Anthropic Claude Platform Release Notes — June 25 to July 15, 2026

Condensed from Anthropic's own Claude Platform release notes, scoped to the window since this KB's prior ingest (2026-06-24). Full changelog lives at the source URL; entries below are selected for relevance to frontier-model releases, agent/harness infrastructure, and platform economics.

Model releases and access

• June 30 — Claude Sonnet 5 launched (claude-sonnet-5): introductory pricing $2/$10 per Mtok through 2026-08-31 (standard $3/$15 after). Supports a 1M-token context window, 128k max output tokens, same tool/platform feature set as Sonnet 4.6 except Priority Tier (not available on Sonnet 5). Three migration-relevant behavior changes: (1) adaptive thinking is on by default and manual extended-thinking budgets are removed (400 error if attempted — this was already deprecated on Sonnet 4.6); (2) non-default sampling parameters (temperature, top_p, top_k) now return a 400 error; (3) new tokenizer produces ~30% more tokens for the same text than the prior tokenizer generation.
• July 1 — Claude Fable 5 and Claude Mythos 5 access restored, following a prior suspension (Anthropic published a public statement on the redeployment).
• Context: Fable 5 / Mythos 5 originally launched June 9 with 1M-token context, 128k max output, always-on adaptive thinking, and the same +30% tokenizer shift as Sonnet 5 later inherited; Fable 5 runs safety classifiers that can return stop_reason: "refusal" with no billing for zero-output refusals, and an opt-in fallbacks parameter re-routes refused requests to another (billed) model.

Agent / harness infrastructure

• July 2 — New agent-memory-2026-07-22 beta header changes memory-store listing behavior for managed agents: results return in a stable server-defined order (order_by/order params now ignored); depth accepts only 0, 1, or omission (other values → 400 error); path_prefix must end in / and matches whole path segments, not substrings. Page cursors issued without the header are invalidated under it. This header replaces managed-agents-2026-04-01 on memory-store endpoints; sending both returns a 400. All major SDKs (Python 0.116.0, TypeScript 0.110.0, Go 1.56.0, Java 2.48.0, Ruby 1.55.0, PHP 0.36.0, C# 12.35.0, CLI 1.16.0) now send the new header by default. The old header's behavior is scheduled to be retired July 22, 2026.
• July 10 — Dreams (research preview) — Anthropic's newer managed-agent surface — now supports Claude Fable 5 and Claude Sonnet 5.
• June 30 (pre-window carryover, still current): Managed Agents session event streams gained event deltas (opt-in, streams partial agent-message text before the full event); session listing gained backward pagination; sessions can override agent config (model/system-prompt/tools/MCP servers/skills) per-session without touching the base agent; vault credentials gained an injection_location setting (header/body/both).

Platform / API economics and admin

• July 14 — Admin API user management (beta, all Claude Enterprise orgs): list/lookup members by email, change roles, remove members, manage invites, manage groups and custom roles. Group/custom-role calls need the anthropic-beta: ce-user-management-2026-07-13 header; member/invite calls need no beta header.
• July 8 — API key expiration: keys can now be created with a preset duration, custom duration, or "Never"; keys with ≥7-day lifetime trigger an email reminder before expiry. Admin API exposes expires_at on key listings.
• June 26 — Rate-limit consolidation: Sonnet and Haiku rate limits now match Opus at every usage tier; usage tiers consolidated from more granular levels into three — Start, Build, Scale. Anthropic states no org receives lower limits than before.
• June 25 — Fast mode for Claude Opus 4.7 deprecated, removal scheduled July 24, 2026 (migrate to Opus 4.8 fast mode).

Limitations

• This is a curated subset of a long-running changelog (626 lines fetched); entries outside the agent/model/harness/economics focus (e.g., minor SDK version bumps, documentation-only clarifications) are omitted here by design, not by omission of the underlying source.
• "Institution" is trivially Anthropic (first-party), but no independent verification of the claimed 54%-type efficiency figures (as with GPT-5.6) is offered by Anthropic in these notes — this source is procedural/API-surface documentation, not a benchmark disclosure.

Why this matters for the KB

This is the highest-authority available source in this ingest batch (first-party, sourceType technical-report) and gives direct primary coverage of two AI LENS required chapters: "Frontier labs" (Sonnet 5 launch, Fable 5/Mythos 5 restoration) and, via the agent-memory beta header and Managed Agents feature stream, the content-focus.md "harness-agentic layer" lane this ingest was weighted toward. The Sonnet 5 tokenizer shift (+30% tokens for the same text) is also a datapoint worth cross-referencing against arXiv:2603.28576's token-pricing analysis (also ingested this pass) — a tokenizer change is a backdoor way effective per-word price can rise even when the quoted per-token price doesn't move.

Source: Claude Platform release notes, Anthropic, accessed 2026-07-20 (content current through 2026-07-15 entries).