cached-input ratio, and it is no longer obvious in either direction. Sonnet 5 also got cheaper to plan around this month, because Anthropic cancelled the September 1 increase that would have taken it to $3 and $15.

One correction to our own framing, because the read column is not the whole cache bill. Anthropic charges to write a cache entry, $2.50 per million at the 5 minute TTL on Sonnet 5 and $4.00 at the hour, and OpenAI charges $6.25 per million on GPT-5.6 Sol. xAI charges nothing to write, and nothing to store. So the table above flatters the models with cheap reads and expensive writes, and the gap narrows or reverses on workloads whose prefixes change often enough that most entries get written more than they get read. If your cache is long-lived and heavily reused, the read column is the one that decides it and Grok 4.6 is the worst of this group. If your prefixes churn, xAI's free writes are worth more than its shallow discount costs you.

Two of these rows have expiry dates attached, which is worth knowing before you build a spreadsheet on them. DeepSeek's rates change on August 16 at 16:00 UTC to a peak and off-peak structure, and the new cache-hit rate is several times the old one. Gemini 3.7 Flash's numbers are promotional through December 31 and every line on the card doubles on January 1. The Grok and Claude rows carry no announced end date.

There is no batch discount, because there is no batch

The usual escape hatch for cost-sensitive work does not exist here. xAI's batch documentation says grok-4.6 and grok-4.5 are not supported for Batch API requests and will be rejected. Not discounted at zero percent, rejected. The pricing page grants 20% off to grok-4.3 and the three grok-4.20 variants, then states that models not listed have no batch discount.

Anthropic, OpenAI, and Google all offer 50% off batch on their current models. If you have overnight work that would happily wait, xAI's only answer is to run it on Grok 4.3 at $1.25 and $2.50, which is a different model with different quality, not a discount on this one. The other direction is Priority Processing. xAI's announcement mentions a fast variant at twice the price but never names it, and Priority Processing is the only 2x option in the docs, so we are joining those two ourselves. There is no separate slug for it. It is a flat 2x multiplier on every token type including cached reads, applied after the caching discount, and you are only billed at 2x when the response confirms the priority tier. Worth noting where that lands you: priority under 200K costs exactly the same as standard above 200K.

The cheaper model is still on the shelf

Grok 4.6 is a real improvement and it is not free, and the pricing structure obscures the second half of that sentence. Nothing on the headline card moved, so every tracker and comparison page will show $2 and $6 and call it a like-for-like upgrade. The actual price is in the cache line, and it lands hardest on the teams who did the most work to control their spend.

So here is the split we would make. If your hit rate is under about 55%, upgrade without thinking about it: the delta is around 10% or less and the hallucination rate alone is worth it. If you are above 85%, measure first, because you are looking at something between 31% and 63% more for input plus whatever the extra verbosity adds, and Grok 4.5 is still sold at the old cache rate with no retirement date announced. That last fact is the useful one. xAI left the cheaper model on the shelf, so this is a choice you get to make per workload rather than a migration you have to accept. Check what your alias resolves to, since grok-build-latest still serves 4.5, then put your own hit rate and token mix into the calculator and compare against the full rate card list. The rate card comparison that matters here is not the one xAI printed.

What each source settles

• xAI: pricing - Every rate in this post for both models, in adjacent rows of one table: Grok 4.6 at $2.00 / $0.50 / $6.00 under 200K and $4.00 / $1.00 / $12.00 above, Grok 4.5 at $2.00 / $0.30 / $6.00 and $4.00 / $0.60 / $12.00. Also the source for the whole-request rebilling footnote, the 20% batch discount list that excludes both models, the Priority Processing 2x multiplier, and the server-side tool rates at $5 per 1,000 web and X search calls.
• xAI: models - The 500,000 token context window and the February 1, 2026 knowledge cutoff. The new xhigh reasoning effort is documented separately under model capabilities, and the alias listing that shows grok-4.5-latest and grok-build-latest still belonging to Grok 4.5, with none on Grok 4.6, is on each model's own page rather than this index. Appending .md to any xAI docs URL returns plain markdown, which is how these numbers were read without the JavaScript render.
• xAI: prompt caching usage and pricing - The load-bearing sentence for the cliff section, stating that long context pricing applies when total prompt tokens including cached tokens exceed the threshold, and that both cached and non-cached tokens then use their long-context rates. Also the absence of any published cache TTL, and the warning about paying full input price on a cache-cold server without a prompt cache key.
• xAI: batch API - The statement that grok-4.6 and grok-4.5 are not currently supported for Batch API requests and will be rejected.
• xAI: introducing Grok 4.6 - The August 12, 2026 dateline, the vendor benchmark table including Terminal-Bench v3.0 at 26% against Grok 4.5's 15.7%, and the reference to a fast variant at twice the price, which the docs never link to anything by name. Priority Processing is the only 2x option xAI sells, so we treat that as the match and say so rather than asserting it.
• Artificial Analysis: Grok 4.6 - The independent measurements in the benchmark table: index 60.92 against 55.76, cost per task $0.8367 against $0.3601, 21,735 output tokens per task against 14,795, hallucination rate 34.29% against 54.15%, Terminal-Bench v2.1 at 88.39% against 81.65%, and the cache-read and cache-write spend split that shows where the cost increase came from. Both models tested at high reasoning effort.
• xAI: May 15 retirement notice - The only retirement page xAI publishes, dated three months before this launch, covering eight older slugs and naming neither Grok 4.5 nor Grok 4.6. This is the basis for saying Grok 4.5 was not retired.
• Anthropic: pricing - Claude Sonnet 5 at $2.00 input, $0.20 cached read, $10.00 output, and the note that the introductory price is now the standard price, which is what cancelled the September 1 increase.
• Google: Gemini API pricing - Gemini 3.7 Flash at $0.75 input and $0.075 cached read, with the December 31, 2026 end date on the promotional rates.
• DeepSeek: pricing - V4 Pro at $0.435 input and $0.003625 cache hit, the rates that expire on August 16, 2026 at 16:00 UTC.
• OpenAI: pricing - GPT-5.6 Sol at $5.00 input, $0.50 cached read, $30.00 output, used for the cache-as-percentage-of-input column.
• Three gaps worth naming. xAI publishes no cache TTL for any model, so every hit rate in this post is a workload property rather than a guaranteed one. xAI publishes no maximum output token count for Grok 4.6, stating only that there is no text output limit, so any numeric output cap you see for the Grok line, including the 16,384 our own catalogue carries, is not a vendor figure. And xAI's reported GDPval-AA v2 score of 1753 does not match the 1746.42 in Artificial Analysis's own data, despite both citing the same benchmark, a small discrepancy we could not resolve and did not use.