Claude Sonnet 5 lands level with Opus 4.8 on an independent benchmark and asks 60% less for it. Two footnotes decide whether the bargain survives September.
Anthropic shipped Sonnet 5 on June 30 at $2 input and $10 output per million tokens, an introductory rate that expires August 31. On GDPval it sits a statistical hair from Opus 4.8 while costing a fraction of it. Then the two footnotes: the price steps up to $3/$15 on September 1, and a new tokenizer counts the same text as roughly 30% more tokens. Read together, they turn a sticker that matches Sonnet 4.6 into a bill that does not.
Image source: Anthropic
The rate card, before and after August 31
| Tier | Input | Output | Cache read | Batch (in/out) |
|---|---|---|---|---|
| Introductory (through Aug 31) | $2 | $10 | $0.20 | $1 / $5 |
| Standard (from Sept 1) | $3 | $15 | $0.30 | $1.50 / $7.50 |
Per million tokens, from Anthropic's official pricing docs. Cache writes run $2.50 (5-minute) and $4 (1-hour) at the intro rate, $3.75 and $6 at standard. Context window is 1M tokens with no long-context surcharge; max output is 128K synchronously, up to 300K on the Batch API beta.
It draws level with Opus on the benchmark that priced it
Anthropic did not bury the headline. On GDPval-AA v2, an economic-value benchmark scored independently by Artificial Analysis, Sonnet 5 posts an Elo of 1618. Opus 4.8 sits at 1615. That is a tie inside the error bars, and it puts a model priced at $2 output-adjacent to nothing on the board except the two Claude models that cost far more. Fable 5 leads the sweep at 1783; Anthropic takes the top three slots outright.
Coding is more mixed, which is the honest read. From the Sonnet 5 system card, run at max thinking effort and averaged over five trials:
| Benchmark | Sonnet 5 | Sonnet 4.6 | GPT-5.5 |
|---|---|---|---|
| SWE-bench Verified | 85.2 | - | - |
| SWE-bench Pro | 63.2 | 58.1 | 58.6 |
| Terminal-Bench 2.1 | 80.4 | 67.0 | 83.4 |
| GDPval-AA v2 (Elo) | 1618 | 1395 | 1509 |
| Humanity's Last Exam (tools) | 57.4 | 46.8 | 52.2 |
Anthropic-reported figures from the Claude Sonnet 5 system card. Competitor numbers are drawn from each developer's own card or leaderboard, not head-to-head reruns. The card does not publish a Sonnet 4.6 SWE-bench Verified figure, so we leave it blank rather than borrow a secondary one. Terminal-Bench 2.1 for GPT-5.5 is its Codex CLI result.
The jump over Sonnet 4.6 is real and broad: GDPval climbs 223 Elo points, Humanity's Last Exam with tools gains eleven, SWE-bench Pro adds five. Against GPT-5.5 it trades wins, losing Terminal-Bench and taking the rest. What you are buying is an Opus-adjacent generalist at a Sonnet price. Whether that stays true after September depends on the two things Anthropic put in the fine print.
The intro price is quietly doing a job
Sonnet 5 switches to the newer Claude tokenizer, the one already running under Opus 4.7 and later, Fable 5, and Mythos 5. Sonnet 4.6 and everything before it use the old one. The switch buys better handling of code and non-English text, and it charges for it in the only currency this site cares about: tokens. Anthropic's own words, from two different pages, are worth quoting because they frame the same fact two ways.
The launch post says the same input "can map to more tokens, roughly 1.0 to 1.35 times depending on the content type." The pricing docs put an average on it: "approximately 30% more tokens for the same text." Both are true. Prose sits near the bottom of that range, code and JSON near the top. You do not choose which; your workload does.
Now line up the dates. Through August 31 you pay $2/$10, a third under Sonnet 4.6's $3/$15. On September 1 you pay exactly Sonnet 4.6's sticker. But you are metering more tokens per unit of real text the whole time. So the intro discount is not generosity; it is roughly the size of the tokenizer tax, set to keep the switch from Sonnet 4.6 close to cost-neutral until the discount ends and the tax stops being offset.
Same sticker as Sonnet 4.6, not the same bill
Here is the comparison the price sheet will not run for you. Take a fixed body of source text that tokenizes to 10M input and 2M output on Sonnet 4.6's old tokenizer, a small production month. Feed the identical text to Sonnet 5 and the token meter reads higher. How much higher depends on whether it is prose or code.
| Same text, one month | Sonnet 4.6 | Sonnet 5 (intro) | Sonnet 5 (standard) |
|---|---|---|---|
| Prose-heavy (~1.0x tokens) | $60 | $40 | $60 |
| Code and JSON (~1.3x tokens) | $60 | $52 | $78 |
Baseline is 10M input and 2M output as counted by Sonnet 4.6. Sonnet 5 columns apply its own token count: the same at 1.0x for prose, 30% more at 1.3x for code and JSON, using the intro $2/$10 and standard $3/$15 rates.
Read the code row, because that is where most agent traffic lives. Today Sonnet 5 does that month for $52 against Sonnet 4.6's $60, a 13% saving that reads like a straight price cut. After September 1 the same month is $78, or 30% more than Sonnet 4.6 for output that is meant to be better but sticker-identical. The prose row is gentler: break-even at standard, a clean third off during the intro. Nobody is being tricked here. But if you migrate off Sonnet 4.6 this summer on the strength of the intro price, put a reminder in the calendar for the last week of August.
Where it lands against the rest of the board
Against Sonnet 4.6 the tokenizer muddies the water. Against the frontier tier it does not, because Opus 4.8, Fable 5, and GPT-5.5 either share the same new tokenizer or count tokens on their own scale, so a straight rate comparison is fair. Here Sonnet 5 is simply the cheap seat with a frontier score.
| Model | Input | Output | 20M in / 4M out |
|---|---|---|---|
| DeepSeek V4 Pro | $0.44 | $0.87 | $12 |
| Gemini 3.5 Flash | $1.50 | $9 | $66 |
| Sonnet 5 (intro) | $2 | $10 | $80 |
| Sonnet 5 (standard) | $3 | $15 | $120 |
| Opus 4.8 | $5 | $25 | $200 |
| GPT-5.5 | $5 | $30 | $220 |
Per million tokens, all list rates. Monthly column is a 20M input, 4M output agent shape with no caching, rounded to the dollar. DeepSeek V4 Pro is $0.435/$0.87.
At standard pricing Sonnet 5 runs that month for $120, which is 40% under Opus 4.8 and 45% under GPT-5.5, for a model that ties one and beats the other depending on the benchmark. It is not the cheapest thing on the board; DeepSeek V4 Pro does the same shape for twelve dollars, and Gemini 3.5 Flash for sixty-six. Those are the models to reach for when the task is routine. Sonnet 5 is the one to reach for when you were about to pay Opus rates and would rather not.
When to make the move
If you are on Opus 4.8 for agentic or general work and not leaning on the very top of its coding range, Sonnet 5 is the obvious trade: near-tie on GDPval, 40% off at standard, 60% off until September, same 1M context, same tokenizer so no counting surprises. This is the clean win and the reason the model exists.
Sonnet 4.6 users have a clock on the decision. Migrating during the intro window is close to free and often a small saving, especially on prose. Migrating in September means paying the tokenizer tax with no discount to hide it, roughly 30% more on code-shaped traffic for a model that is better but not sticker-cheaper. That can absolutely be worth it for the capability jump; just make it a decision, not a surprise on the invoice.
And if your task is genuinely routine, none of this is your fight. DeepSeek V4 Pro and Gemini 3.5 Flash sit an order of magnitude below Sonnet 5 and will not notice the difference on a classification or extraction job. Drop your real token mix into a calculator before you commit to any of them, because the only number that matters is the one your own workload produces.
Sources
- - Claude Sonnet 5 announcement: anthropic.com
- - Anthropic pricing docs (intro/standard, cache, batch): platform.claude.com
- - Model reference (context, tokenizer): platform.claude.com/models
- - Claude Sonnet 5 system card (benchmarks): anthropic.com/system-card
- - OpenAI API pricing (GPT-5.5): developers.openai.com
- - Google Gemini API pricing (3.5 Flash): ai.google.dev
- - DeepSeek API pricing (V4 Pro): api-docs.deepseek.com
- - TokenCost pricing page: tokencost.app/pricing