ApiDelta — May 7, 2026

Anthropic expanded 1 M‑token context windows to GA models and added a rate‑limits endpoint, while several older Claude models are now retired; Mistral released Vibe 2.0 and updated its Python SDK.

1 M‑token context now generally available for Claude Opus 4.6/4.7 and Sonnet 4.6

Requests up to 1 million tokens no longer require a beta header and are billed at the standard $5 / $25 per MTok rate. Calls that exceed 200 k tokens will succeed automatically on these models, simplifying prompt design for long‑form tasks. Source

Claude Opus 4.7 launched as the top‑tier GA model

Opus 4.7 is now the default “most capable” model for complex reasoning and agentic coding, priced identically to Opus 4.6. The release includes a new tokenizer and documented capability improvements; existing Opus 4.6 calls can be switched to 4.7 without code changes. Source

Rate Limits API added for programmatic quota inspection

Admins can query organization‑ and workspace‑level rate limits via a new endpoint, enabling automated monitoring and dynamic throttling in production pipelines. Source

Three Claude models officially retired

- claude-3-5-haiku-20241022 (retired Dec 2025, removal Feb 19 2026) - claude-3-7-sonnet-20250219 (retired Oct 2025, removal Feb 19 2026) - claude-3-haiku-20240307 (retired Feb 19 2026, removal Apr 20 2026)

Calls to these identifiers now return errors; migrate to the latest Opus or Sonnet versions. Source

Mistral Vibe 2.0 released, API now paid for Devstral 2.0

Vibe 2.0 introduces updated document annotation capabilities and removes the previous 8‑page limit. Access to Devstral 2.0 is shifted to a paid tier, so developers must update billing settings to continue using it. Source

Mistral Python SDK v2.4.3 adds RAG registration and removes OCR request ID

- Breaking: mistral.ocr.process().request.id removed. - New: mistral.beta.rag.ingestionpipelineconfigurations.register() added for custom RAG pipelines.

Upgrade the SDK and adjust any OCR handling code accordingly. Source