Three Claude API deprecations hit between April 19 and June 15, 2026. Here's the plain-English checklist: what's breaking, what to swap it to, and what you can ignore.
Last updated: Apr 18, 2026
Requests to claude-3-haiku-20240307 will start returning errors on Sunday. If you have this model ID anywhere in your codebase, CI/CD config, or infrastructure, swap it out before the weekend.
What to check: grep your repos for the old model string. Check Bedrock/Vertex configs. Run a staging test with the new model ID before Sunday — Haiku 4.5 is near-frontier intelligence at Haiku speed, but output formatting can shift subtly.
The context-1m-2025-08-07 beta header stops working for Sonnet 4 and Sonnet 4.5. Requests over 200k tokens will return an error after April 30.
Good news: Sonnet 4.6 and Opus 4.6+ have 1M context built in at standard pricing — no beta header needed. Just update the model ID and drop the header.
What to check: search your API calls for context-1m-2025-08-07. Every one of those call sites needs to swap model and drop the header. Test long-context retrieval quality after the swap — 1M capacity doesn't mean throw everything in. Relevance still matters.
Anthropic deprecated both models on April 14, 2026. They keep running until June 15 — plenty of runway, but don't wait until the last week. Start rolling production traffic over now.
Heads up on Opus 4.7: Anthropic notes Opus 4.7 may require prompt tweaks compared to Opus 4. It's a stronger model — better self-verification, better vision, better long-running agent behavior — but your existing Opus 4 prompts weren't necessarily tuned for it. Test before you ship.
claude-3-haiku-20240307, claude-sonnet-4-20250514, claude-opus-4-20250514, and context-1m-2025-08-07.betas: ["context-1m-2025-08-07"] config once everything is on Sonnet 4.6 or Opus 4.7.If you're only using Claude in the web app (claude.ai), Projects, Cowork, or the Desktop app: nothing. Anthropic manages model selection on your behalf in those surfaces. You're already on the latest models.
If you're using Claude Code: nothing. Claude Code picks models automatically based on the task — it's already migrated.
If you're on Enterprise with SSO + managed models: check with your admin, but most Enterprise deployments auto-migrate.
Deprecations are going to keep happening. Anthropic ships a new flagship model roughly every two months, and older models get retired on a ~6–12 month cycle. The fastest teams build one migration runbook once and reuse it each cycle. If this page helped, we'll keep it current — bookmark it.