1 September 2026 was the date a long-announced price increase for Claude Sonnet 5, Anthropic’s general-purpose model, was due to take effect. It will not. The introductory rate of $2 per million input tokens and $10 per million output tokens becomes the standard price. The decision, confirmed on 10 August, spares teams building on the API a 50% jump that had been on the calendar.
A 50% increase, scheduled and then withdrawn
At launch, Claude Sonnet 5 shipped at what Anthropic called an introductory rate. The company had said it would move from $2 and $10 to $3 and $15 per million tokens on 1 September 2026. On 10 August, Anthropic reversed that timeline: the introductory rate becomes the reference price, with no end date. The September deadline therefore passes with no effect on anyone’s bill.
For a production workload, the gap was far from trivial. Moving from $2 to $3 on input and from $10 to $15 on output was exactly a 50% increase on every billed token, at constant volume.
The increase scheduled for 1 September 2026, from $2 and $10 to $3 and $15 per million tokens, will not happen.
The Claude Sonnet 5 price grid
The standard rate comes with the platform’s usual discounts. Batch processing halves the cost in exchange for asynchronous handling. Prompt caching sharply reduces the cost of the prompt segments reused from one request to the next.
| Billing mode | Input ($ / M tokens) | Output ($ / M tokens) |
|---|---|---|
| Standard (synchronous) | 2.00 | 10.00 |
| Batch processing (−50%) | 1.00 | 5.00 |
| Cache — read | 0.20 | — |
| Cache — write | 2.50 | — |
The context window reaches one million tokens. Output is capped at 128,000 tokens on the synchronous API, and up to 300,000 tokens in batch mode via a dedicated beta header.
What $2 and $10 really cover
The headline price does not tell the whole story. On output, a token is worth five times an input token: long generations are what weigh on the bill, not the size of the context sent. A simple estimate helps frame a budget before putting an assistant into production.
# Estimate the monthly cost of a coding assistant on Claude Sonnet 5
INPUT_PRICE = 2 / 1_000_000 # 2 USD per million input tokens
OUTPUT_PRICE = 10 / 1_000_000 # 10 USD per million output tokens
requests_per_day = 4000
input_tokens = 3000 # context + instruction
output_tokens = 600 # generated answer
cost_day = requests_per_day * (input_tokens * INPUT_PRICE + output_tokens * OUTPUT_PRICE)
print(round(cost_day * 30, 2), "USD/month") # ~ 1440.0, before cache discount
A signal about where inference prices are heading
Walking back an increase that had already been announced is not a neutral move. It lands in a market where several providers align their mid-tier models on similar rates, and where inference cost keeps falling as hardware and software optimisations advance. For teams, stability often matters more than a few cents per million tokens: a rate that does not move lets a budget be committed for twelve months with no revision clause.
What to remember
Claude Sonnet 5 stays at $2 per million input tokens and $10 on output, now with no increase on the horizon. Batch processing and prompt caching remain the two most effective levers for trimming a bill. Projects already costed on the introductory rate need no adjustment.
On my own projects, price predictability matters more than the price itself. I have watched coding-assistant budgets blow up not because the per-token rate was high, but because an increase landed in the middle of a quarter that was already committed. An introductory rate turned into a permanent one is good news mainly because it removes a variable when sizing a budget. The real discipline still stands: watch the output and turn on caching, that is where most of the bill is decided. — Simon Janvier
Pour aller plus loin
Primary source: the official Claude platform pricing page.
