deniedModels lets admins blocklist specific models even when availableModels would otherwise allow them, and availableModelsMatch: "exact" prevents new model releases from auto-entering the allowed set. The /doctor prompt-audit command — which audits CLAUDE.md files, skills, and commands for patterns written for older models — is the practical complement: a way to surface prompts that were tuned for Claude 3.x or Sonnet 4.x but are now running on Fable 5.1 or Opus 5.5 with potentially different behavior. Together, these three additions suggest Anthropic is building out the enterprise operator layer of Claude Code in earnest. Separately, pydantic-ai v2.51.0 adds OpenAI's GPT-Live realtime API as a first-class model type, and Cline Desktop v0.0.37 patches a Bedrock reasoning config error that was breaking with a cryptic "Unknown parameter" message.deniedModels managed setting, availableModelsMatch: "exact" for strict version locking, /doctor prompt-audit to catch stale model-specific prompts, and OTEL logging of MCP tool content; the enterprise model governance story in one releaseA quiet-news day — reaching into the watchlist for a tool that's been genuinely underrated.
bash <(curl --proto '=https' --tlsv1.2 -sSf https://setup.atuin.sh)
atuin search --exit 1 # find all commands that failed
atuin search git push # find all git push commands with timestamps and dirs
atuin stats # show your most-used commands
No breaking changes this period.
No model releases within the 24-hour scan window. (Claude Opus 5.5, GPT-6 Sol/Luna, and Meta Muse Spark 1.3 were covered in prior digests; Claude Sonnet 5.5 and Haiku 5.5 are in Worth Watching.)
deniedModels, availableModelsMatch: "exact"), a new /doctor prompt-audit command to audit CLAUDE.md files and skills for stale model-specific patterns, OTEL logging of MCP tool content (opt-in), Amazon Bedrock's Mantle endpoint as an upstream provider, multiple performance improvements (startup latency, first-reply latency), and approximately 25 bug fixes across SDK sessions, MCP, plugin management, Vim mode, and platform-specific paths.deniedModels, availableModelsMatch: "exact"), a prompt-auditing tool, OTEL MCP tracing, and significant startup/latency wins.deniedModels managed setting (enterprise admins). New in .claude/settings.json under managed.deniedModels: an array of model IDs that are blocked even if availableModels would otherwise allow them. Use this when you want to permit a model family broadly but block a specific version — e.g., allow all claude-* models but block claude-haiku-4-5-20251001 due to a known behavior difference. This is additive to (not a replacement for) the existing availableModels list. (2) availableModelsMatch: "exact" managed setting (enterprise admins). Prevents new model releases from automatically entering the allowed set: when a new Anthropic model ships, it stays blocked until an admin explicitly adds its ID to availableModels. Previously, new models inherited access if they matched a prefix in the allowed list. For regulated environments where every model change needs a formal review cycle, this is the flag you need. (3) /doctor prompt-audit (all users). Run /doctor prompt-audit (or /checkup prompt-audit) to scan your CLAUDE.md files, skills, agents, and commands for patterns known to have been written for older Claude models — things like hardcoded version strings, deprecated tool call formats, or instructions that assume Sonnet 3.x behavior. The audit outputs a list of flagged items with explanations. If you maintain CLAUDE.md files that haven't been reviewed since the Fable/Opus 5.x era, run this first. (4) OTEL MCP tool content (observability). Set OTEL_LOG_TOOL_CONTENT=1 to include MCP tool inputs and outputs in OpenTelemetry span events. Previously, OTEL spans captured tool calls but not their content. Combined with a collector like Honeycomb or Jaeger, this enables request-level tracing of what MCP tools received and returned — useful for debugging multi-step agent workflows where a downstream tool produced unexpected output. (5) Bedrock Mantle endpoint. mantle is now an upstream provider in the Claude apps gateway, pointing at Amazon Bedrock's Mantle endpoint. If your org runs Claude Code through a Bedrock-based gateway, check whether your enterprise Claude Code deployment should route via mantle. (6) Performance wins. First-reply latency improved (pattern-compile step now runs while reply streams). claude -p startup is faster (interactive UI no longer loaded for piped/programmatic use). First-request latency improved through reused preconnected API connections. Expect a noticeable startup-time improvement on claude -p workflows. (7) --system-prompt now accepts both text and -file together. --system-prompt text -system-prompt-file myfile.md is now valid — file content prepends the inline text. Previously these couldn't be combined. (8) Windows PowerShell fix. The Bash tool no longer allows deletion of drive roots and Windows protected system directories — a meaningful safety fix for Windows Claude Code users running agents with broad file permissions.-p flag for scripted/programmatic invocations (startup speedup); you run Claude Code on Windows.EffortQuick — update Claude Code; then opt-in to OTEL_LOG_TOOL_CONTENT=1 and availableModelsMatch: "exact" as needed; run /doctor prompt-audit on existing CLAUDE.md files once after update.OpenAILiveModel for real-time interactions with OpenAI's GPT-Live platform, exposes context_window_used on realtime sessions, caches the agent graph instead of rebuilding on every Agent.run() invocation, reduces RunContext copying overhead in capability hooks, and aligns dated gemini-3.8-live identifiers.OpenAILiveModel enables real-time GPT-Live sessions. Import from pydantic_ai.models.openai_live. The interface mirrors the existing GeminiLiveModel — declare it as your agent's model and pydantic-ai handles the WebSocket lifecycle, reconnect, and typed tool binding. If you've already built a realtime agent on Gemini Live with pydantic-ai, the migration surface is a model swap and identifier change. (2) context_window_used on realtime sessions. Available on RunResult.usage.context_window_used for realtime sessions. The field is derived from GPT-Live's reported ratio when available, falling back to response usage data. Useful for monitoring context pressure in long-running realtime sessions where you can't trivially count tokens yourself. (3) Agent graph cache. If you're calling Agent.run() frequently in a hot path (e.g., per-request in a web server or per-event in a streaming pipeline), v2.51.0 removes the per-invocation graph rebuild. The graph is built once and cached. Benchmark the before/after on your workload — the improvement is larger for agents with complex tool graphs. (4) Gemini Live compatibility. gemini-3.8-live dated identifiers (e.g., gemini-3.8-live-001) are now correctly matched. Connections to Gemini 3.1 Flash Live and 3.8 Live with incompatible configurations are rejected at connection time rather than silently failing mid-session.Agent.run() at high frequency in a server or streaming context; you use Gemini Live models with dated version identifiers.EffortQuick — update pydantic-ai; use OpenAILiveModel for GPT-Live realtime sessions.none/disabled), previous builds (v0.0.36 and earlier) failed with "Unknown parameter: 'reasoningConfig'" — a silent API rejection that produced no useful error message in the UI. v0.0.37 fixes the parameter mapping. After updating, verify reasoning effort is set to your intended level in the Bedrock provider settings panel. (2) About page and What's New. If you're managing Cline Desktop deployments and users have been asking "what version am I on?", the About page in Settings now shows version, build, and an update check. The "What's new" dialog auto-triggers after updates — useful if you push Cline Desktop via MDM and users need visibility into what changed.No new research papers within the 24-hour scan window. (arxiv.org is egress-blocked; secondary search confirmed no high-signal papers dated September 26 in cs.AI or cs.CL. The nearest candidates — "Policy as Code: CAR-bench winner" arXiv:2609.29251v1 and others in DailyArXiv issue #530 — are dated September 24, outside the window. See near-misses.)
No confirmed new leaderboard movements for September 26. lmarena.ai egress-blocked. SWE-bench Pro September snapshot confirmed in the September 25 digest (Claude Opus 5.5 at 89.9% — no new entrants confirmed for today).
deniedModels (blocklist specific model versions), availableModelsMatch: "exact" (prevent new models from auto-entering the approved set), and /doctor prompt-audit (scan prompts for stale model-specific patterns). None of these make Claude Code more capable for a solo developer — they address the fleet-management pain of organizations where a new model release can silently change behavior across hundreds of engineer sessions. The same week, Meta's Model API went GA with a contributor tier that uses your prompts for training, and Cline Desktop v0.0.37 fixes a Bedrock reasoning config error — both pointing at a maturing enterprise deployment layer across coding-agent tools.availableModelsMatch: "exact" is the coding-agent equivalent of pinning a Docker image tag. Expect the other major coding agents (Cline, OpenHands, Codex) to add similar managed/enterprise settings in the next few releases — either as built-in capabilities or as deployment-configuration hooks.Nothing cleared the quality bar this period. simonwillison.net had no Sep 26 posts. Nathan Lambert, Eugene Yan, Sebastian Raschka: no Sep 26 posts confirmed. No HN threads with score >200 and Sep 26 technical depth confirmed.
No additional [NOTABLE] items this period.
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 the 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.