Claude Sonnet 5 Pricing: Your $2/$10 Window Closes Aug 31
Claude Sonnet 5's introductory API pricing of $2 per million input tokens and $10 per million output ends August 31, 2026, stepping up 50% to the standard $3/$15 on September 1. That intro window is the only period the model is both cheaper and stronger than Sonnet 4.6, which already costs $3/$15. Prompt caching and the Batch API stack with the discount, pulling effective input rates as low as $0.20 per million. For PMs, the move this quarter is to run migrations and evals at intro rates while modeling all unit economics at the standard price.
If I were running a product on Sonnet 4.6 right now, here's the play. This week, stand up a Sonnet 5 eval branch and push your full regression suite through the Batch API at intro rates — $1 per million input, $5 per million output means the entire exercise costs half of what on-demand would, on top of a rate card that itself disappears August 31. You're effectively being paid to de-risk your own migration. Second, rebuild your cost model twice — once at $2/$10, once at $3/$15 — and socialize the $3/$15 number internally. The most common failure mode with promo pricing is a team that anchors its margins to the intro rate, ships customer pricing on that basis, and eats the step-up in Q4. Third, if your prompts share long static prefixes, implement prompt caching now: at standard rates the 90% read discount is what keeps September's increase from being a true 50%, and cached reads not counting against input rate limits is free throughput on the Claude API. To be fair, a $300-a-month delta on a mid-sized workload won't sink anyone — the danger isn't the invoice, it's the mispriced commitments downstream of it. You have until August 31, and every week of delay is discounted eval budget you don't get back.
Claude Sonnet 5's intro pricing of $2/$10 per million tokens ends August 31, 2026; standard $3/$15 rates — a 50% step-up on both sides — take effect September 1.
You have 48 days. Claude Sonnet 5's introductory pricing — $2 per million input tokens, $10 per million output, per Anthropic's pricing docs — ends August 31, 2026. On September 1 it steps up to the standard $3/$15, a 50% increase on both sides of the meter, announced in advance with a hard date attached.
My read: this is the cheapest frontier-class capacity you'll see all year, and most teams will waste the window by treating it as a curiosity instead of a budget line.
Here's what you need to know, and what the arithmetic actually says.
The deadline is the story, not the price
Sonnet 5 shipped June 30 on the Claude API, AWS Bedrock, and Google Vertex, and the $2/$10 intro rate has a published stop date: August 31. From September 1 you pay $3/$15 — which, not coincidentally, is exactly what Claude Sonnet 4.6 costs today. That makes the intro window the only stretch where Sonnet 5 is simultaneously cheaper than its predecessor and better than it: 85.2% on SWE-bench Verified against 4.6's 79.6%, as of July 2026 per the llm-stats board (self-reported scores, worth remembering).
After the deadline you're paying 4.6 money for a stronger model. Still a fine trade — just no longer a discount.
Why hand out a 33% cut on your newest workhorse seven weeks before quietly taking it back? The hub piece makes the case that this is IPO-timed aggression, and I won't re-argue it here. What matters for you is the mechanism. Intro pricing is built to get Sonnet 5 into your production traffic before you've had time to deliberate.
It's the streaming-service playbook: hook you at the promo rate and bet that once you're wired in, nobody churns over a step-up.
That bet usually pays, because migration friction is real — which is exactly why you should do the migration work now, while the discount subsidizes your evals. The Sonnet 4.6 migration guide covers the mechanics; the short version is that every eval run you do in July costs a third less on input than the same run in September.
What the step-up does to a real bill
Take a plausible production workload: 200M input tokens and 20M output tokens a month — a mid-sized agent feature or a busy copilot.
At intro rates, 200 × $2 plus 20 × $10 comes to $600 a month. At standard rates, 200 × $3 plus 20 × $15 comes to $900 a month. Same traffic, $300 more per month — $3,600 a year — the moment the calendar flips.
Run your own token mix through the cost calculator at both rate cards, and use the token counter if you need to price individual prompts first. Whatever number falls out at standard rates is the one that belongs in your Q4 budget. The intro figure is a subsidy with an end date, not a baseline.
Caching and Batch stack with the discount
The step-up looks less brutal once you layer in Anthropic's two cost levers, both of which apply during and after the intro window.
| Rate card | Input / 1M | Output / 1M |
|---|---|---|
| Intro (through Aug 31) | $2.00 | $10.00 |
| Standard (from Sep 1) | $3.00 | $15.00 |
| Cache read (intro / standard) | as low as $0.20 / $0.30 | — |
| Cache write (intro / standard) | $2.50 / $3.75 | — |
| Batch (intro / standard) | $1.00 / $1.50 | $5.00 / $7.50 |
Prompt caching prices cache reads up to 90% below base input, with writes billed at 1.25× base. And per Anthropic's docs, cached reads don't count toward input-token rate limits on the Claude API — a documented advantage neither OpenAI nor Google currently matches, and one that matters as much for throughput as for cost. If 150M of that 200M-input example lands as cache reads at the full discount, your September input bill drops from $600 to about $195.
The Batch API takes 50% off both input and output for anything that can wait, and with the output-300k-2026-03-24 beta header it returns up to 300K output tokens per request against the standard 128K. Batch at intro rates — $1 in, $5 out — is the cheapest Sonnet 5 gets. If you're sitting on a backlog of offline work — codegen sweeps, document processing, eval suites — the next seven weeks are the time to burn through it.
Against Terra and Gemini, the clock cuts both ways
GPT-5.6 Terra, GA since July 9, runs $2.50/$15 per OpenAI's pricing page. Gemini 3.1 Pro is $2/$12, rising to $4/$18 above 200K input tokens, per Google. At intro rates Sonnet 5 undercuts both on output — $10 against $15 and $12 — while matching Gemini on input.
At standard rates it becomes the most expensive input of the three. A model decision you make in July can flip on you by October, so run the comparison at both prices before you commit.
To be fair to Anthropic: publishing the deadline up front is more honest than the industry norm. There's no silent repricing here, no usage-tier fine print — you know today, to the dollar, what you'll pay in Q4. And if your question is whether to spend more rather than less, Opus 4.8 at $5/$25 buys +3.4 SWE-bench points for 2.5× Sonnet's intro price; the Opus comparison runs that trade in full.
The harder question isn't whether $2/$10 is a good deal — it obviously is. It's whether anything you commit to in the next seven weeks — your product's pricing, your margin model, the unit economics you show your board — is quietly anchored to a number that stops existing on September 1.
Watch for: whether Anthropic extends the intro window past August 31. A quiet extension would tell you adoption came in under plan. A firm cutoff tells you the discount already did its job.
Get this in your inbox. AI Rundown Daily delivers original briefings every morning — free. Subscribe →
Frequently Asked Questions
It's deliberate strategy — the launch discount is widely read as IPO-timed, a case our hub article makes in full. But the money is real: a workload of 200M input and 20M output tokens a month costs $600 at intro rates versus $900 at standard, per Anthropic's published pricing. Gimmick or not, evals and Batch backlogs run through the window cost genuinely less. The unusually honest part is that Anthropic printed the August 31 deadline up front instead of repricing quietly later.
Both levers apply under the intro and standard rate cards. Prompt caching bills cache reads up to 90% below base input — as low as $0.20 per million during the intro window, $0.30 after — with writes at 1.25× base, and on the Claude API cached reads don't count toward input-token rate limits. The Batch API takes 50% off both input and output, so intro-rate batch work runs at $1/$5 per million, and the output-300k-2026-03-24 beta header lifts batch output to 300K tokens per request. Stacked correctly, your effective September increase lands well under the headline 50%.
Anchoring your own economics to a temporary number. On September 1 the same traffic costs 50% more per token, so any customer pricing, margin model, or board projection built on $2/$10 quietly breaks. There's also a competitive wrinkle: at standard rates Sonnet 5 carries the most expensive input of its peer group — $3 versus GPT-5.6 Terra's $2.50 and Gemini 3.1 Pro's $2, as of July 2026 — so a model comparison run in July should be re-run at standard prices before you commit.
