sora-2, sora-2-pro, or the Videos API is broken right now), and Anthropic expands billing to cover pre-output refusals in the bio, frontier_llm, and reasoning_extraction categories — a change that quietly recalculates the cost floor for any app built on Claude Fable or Opus models. On the tooling side, Cline Desktop reaches Linux today, closing the last major platform gap for the most widely used open-source coding agent harness. And a new benchmark, KEX-bench, puts numbers on something the security community has suspected: frontier coding agents solve 56% of Linux kernel exploit generation tasks without any reference proof-of-concept.sora-2, sora-2-pro, or OpenAI's Videos API endpoint, it is broken right now; migrate to an alternative video generation APIbio, frontier_llm, reasoning_extraction categories now billed at full model rates; update cost projections for Fable/Opus-based appsdocker run -d --name ai-memory -p 127.0.0.1:49374:49374 \
-v ai-memory-data:/data akitaonrails/ai-memory:latest
ai-memory install-mcp --client claude-code --apply
ai-memory install-hooks --agent claude-code --apply
sora-2, sora-2-pro, and any calls to the Videos API endpoint are no longer functional as of September 24, 2026.sora-2, sora-2-pro, or the OpenAI Videos API endpoint has been broken since today's UTC rollover; OpenAI's consumer-facing Sora app had already shut down April 26, 2026.sora-2, sora-2-pro, videos.generate, and any calls to https://api.openai.com/v1/video/* — these are all dead endpoints as of today. (2) Migration options for AI video generation: Google's Veo API (via Gemini API — veo-2-0826 and veo-3-0725 are the current stable versions), Runway ML API, and Pika API are the three most commonly cited alternatives. Veo 3 is closest to Sora's creative quality on narrative clips and has Google's API SLA backing it. (3) Sora's research team continues internally (world simulation for robotics), but OpenAI has no public timeline for any new video generation API. (4) If you archived any Sora outputs: the output files from prior calls are unchanged; only new generation requests fail. (5) No credits for unused Sora API prepayment have been announced; contact OpenAI support if you held a balance.bio, frontier_llm, reasoning_extraction) are now billed at the requested model's standard rates. Before today, any refusal that occurred before any output tokens were returned was not billed. The change applies to all platforms (API, Bedrock, Vertex).stop_reason: "refusal" with stop_details.category: "bio" (or "frontier_llm" or "reasoning_extraction") before producing any output now incurs the full input-token cost at the Fable 5.1 rate ($3/MTok).bio category covers requests related to bioweapons and CBRN materials; frontier_llm covers requests to extract training data, weights, or detailed training procedures from frontier models; reasoning_extraction covers attempts to elicit full internal reasoning chains in ways that circumvent safety measures. If your app serves researchers, enterprise users, or includes open-ended question fields where users could probe these areas, your refusal rate in these categories will now accrue costs. (2) Check your fallback setup. Anthropic's fallbacks: "default" SDK parameter (added in the fable-5-fallback beta) handles refused requests by automatically retrying on a fallback model; with today's billing change, a Fable 5.1 refusal now costs the Fable 5.1 input-token rate before the fallback model's charge is incurred. Review the SDK middleware docs to understand total cost impact. (3) Mid-stream refusals (stop_reason: "refusal" with some output already returned) were already billed before today — no change there. (4) Refusals in all other categories (those with stop_details.category outside these three, or with a null category) remain unbilled before output. (5) For cost model updates, see: https://platform.claude.com/docs/en/build-with-claude/refusals-and-fallback#how-refusals-are-billedmaxProseWidth setting, session continuation bug fixes, and extended thinking preservation fixes for slash commands.--continue and --resume no longer re-send prior messages in altered form; extended thinking blocks no longer vanish when running /model, /rename, or /artifacts commands mid-session.--api direct and telemetry disabled (the common configuration for enterprise CI), auto mode now uses the server-side classifier by default instead of the client-side classifier. This changes which shell commands are auto-approved. Verify your CI workflows still behave as expected after upgrading — particularly any workflows that rely on auto-approving specific Bash patterns. (2) Session continuation bugs fixed. The --continue and --resume flags were re-sending earlier messages in slightly altered form, which could break prompt cache hits and distort API reasoning. This has been a silent cost-inflation bug for anyone running multi-session agent pipelines: upgrade to fix it. (3) Extended thinking now survives slash commands. Thinking blocks were being silently dropped when you ran /model, /rename, or /artifacts commands while Claude was working. This is now fixed — critical for workflows that switch models mid-session on complex reasoning tasks. (4) maxProseWidth setting: Add "maxProseWidth": 100 to ~/.claude/settings.json to cap Claude's prose output width in wide terminals while keeping tables and code blocks full-width. (5) Skills in anthropic-skills or claude-ai namespaces no longer auto-load unless published as a named plugin — check any custom skills using these namespace prefixes.--continue/--resume flags in multi-session agent pipelines; you run Claude Code in auto mode with telemetry disabled; you use extended thinking with slash commands; you have custom skills in anthropic-skills or claude-ai namespaces.EffortQuick — claude update or bump your version pin; verify auto mode behavior in CI after upgrading..deb and .rpm packages for Linux (x64), completing native desktop coverage across macOS, Windows, and Linux. Simultaneously, CLI v4.1.21 adds the ai& provider (Japanese open-weight models via an OpenAI-compatible endpoint), expands the model catalog to 6,386 models across 209 providers, and sets Claude Opus 5.5 as the default for 11 providers.apt install cline-desktop_0.0.35_amd64.deb or the equivalent RPM. This brings plugin slash commands (previously broken in Desktop), provider-specific reasoning-effort memory, and the Diagnostics export feature (Settings → export a support-safe text file with logs and config) to Linux users. (2) Diagnostics export is worth knowing. When filing a GitHub issue for a Cline Desktop problem, you can now export a single text file with version, OS, settings, logs, and session manifests with sensitive data stripped — substantially faster than manual log collection. (3) Claude Opus 5.5 is now the default for 11 providers. If you use any of the 11 affected providers in Cline and haven't pinned a specific model, you'll be on Opus 5.5 after upgrading — check costs and expected behavior before updating in production workflows. (4) Local model session stability improved. Long Ollama/LM Studio/llama.cpp sessions that previously failed at the model's output-token limit now compact and retry instead of dying — reducing silent session truncation on local model workflows. (5) New ai& provider serves open-weight Japanese language models via an OpenAI-compatible endpoint — the first Japan-focused provider in the Cline catalog.ai-memory install-mcp --client <harness> --apply. The zero-LLM offline mode (FTS5 full-text search) works out of the box without any API key — add Anthropic/OpenAI keys only if you want semantic ranking. Typed handoffs (ai-memory handoff create) give you a claim-once protocol for parallel agent sessions: agent A creates a handoff, agent B claims it, both sides see the claimed state. For teams: the server supports per-project shared memory with user attribution and audit logs. Note: star-history.com was egress-blocked in this scan — momentum is from GitHub Trending weekly data only; validate the growth curve before treating this as confirmed sustained growth.SWE-bench Pro (September 2026 update): Claude Opus 5.5 leads the BenchLM SWE-bench Pro table at 89.9%, ahead of Claude Fable 5.1 (81.2%) and Claude Mythos 5 (80.3%), across 75 evaluated models. Separately, on Scale's standardized public/commercial split: Muse Spark 1.1 leads at 61.5% (public) / 51.5% (commercial), with GPT-5.4 xHigh at 59.1% and Claude Opus 4.6 at 47.1%. Sources: benchlm.ai/benchmarks/swe-bench-pro; morphllm.com/swe-bench-pro — September 2026 updates
compact_before_next_turn() addition.Nothing cleared the quality bar this period. simonwillison.net was egress-blocked. Nathan Lambert, Eugene Yan, and Sebastian Raschka had no Sep 24 posts confirmed. No HN threads with score >200 and new Sep 24 technical depth confirmed.
GitHubCopilotOAuthFlow device authorization, BoolCriteria for boolean fields, TypeSafeModel whole-number choices; fixes streamed OpenAI logprobs, GPT models on Bedrock, Gemini Live audio. [github.com/pydantic/pydantic-ai/releases/tag/v2.49.0]Anthropic confirmed both follow Opus 5.5 "within weeks" of the September 22 release. No pricing or specific dates announced. [anthropic.com/claude-opus-5-5]
Migrate to antigravity-preview-09-2026, rename snake_case built-in tool parameters to PascalCase, and update file-edit handling. See September 18 digest Breaking Changes for the full migration guide. [ai.google.dev/gemini-api/docs/changelog]
Standard pricing: $10/$1 (cached)/$50 per MTok. Long context (>272K tokens): $20/$75 per request. Free billing period ends October 4. [platform.openai.com/docs/changelog]
Filtered from 30+ primary sources against a published quality rubric. No press releases, no fluff — only what changes what you build.