🚨 Breaking
- Claude Sonnet 4 and Claude Opus 4 retire June 15, 2026. Requests to
claude-sonnet-4-20250514andclaude-opus-4-20250514will error after that date. Migrate to Claude Sonnet 4.6 or Claude Opus 4.8 immediately. Source - Claude Opus 4.1 deprecated – retired on August 5, 2026. Model ID
claude-opus-4-1-20250805. Migrate to Claude Opus 4.8. Source - Claude Opus 4.7 includes API breaking changes vs Opus 4.6. Check migration guide.
- Claude Haiku 3 (
claude-3-haiku-20240307) already retired – requests return errors. Source - 1M-context beta retired for Sonnet 4.5/4 – beta header has no effect; requests above 200k fail. Use Sonnet 4.6 or Opus 4.6+ for 1M context. Source
🗑️ Dépréciations
- Claude Opus 4.1 – deprecation announced June 5, 2026; retirement August 5, 2026. Source
💰 Pricing
No pricing changes reported.
🆕 Nouveautés
- Claude Opus 4.8 – new most capable model with 1M context, 128k output tokens, same tools as Opus 4.7. Available on API, Bedrock, Vertex, Foundry (200k). Source
- Claude Opus 4.7 – launched April 16; includes breaking changes vs 4.6. Source
- SGLang v0.5.13 – adds support for Nemotron 3 Ultra, Step-3.7-Flash, Command A+; new diffusion models. Source
- vLLM v0.23.0 – DeepSeek-V4 matures across backends, 408 commits, 200 contributors. Source
- llama.cpp b9616 – CI fix for release. Source
🌐 Actualité IA
- Recoverable Visual Token Routing for VLMs (Hugging Face paper 2606.12412) – proposes rerouting instead of removing visual tokens to reduce KV-cache memory without losing information. Source
💡 Conseil du jour
Immediate action: Migrate all calls using claude-sonnet-4-20250514 and claude-opus-4-20250514 before June 15, 2026. Update model IDs to claude-sonnet-4-6 and claude-opus-4-8. Also start testing Opus 4.8 or 4.7 for Opus 4.1 workloads to meet the August 5 retirement. Pin model versions in your config and deploy with a canary to catch any subtle output differences.