claude-sonnet-4-5-20250929 gets its November 30 retirement notice — two months out, migrate to Sonnet 5.5 now. Claude Code v2.1.286 drops with three separate log-redaction security patches (zero-width-space credential leaks, partial Bearer token masking, MCP error messages exposing secrets) plus a gateway billing bug that was charging cache writes at the wrong rate. Separately, Gemini CLI quietly gains autonomous plan execution in non-interactive mode, which is the capability gate for using it unattended in CI/CD pipelines. No new model launches today; the action is in hardening and cleanup.claude-sonnet-4-5-20250929) retires November 30 — if you haven't migrated to Sonnet 5.5, you have 61 days; read the five breaking changes in the September 28 digest firstgit clone https://github.com/tobi/disktree && cd disktree
cargo build --release
./target/release/disktree ~
claude-sonnet-4-5-20250929 — the September 2025 snapshot of Claude Sonnet 4-5. After this date, API calls to this model ID will return errors.claude-sonnet-4-5-20250929 retires November 30, 2026 — 61 days from today; Anthropic's recommended replacement is claude-sonnet-5-5, which launched September 28 with five breaking API changes (see September 28 digest).claude-sonnet-4-5-20250929 directly (as a pinned snapshot ID, not the rolling alias), the model stops responding on November 30. Migration path: target claude-sonnet-5-5. However, before you do: the September 28 digest documented five breaking API changes between Sonnet 5 and Sonnet 5.5 — thinking: disabled, tool_choice: any/tool, computer_20251124 toolset, old-generation advisor pairings, and conversation-bound thinking blocks. Run through that checklist before swapping the model ID. If you need more runway, the older claude-sonnet-5 is not yet deprecated and remains available. Also check personal/sources/ai-dev-sources.md model deprecations page for the full schedule; further 4.x retirements are likely to follow within the next 30–60 days.claude-sonnet-4-5-20250929 in production code, SDKs, or CI pipelines; you have hardcoded model IDs (not rolling aliases) that haven't been audited since Sonnet 5.5 launched.EffortModerate — the model ID swap is Quick, but the Sonnet 5.5 migration checklist (5 potential breaking changes) requires code review.No new model releases today.
No new API or SDK releases today.
arXiv.org was egress-blocked; full paper text for cs.AI/cs.CL submissions could not be read. See Near-misses for details. One candidate worth tracking: EngiWorld: What Can Frontier Agents Deliver in Professional Engineering Environments? (arXiv:2609.37686) — see Trends section.
%2B, %2F) were only partially obscured in logs — the unencoded portion after the last encoded character was visible. Affects anyone using MCP servers with OAuth-style auth tokens.Bearer <key>). Fixed: error messages now sanitize these patterns.Additionally: the Claude apps gateway was pricing 1-hour prompt cache writes at the cheaper 5-minute rate, resulting in underbilling. This is now corrected — gateway spend meters will show higher costs for cache writes going forward. Review your recent gateway billing if cost predictability matters.
New features: VSCode extension gains a bookmarks feature to save and revisit Claude's responses within a conversation; Claude Tag (Slack) can now set spend limits directly from a channel ID or Slack link.
--no-interactive or piped stdin) would stall on plan execution steps that normally require user confirmation. This blocked use cases like: running Gemini CLI as a CI step, scripting multi-step agent tasks, or integrating it as a backend tool in orchestration pipelines. The autonomous plan execution change removes this gate — Gemini CLI will now proceed through plan steps without prompting when invoked non-interactively. The companion fix (folder trust state propagation in headless mode) is equally important: without it, agents running in a subdirectory outside the initially-trusted root would hit trust failures mid-run. Together these two changes make Gemini CLI a viable tool for unattended agent automation. Note: this is a nightly (pre-release) build — use in CI should be pinned to a stable tag once one ships with these changes. The A2A server V1→V2 settings migration is relevant for teams using the experimental agent-to-agent protocol integration.cargo build --release, then run disktree <path>. The interactive treemap lets you navigate into subdirectories, see individual large files within their directory context (an improvement over ncdu which requires drilling down), and mark/delete directly without leaving the UI. Free space is always visible. If you're regularly cleaning up after agent sessions or managing large model/artifact caches, this is faster than the du -sh | sort -rh | head + manual drill pattern.No confirmed new leaderboard entries or SOTA changes within the 24-hour scan window. lmarena.ai was egress-blocked. The September 2026 Arena snapshot (claude-opus-5-max at ~1505–1531 ELO; highest open-weight model kimi-k3-max at ~1476) was not updated for September 30 specifically.
Nothing cleared the quality bar this period. HN threads on Claude Code v2.1.286 security fixes were not yet at score >200 at time of scan. simonwillison.net egress-blocked.
Announced at DevDay September 29 but not yet available in the API as of September 30. Watch platform.openai.com/docs/changelog for release. [openai.com/index/devday-2026-recap/]
Limited preview only as of September 30; no public endpoint, model ID, or price published yet. 10× faster than GPT-6 Luna standard API (~150ms latency). [openai.com/index/devday-2026-recap/]
No model card, pricing, or date as of September 30. The Opus 5.5 → Sonnet 5.5 gap was 6 days; Haiku 5.5 timeline remains unconfirmed. [Anthropic Sonnet 5.5 release, September 28, 2026]
Free billing period ends October 4. Standard pricing applies from October 5. [platform.openai.com/docs/changelog]
Migrate to antigravity-preview-09-2026; rename snake_case built-in tool parameters to PascalCase. [ai.google.dev/gemini-api/docs/changelog]
Filtered from 30+ primary sources against a published quality rubric. No press releases, no fluff — only what changes what you build.