Back to sources

AI API

Anthropic change intelligence

Claude model, tool use, API, rate limit, SDK, and capability updates that affect AI product behavior.

Monitor cadence

Daily

Primary audience

AI app developers and platform teams

Official changelog

Watched topics

ModelsTool useRate limitsSDKsMigration

Latest reviewed changes

Read from the official changelog and annotated with impact and deadlines. Last checked 2026-07-16.

Watch this source
HighDeprecation

Claude Opus 4.1 deprecated, retires August 5, 2026; sampler params return 400 on new models

claude-opus-4-1-20250805 was deprecated June 5, 2026 and retires August 5, 2026 (replacement: claude-opus-4-8); claude-mythos-preview retired June 30, 2026 in favor of claude-mythos-5. Separately, temperature/top_p/top_k return a 400 error when set to non-default values on Opus 4.7+, Opus 4.8, and Sonnet 5.

Deadline: 2026-08-05

Action note

Audit API usage for claude-opus-4-1-20250805 and mythos-preview calls, migrate to claude-opus-4-8 / claude-mythos-5 before 2026-08-05, and strip non-default temperature/top_p/top_k from requests to newer models.

View original source
HighPricing

Claude Fable 5 / Mythos 5 use new tokenizer: ~30% more tokens for the same text

Claude Fable 5 and Mythos 5 use the tokenizer introduced with Opus 4.7, which produces roughly 30% more tokens than earlier models for the same text, and billing reflects the higher counts. Token counts measured on older models must not be reused; the free count_tokens endpoint returns counts for the model you pass.

Action note

Before migrating any workload to Fable 5 or Mythos 5, recount your prompts via /v1/messages/count_tokens with the target model and re-budget costs and context-window fit for the ~30% token increase.

View original source
MediumSchema/API

Rate Limits Admin API: programmatically read org and workspace rate limits

New read-only Admin API endpoints (GET /v1/organizations/rate_limits and per-workspace variants) return configured limits per rate-limit group with values like requests_per_minute and input/output tokens per minute. Requires an Admin API key; limits can be read but not updated via the API.

Action note

Replace hardcoded rate-limit values in gateways and proxies with a scheduled fetch of /v1/organizations/rate_limits, and pair it with the Usage API for limit-utilization alerting.

View original source
MediumSchema/API

Usage & Cost Admin API adds geo, context-window, and fast-mode reporting dimensions

Admin-key endpoints /v1/organizations/usage_report/messages and /v1/organizations/cost_report report token usage (1m/1h/1d buckets, groupable by api_key, workspace, model, service_tier, context_window, inference_geo) and daily USD costs. Data lands within about 5 minutes; not available on Claude Platform on AWS.

Action note

Build cost-reconciliation and per-workspace chargeback jobs against usage_report/messages and cost_report with an Admin API key.

View original source