← Back to all deprecations

Migrating off Claude Sonnet 4

Claude Sonnet 4 (Anthropic) — no announced shutdown date, but superseded. Replaced by Claude Sonnet 4.6. Source · Verified 2026-08-14.

Claude Sonnet 4 has no announced Anthropic shutdown date but is directly superseded by Sonnet 4.6, a strict upgrade at the same position in Anthropic's lineup. This is the cleanest 1:1 migration on this page — same tier, same provider, same positioning.

What should I use instead of Claude Sonnet 4?

Claude Sonnet 4, from Anthropic, has no announced shutdown date but is superseded. The recommended successor is Claude Sonnet 4.6, priced 0% more at blended (3:1) rates — swap the model id in your API calls and test side by side before cutting over production traffic.

Verified 2026-08-14 — source

The swap

// All AI Ask API
{
-  "model": "claude-sonnet-4",
+  "model": "claude-sonnet-4-6",
   "messages": [...]
}

Same change applies if you call Anthropic directly with its own SDK — only the model identifier changes; request/response shape is otherwise unaffected for a same-provider swap.

What actually changes

Claude Sonnet 4Claude Sonnet 4.6
Blended $/M$6.00$6.00
Context window200,000 tokens300,000 tokens
Max output32,000 tokens64,000 tokens
Modalitiestext, visiontext, vision
Tokens / secunchanged / not measured76

Pricing from our live pricing registry; speed from our benchmark leaderboard where measured.

Gotchas

  • Sonnet 4.6's context window is 50% larger (300K vs 200K tokens). If you were chunking long documents specifically to stay under Sonnet 4's 200K limit, that chunking logic can likely be relaxed, not just left in place as a safety margin.
  • Extended-thinking (reasoning) token accounting differs slightly between the two model generations — if you set a hard `max_tokens` budget that included thinking tokens for Sonnet 4, re-verify it against Sonnet 4.6's documented token accounting rather than assuming it transfers unchanged.
  • Sonnet 4.6 is priced the same or lower than Sonnet 4 depending on your input:output mix — this is a rare case where the upgrade doesn't cost you more, so there's little reason to delay the switch.

Alternatives beyond the obvious successor

GPT-5.6 Terra$5.63/M blended

OpenAI's equivalent "balanced" tier, if you want a cross-provider comparison before committing.

Gemini 3.6 Flash$3.00/M blended

Faster and larger context window than either Sonnet generation, at a lower price.

GLM-5.2$2.15/M blended

Open-weights alternative with a 1M-token context window at a fraction of the cost, if coding is the primary workload.

Not tied to Anthropic? See cross-vendor alternatives to Claude Sonnet 4.6.

FAQ

Is Claude Sonnet 4 being shut down?

Claude Sonnet 4 has no announced shutdown date, but it is superseded and no longer the recommended choice. Source: https://docs.anthropic.com/en/docs/about-claude/model-deprecations.

What should I switch to from Claude Sonnet 4?

Claude Sonnet 4.6 is the direct successor. See "The swap" below for the exact model id change.

Will switching cost more or less?

The successor is 0% more expensive at blended (3:1) pricing. See the comparison table for exact per-million-token numbers.

Decision and evidence guide. Verified 2026-08-14. These are dated, route-local references; unavailable values are not inferred.

Anthropic native vs cloud-hosted Sonnet 4 lifecycle isolation board

Frozen fixture board. Formula / decision rule: resolved = provider surface (anthropic.com / Bedrock / Vertex) + exact model version + notice + shutdown evidence Boundary: Anthropic API retirement does not automatically trigger Bedrock or Vertex retirement.

Frozen fixtureJoined inputs and observationCalculated resultState
claude-sonnet-4-20250514 on anthropic.com APIhost=api.anthropic.com; model=claude-sonnet-4-20250514; notice=docs.anthropic.com/models; shutdown=Unavailable specific date per public docs 2026-08-14
Anthropic dates model versions by release snapshot (YYYYMMDD); check the models page for current status.
shutdown=check Anthropic models page; do not invent dateUNAVAILABLE — check models page.
anthropic.claude-sonnet-4 on Amazon Bedrockhost=bedrock; model-id=anthropic.claude-sonnet-4-20250514-v1:0; aws-notice=separate; own-retirement=aws retirement date
Bedrock model retirement is announced by AWS separately from Anthropic native API notices.
check AWS documentation for Bedrock retirement timelinePASS WITH SEPARATION — Bedrock-local.
claude-sonnet-4@001 on Vertex AIhost=vertex-ai; version=@001; google-notice=separate; own-retirement=google retirement date
Vertex AI publisher model deprecation is controlled by Google Cloud.
check Google Cloud release notes for Vertex Claude deprecationPASS WITH SEPARATION — Vertex-local.

Provenance: claude-sonnet-4 module 1 first-party evidence, surface verification date 2026-08-14. Anthropic model deprecation notices. Missing joins fail closed.

Extended thinking mode migration compatibility receipt

Frozen fixture board. Formula / decision rule: compatible = new model supports extended thinking + budget_tokens parameter + response streaming format unchanged Boundary: Extended thinking block format and streaming behavior must be verified on each successor.

Frozen fixtureJoined inputs and observationCalculated resultState
claude-sonnet-4 extended thinking to claude-sonnet-5old=claude-sonnet-4-20250514; thinking=extended; param=budget_tokens; successor=claude-sonnet-5; thinking support=documented; format=same
Claude Sonnet 5 documents extended thinking compatibility; format is preserved.
migration=compatible for thinking workloads; verify budget_tokens behaviorCOMPATIBLE — verify budget behavior.
Extended thinking + tool use combinedold=claude-sonnet-4; tools=parallel; thinking=before tools; successor=claude-sonnet-5; tool+thinking order=verify
Tool use during extended thinking has specific ordering; verify on successor.
test tool+thinking combination before promotingVERIFY — tool+thinking ordering.
Non-thinking production workload to claude-haiku-3-5 cost reductionold=claude-sonnet-4; thinking=disabled; task=summarisation; haiku class=lower; quality delta=required
For non-reasoning tasks, Haiku may qualify with quality testing, significantly reducing cost.
quality gate required; cost delta = check /llm-api-pricingCONDITIONAL — quality gate required.

Provenance: claude-sonnet-4 module 2 first-party evidence, surface verification date 2026-08-14. Anthropic model deprecation notices. Missing joins fail closed.

Computer use and tool use parity receipt for successor models

Frozen fixture board. Formula / decision rule: parity = computer_use_supported + same tool schema + same beta header + identical result format Boundary: Computer use is a beta feature; parity cannot be assumed across model versions without explicit documentation.

Frozen fixtureJoined inputs and observationCalculated resultState
Computer use beta · claude-sonnet-4 to claude-sonnet-5old=claude-sonnet-4; computer-use=beta header; successor=claude-sonnet-5; header=anthropic-beta: computer-use-2025-01-22; parity=documented
Claude Sonnet 5 inherits computer use support with the same beta header.
computer use migration=compatible; re-test automation scenariosCOMPATIBLE — re-test required.
Tool use with parallel calls · JSON schema strict modeold=claude-sonnet-4; tools=parallel; schema=strict; successor=claude-sonnet-5; parallel=documented; strict=documented
Parallel tool calls and strict JSON schema are documented on Claude Sonnet 5.
parity=documented; run schema regression before promotingCOMPATIBLE — regression test.
Custom tool using claude-sonnet-4 only capabilityold=claude-sonnet-4; tool=custom capability; successor=check docs; parity=Unavailable until verified
Custom tools requiring undocumented capabilities need explicit verification on the successor.
parity=Unavailable; document specific capability requirement and verifyUNAVAILABLE — capability join required.

Provenance: claude-sonnet-4 module 3 first-party evidence, surface verification date 2026-08-14. Anthropic model deprecation notices. Missing joins fail closed.

Run this scenario →

Run the old prompt against both models, side by side

The only real way to know if a migration is safe is to test it. One workspace, both models, your actual prompts.

Try It Free