Claude Sonnet 5 Still Gets a September API Price Hike: What to Budget

Correction — August 15, 2026: An earlier version of this article incorrectly said Anthropic had cancelled the September 1 API price increase. Anthropic's current pricing page, Sonnet 5 documentation, migration guide, and launch announcement still show $2/MTok input and $10/MTok output through August 31, 2026, followed by $3/MTok input and $15/MTok output starting September 1, 2026. The article has been corrected.

If an API budget assumes Claude Sonnet 5 will stay at $2/$10 per million tokens after August 31, update it.

Anthropic currently lists the $2 input / $10 output rate as introductory pricing through August 31, 2026. On September 1, 2026, the standard first-party Claude API rate is scheduled to become $3 per million input tokens and $15 per million output tokens.

That is a 50% increase in the per-token rate for the same Sonnet 5 token volume.

There is a second budgeting wrinkle: Sonnet 5 uses a newer tokenizer that Anthropic says produces roughly 30% more tokens for the same text than Sonnet 4.6, with the exact difference depending on the content and workload. So migration costs should be measured from real Sonnet 5 token counts rather than copied from a Sonnet 4.6 bill.

Pricing check — August 15, 2026: Anthropic's live Claude Platform pricing documentation and current Sonnet 5 documentation both still show the September 1 increase.

The current Sonnet 5 price schedule

The September change is a 50% increase because both token rates are multiplied by 1.5:

• $2 → $3 input
• $10 → $15 output

This is an API token-pricing change. It is not a claim that Claude consumer subscription prices are increasing by 50%.

What the same Sonnet 5 workload costs before and after September 1

For ordinary first-party Claude API usage, before caching or other modifiers:

cost = (input MTok × input rate) + (output MTok × output rate)

Those examples assume the same Sonnet 5 token counts on both sides of September 1. Real bills can differ because caching, batching, data residency, retries, tools, agent loops, output length and other workload behavior also matter.

The tokenizer catch matters when comparing Sonnet 5 with Sonnet 4.6

Anthropic says Claude Sonnet 5 uses a newer tokenizer that produces approximately 30% more tokens for the same text than Sonnet 4.6. Its migration guidance explicitly recommends recounting prompts instead of reusing old token budgets.

That creates two separate comparisons that should not be mixed together.

Sonnet 5 in August vs. Sonnet 5 in September

The tokenizer is the same model tokenizer on both sides. If the application uses the same number of tokens, the September bill is simply 50% higher at the scheduled rate.

Sonnet 4.6 vs. Sonnet 5 after September 1

The listed base rates are both $3/$15, but equivalent text may tokenize into more tokens on Sonnet 5. That means "same price per token" does not necessarily mean "same price for the same text or task."

For example, if a particular input becomes 30% more tokens under Sonnet 5 and everything else stays equal, that part of the request would cost about 30% more at the same per-token rate. Real output behavior may differ too, so this is only an illustration.

The practical metric is therefore:

Measure dollars per successful production task, not only dollars per million tokens.

Prompt caching also becomes more expensive on September 1

Anthropic prices prompt caching as a multiplier of the base input-token rate. Its current pricing table shows these Sonnet 5 amounts:

Caching can still substantially reduce the cost of repeatedly processing a large stable prefix, but the cached-token rates rise with the scheduled base-price change.

Batch processing still cuts the token rate in half

Anthropic's Batch API applies a 50% discount to input and output tokens.

For Sonnet 5, the current published batch schedule is:

For non-interactive workloads such as offline classification, evaluation runs, document processing or nightly enrichment, batch economics may matter more than the headline standard rate.

US-only inference adds a 1.1× multiplier

Anthropic's current pricing documentation says Claude 4.6 and later models incur a 1.1× pricing multiplier when requests use US-only inference via inference_geo: "us". Global routing remains the standard-price default.

For Sonnet 5 that means, before other applicable modifiers:

Anthropic says the residency multiplier also applies to cache writes and cache reads.

Partner-operated cloud pricing needs its own check

Sonnet 5 is available through several platforms, but Anthropic's first-party API table should not be blindly copied into every cloud budget.

Anthropic's pricing documentation says partner-operated platforms such as Amazon Bedrock and Google Cloud have independent regional/platform pricing. If the bill comes from a partner platform, check that provider's current rate card and any private offer before finalizing a budget.

A better migration budget than "tokens × list price"

If a team is moving from Sonnet 4.6 to Sonnet 5, use a production-shaped sample rather than extrapolating from one prompt.

1. Capture representative requests

Take a privacy-approved sample that reflects the actual mix of short prompts, long context, tool calls, document inputs and agentic workflows.

2. Recount with the Sonnet 5 tokenizer

Do not reuse Sonnet 4.6 token counts. The tokenizer changed, and Anthropic explicitly warns that the same text can produce more tokens.

3. Include the full request lifecycle

Measure retries, tool-result context, agent loops, cache writes, cache hits and output tokens—not only the initial prompt.

4. Price both sides of September 1

For the same measured Sonnet 5 workload, calculate the bill once at $2/$10 and again at $3/$15. That gives a clean view of the scheduled rate change.

5. Divide by successful tasks

A useful application metric is:

cost per accepted task = total API cost / tasks that meet your acceptance criteria

For coding, that could mean a patch that passes tests. For extraction, a row that passes validation. For support automation, a response that does not require human correction.

Should a Sonnet 4.6 application move purely because Sonnet 5 is cheaper today?

Probably not purely for the temporary sticker price.

Through August 31, Sonnet 5 has a lower per-token introductory rate than Sonnet 4.6. Starting September 1, the published standard rates are the same $3/$15. At that point, the newer tokenizer means token volume can be higher for equivalent text, while capability or workflow efficiency may offset some of that in practice.

There are also migration behavior changes to test. Anthropic's Sonnet 5 documentation says adaptive thinking is on by default, manual extended-thinking budgets are not supported, and non-default temperature, top_p or top_k settings return errors.

What to watch before September 1

Pricing pages can change, so recheck the live Anthropic rate card shortly before the date rather than treating this article as a permanent contract.

As of August 15, however, the official documentation is consistent: the introductory $2/$10 Sonnet 5 rate runs through August 31, and $3/$15 is scheduled to begin September 1, 2026.

The other useful things to watch are actual Sonnet 5 token counts, output behavior, cache-hit rates, batch eligibility and whether the model completes tasks with fewer retries or agent steps.

Conclusion

Claude Sonnet 5's September API price increase is still scheduled.

The first-party rate remains $2/MTok input and $10/MTok output through August 31, 2026, then moves to $3/MTok input and $15/MTok output on September 1 according to Anthropic's current documentation.

For the same Sonnet 5 token volume, that is a 50% increase. For a migration from Sonnet 4.6, the newer tokenizer adds another reason to benchmark real production traffic rather than comparing list prices alone.

Update the rate card, recount tokens, and measure cost per successful task before locking the September budget.

Sources

Checked August 15, 2026:

• Anthropic — Claude Platform pricing
• Anthropic — What's new in Claude Sonnet 5
• Anthropic — Claude model migration guide
• Anthropic — Introducing Claude Sonnet 5
• Anthropic — Prompt caching