Claude Sonnet 4 retires June 15, 2026: what breaks and how to migrate

The most widely deployed Claude of 2025 is now retired — its API IDs stopped responding on June 15, 2026. The good news: its replacement costs the same and beats it everywhere.

By the benchr team · · Verified against Anthropic's deprecation docs, June 16, 2026 · View changelog

Retirement date2026, Claude API
Input / 1M, both modelsprice unchanged
Sonnet 4.6 SWE-benchvs 72.7% on Sonnet 4
Context growth200K → 1M tokens

Who this hits

Sonnet 4 shipped May 2025 and spent close to a year as the default recommendation for production API work — cheap enough to run at volume, strong enough to trust with real tasks. That's exactly why this retirement matters more than most. If your team built anything on Claude between mid-2025 and early 2026, odds are decent the model string says claude-sonnet-4-20250514 or the claude-sonnet-4-0 alias.

Anthropic announced the deprecation on April 14, 2026 and emails affected accounts, but notification emails go to whoever owns the API key. The engineer who pinned that model ID may have left the company. Grep for it.

The swap, in numbers

Sonnet 4 vs Sonnet 4.6 (verified June 12, 2026)
SpecSonnet 4 (retiring)Sonnet 4.6 (replacement)
Input / 1M tokens$3.00$3.00
Output / 1M tokens$15.00$15.00
SWE-bench Verified72.7%79.6%
Context window200K1,000,000
Max output64K64K

This is the rare migration with no trade-off row. Same bill, better scores, five times the context. The only real work is verification, not negotiation.

Migration checklist

1. Find every reference. Search code, config files, environment variables, and infrastructure-as-code for claude-sonnet-4-20250514 and claude-sonnet-4-0. Don't forget prompt-management tools and eval harnesses.

2. Confirm what's still calling it. Anthropic's console exports usage as CSV, broken down by API key and model. If a deprecated ID shows traffic, trace the key.

3. Swap and re-run your evals. Sonnet 4.6 follows instructions more tightly than Sonnet 4. Prompts that compensated for the older model's quirks (forced formatting, repeated constraints) sometimes over-correct on the newer one. Run your test suite before and after.

4. Watch token counts for a week. Newer Claude models tokenize differently, and long-context behavior changes how you batch documents. Costs stay flat per token, but token counts per request can shift. The calculator helps you re-budget if volumes move.

5. Check partner platforms separately. The June 15 date covers the Claude API, AWS, and Microsoft Foundry. Amazon Bedrock and Vertex AI publish their own retirement schedules. Don't assume the same date.

Frequently asked

What happens to claude-sonnet-4 API calls after June 15, 2026?

They fail. The retirement covers the Claude API, Claude Platform on AWS, and Microsoft Foundry. Bedrock and Vertex AI run separate schedules; verify those in their own model tables.

Does Sonnet 4.6 cost more?

No. Both models bill $3/$15 per million tokens. The upgrade is free in pricing terms.

Is it a drop-in swap?

Mostly. Change the model string and most workloads run fine. Re-test formatting-sensitive prompts and anything that assumes a 200K context ceiling, since 4.6 accepts 1M tokens.

Changelog

  • — Status updated to retired. The June 15, 2026 shutdown has passed and was re-confirmed on Anthropic's deprecation table; calls to the old IDs now fail.
  • — Published. Dates and prices verified against platform.claude.com deprecation and pricing docs.

Sources

  • Anthropic model deprecations — platform.claude.com/docs/en/about-claude/model-deprecations (verified June 12, 2026)
  • Anthropic Claude 4 announcement (Sonnet 4 launch specs) — anthropic.com/news/claude-4
  • benchr deprecations dataset · model-figures.json