Anthropic canceled a price increase before it happened. On August 11, 2026, the company announced that Claude Sonnet 5's introductory pricing — $2 per million input tokens and $10 per million output tokens — is now permanent. The original terms ran "through August 31," with an implied step up to standard Sonnet rates on September 1. That step is gone.
The announcement drew roughly 1.5 million views within hours. The replies were not a celebration. They were arithmetic.
TL;DR — what changed and what didn't
| Question | Direct answer |
|---|---|
| What's the new price? | Nothing is new — $2/M input, $10/M output, unchanged |
| What actually changed? | The expiry date. Introductory pricing became permanent |
| When was the increase scheduled? | September 1, 2026, after the "through August 31" window |
| When did Sonnet 5 launch? | June 2026, billed as Anthropic's "most agentic Sonnet yet" |
| Is it the cheapest frontier model? | No — GPT-5.6 Luna and DeepSeek V4 Flash undercut it on list price |
| Does cheap-per-token mean cheap-per-job? | No. Token consumption per task varies more than list price does |
| What should you do? | Benchmark cost per completed task on your own workload, not per-token rates |
What Anthropic actually said
The wording matters here, so it's worth quoting precisely. Anthropic's post read: "We're making Claude Sonnet 5's introductory pricing permanent. We launched Sonnet 5 in June at $2 per million input tokens and $10 per million output tokens through August 31, and that price will remain unchanged."
Two things follow from that sentence. First, this is not a price cut — nobody's bill goes down tomorrow. Second, it is a withdrawn increase, which for anyone who built a Q4 budget assuming standard Sonnet rates from September is a real, if quiet, saving.
Sonnet 5 launched at the end of June 2026 with a specific pitch: an agentic model that "makes plans, uses tools like browsers and terminals, and runs autonomously at a level that just a few months ago required larger and more expensive models." That framing is the whole story behind the pricing decision. Anthropic priced Sonnet 5 to displace Opus-class spend on long-running agent work, and dropping the September step-up keeps that displacement math intact.
Why the replies were skeptical
The top responses to the announcement did not argue that $2/$10 is expensive in the abstract. They argued that it's expensive relative to what's on the shelf next to it in August 2026:
- One widely-shared reply put GPT-5.6 Luna at $1.2, versus Sonnet 5's $2 input rate.
- Another claimed a competing model "benches SIGNIFICANTLY better" at $0.14 per million input tokens on its official API — the same figure DeepSeek has quoted for V4 Flash 0731.
- A third made the point that lands hardest for anyone running agents in production: "With the amount of tokens Sonnet 5 uses, this introduction pricing is still TOO expensive."
That last complaint is the one worth taking seriously, and it's not really about pricing at all.
The number that actually decides your bill
Per-token price is the denominator. Tokens consumed is the numerator. For agentic workloads, the numerator moves far more than the denominator does.
Artificial Analysis publishes a metric built exactly for this: cost per Intelligence Index task — the dollars it takes a given model to work through their evaluation suite end to end, reasoning tokens and all. A chart from that dataset circulating alongside the pricing announcement put GPT-5.6 Luna (max) at $0.05 per task against Claude Sonnet 5 (max) at $1.72, with Sonnet 5's total broken into four stacked components ($0.69, $0.28, $0.58, $0.15).
Hold that comparison at arm's length — it's a third-party benchmark on a third-party task mix, and neither model's score on that index is visible in the price bar alone. But the structural point survives any quibble about methodology: a 1.7x gap in list price produced a far larger gap in cost per completed task. That difference is token consumption, not pricing policy, and no permanent-rate announcement touches it.
This is the same trap we mapped in context window pricing decoded and in why AI companies want you using agents: agentic harnesses re-send context, spawn subagents, retry tool calls, and think out loud. Every one of those multiplies the numerator. If you are evaluating models on the pricing page, you are reading the least variable term in the equation.
| What you compare | Typical spread across frontier models | How much it moves your bill |
|---|---|---|
| Input price per million tokens | ~10-15x | Low — most agent spend is not raw input |
| Output price per million tokens | ~3-5x | Medium — reasoning-heavy models emit a lot |
| Cached input price | ~50-100x | High if your harness caches well |
| Tokens burned per completed task | Can exceed 20x | Highest — this is the term that dominates |
If you haven't instrumented the bottom row for your own workload, a permanent price is reassurance about a variable you weren't controlling anyway. Our guide to prompt caching for LLM cost optimization covers the highest-leverage fix, and Databricks' account of managing AI coding costs at scale shows what the measurement discipline looks like at enterprise volume.
The price war context
The other reply worth quoting came from an industry analyst: "But but but people said prices wouldn't come down and tokens would only get more expensive and now Anthropic and OpenAI are lowering prices of models — that's simply not possible."
They're right about the direction, and the last six weeks make the pattern clear:
- OpenAI cut rates on GPT-5.6 Luna and Terra — covered in our Luna and Terra price cut breakdown.
- Chinese labs kept the floor low, with DeepSeek's V4 Flash at $0.14/M input and a permanent 75% discount on V4-Pro back in May.
- Anthropic declined to raise Sonnet 5, which in a market moving this fast is functionally a cut against the counterfactual.
There's a caveat embedded in point 2 that applies directly to point 3. DeepSeek's "permanent" discount lasted until it didn't — by August 6, 2026, the company was warning developers of a significant price increase with no numbers attached. Permanent, in AI API pricing, means "until serving economics say otherwise." Anthropic's language is stronger than an extension with a fresh deadline, and that's genuinely useful for planning. It is not a contract.
What "permanent" is actually worth to you
Concretely, three things:
- Budget certainty through Q4. If you modeled September onward at standard Sonnet rates, you can reclaim that delta. Do the math on your last 30 days of Sonnet 5 spend before you assume it's trivial.
- No forced migration in August. Teams that had a "evaluate alternatives before September 1" ticket open can close it — or better, keep it open and run the evaluation anyway, because the cost-per-task question was never about the deadline.
- A stable denominator for your own benchmarking. A fixed rate makes cost-per-task measurements comparable across months, which matters more than the rate itself.
How to actually evaluate this, in four steps
- Pick three candidate models and one task suite drawn from your real workload — not a public benchmark. Our guide on how to read AI benchmarks explains why public suites mislead here.
- Run each to completion with your actual harness, tools, and system prompt. Log total input, cached input, and output tokens per task.
- Compute dollars per successfully completed task, counting failed and retried runs against the model that failed them. This is where verbose models lose the advantage their list price suggests.
- Re-run quarterly. Both the prices and the token-efficiency of these models moved substantially in the last two months alone.
Bottom line
Anthropic locked Claude Sonnet 5 at $2/$10 per million tokens indefinitely and canceled the September 1 increase. That's a real, if modest, win for anyone budgeting Q4 Claude spend, and it signals that competitive pressure from OpenAI's price cuts and Chinese open-weight labs is binding on Anthropic's pricing decisions.
It does not make Sonnet 5 the cheapest option, and the loudest criticism — that Sonnet 5's token consumption erases its list-price competitiveness — is a measurement question that the announcement doesn't answer. Run the cost-per-task numbers on your own workload. That's the figure that shows up on your invoice.
Related on explainx.ai
- OpenAI's GPT-5.6 Luna and Terra price cuts — the move Anthropic is responding to
- DeepSeek warns of a significant API price increase — what "permanent" pricing is worth in practice
- Context window pricing decoded — where long-context billing actually bites
- Prompt caching for LLM cost optimization — the highest-leverage lever on agentic spend
- Why AI companies want you using agents: token economics
- Databricks on managing AI coding costs at scale
- AI token pricing explained — the fundamentals, if per-token math is new to you
- Claude Code pricing guide
- How to read AI benchmarks — why cost-per-task beats leaderboard rank
- DeepSeek V4 Flash 0731 on ARC-AGI: cost per task
Primary sources: Anthropic's Claude account announcement on X, August 11, 2026 (12:33 AM UTC) · Claude Sonnet 5 launch announcement, June 30, 2026 · Artificial Analysis cost-per-Intelligence-Index-task figures as circulated August 11, 2026 · Anthropic API pricing
Accurate as of August 11, 2026. Competitor rates cited here are drawn from public developer posts and third-party benchmarks, not from those providers' official pricing pages — verify current rates directly before budgeting production workloads. Follow @explainx_ai for updates.