← All digests
AI Developer Digest

Tue, Sep 22, 2026

12 signals that cleared the gate41 scanned26 min read
The Signal — start here
Anthropic fired the day's opening shot: Claude Opus 5.5 ships with Fable 5.1-class performance at 40% lower cost than Opus 5, accompanied by a full-stack coordinated release — SDK, Claude Code, and platform notes all land the same morning. The three breaking changes to the Opus 5.5 API (thinking mode, tool_choice, computer-use toolset) require immediate code review for any team migrating. Alongside that: OpenAI Codex 0.156.0 goes stable with voice conversations by default and a fullscreen TUI — the first non-alpha with UX-facing features. And the technical story of the week belongs to Devin: Cognition engineer Eric Lu directed the agent to build a GPU number-field-sieve siever from scratch in nine hours, then used it to factor RSA-260 — a 862-bit number that had stood for 36 years. The security implications for developers using 1,024-bit RSA are real and immediate.
Must-reads today
1
Claude Opus 5.5 breaking API changes — three changes to thinking, tool_choice, and computer-use toolset require code review before migrating; cache reads drop 60%
2
Devin factors RSA-260 — Cognition's agent built the GPU siever in 9 hours; the implication is that 1,024-bit RSA is now feasible to factor for ~$30M at market GPU prices
3
Codex 0.156.0 stable — first non-alpha with voice-by-default, /tui fullscreen mode, and worktree support enabled globally; catches up materially to Claude Code's UX

🏆 Repo of the Day

1

google/ax — Kubernetes-native declarative runtime for autonomous agent workloads

go install github.com/google/ax/cmd/ax@latest

Define a task declaratively

cat > task.yaml << 'EOF'

apiVersion: ax.google.com/v1

kind: Task

metadata:

name: my-agent-task

spec:

model:

provider: anthropic

name: claude-opus-5-5

workspace:

repos:

branch: main

gateway:

allowlist:

  • api.anthropic.com

  • github.com

EOF

ax apply -f task.yaml

ax watch my-agent-task

Breaking Changes

1
●Breaking

Claude Opus 5.5 — three API changes required when migrating from Opus 5 or earlier models

What changed
Claude Opus 5.5 introduces three behavioral changes that return HTTP 400 errors for request patterns that previously worked on Opus 5 and earlier models. These apply only when using claude-opus-5-5; existing code targeting other Claude models is unaffected.
TL;DR
Disabling thinking, non-auto tool_choice modes, and the legacy computer_20251124 tool all return 400 errors on Opus 5.5 — three separate code paths to audit before migrating.
Developer signal
(1) Thinking cannot be disabled. thinking: {"type": "disabled"} returns a 400 error. Opus 5.5 uses always-on adaptive thinking; use the effort parameter to control depth instead. If you previously passed thinking: disabled to reduce cost or latency, remove that parameter and tune effort to the appropriate level. (2) tool_choice types any and tool return 400. Pass tool_choice: {"type": "auto"} with strict tool-use settings instead. The auto type with a strict schema enforces the same behavior — check your tool call construction if you relied on any for guaranteed tool invocation. (3) computer_20251124 tool returns 400 on Claude API and Google Cloud. Migrate to the computer_toolset_20260801 toolset. Exception: computer_20251124 continues working on Amazon Bedrock — do not remove it from Bedrock-targeted code until AWS publishes a separate deprecation notice. Audit all three changes before switching a pipeline to claude-opus-5-5.

Affects you ifYou call the Claude API or Google Cloud AI with any of: thinking: {"type": "disabled"}; tool_choice: {"type": "any"} or {"type": "tool"}; the computer_20251124 tool type.EffortModerate — three distinct parameter changes; straightforward but requires systematic audit of every request construction path in your codebase.

Model Releases

1
High

Claude Opus 5.5 — Fable 5.1-class performance at 40% lower cost; Terminal-Bench 4.0 at 66.4%

What changed
Anthropic released Claude Opus 5.5 (claude-opus-5-5), replacing Opus 5 as the flagship Opus-tier model. Input tokens drop from $5 to $4/MTok, output from $25 to $20/MTok, and cache reads from $0.50 to $0.20/MTok. The model runs 30% faster than Opus 5 and uses always-on adaptive thinking.
TL;DR
Claude Opus 5.5 scores 66.4% on Terminal-Bench 4.0 (Opus 5: 52.3%) and 1846 Elo on GDPval-AA v2.1 (Fable 5.1: 1735, Opus 5: 1708), priced at $4/$20 per MTok with cache reads at $0.20/MTok — 60% cheaper cached.
Developer signal
(1) Swap model ID for an immediate cost cut. Changing claude-opus-5 to claude-opus-5-5 in any pipeline drops your bill 20% on uncached tokens and 60% on cache reads with comparable output quality — but audit the three breaking changes in the [BREAKING] section first. (2) The agentic coding benchmark is the headline number. Terminal-Bench 4.0 at 66.4% is a 14-point jump over Opus 5 (52.3%) and surpasses Fable 5.1 on that specific benchmark. For long-horizon coding tasks — scaffolding, multi-file refactors, migration scripts — Opus 5.5 is now Anthropic's recommended model. (3) Always-on adaptive thinking changes prompt design. Thinking cannot be disabled (see [BREAKING]), but you can tune depth via effort. If you were using thinking: disabled to speed up simple tasks, model those with a smaller Claude model (Sonnet, Haiku) instead; Opus 5.5 is optimized for complex work that benefits from reasoning. (4) 1M context, 128k max output. These are the same as Fable 5.1 — no context window surprises when migrating Fable pipelines to Opus 5.5. (5) Distillation protection. A "preserved thinking safeguard" prevents API users from extracting reasoning traces — plan accordingly if your pipeline was inspecting raw thinking blocks for prompt optimization. (6) Claude Sonnet 5.5 and Haiku 5.5 ship within weeks, suggesting the full Claude 5.5 family completes this month.

Affects you ifYou run Claude Opus 5 in any pipeline; you use claude-opus-5 in cost-sensitive long-horizon tasks; you build with computer use or thinking-intensive agents; you use the Claude API, Amazon Bedrock, Google Cloud, or Microsoft Foundry (all supported).EffortModerate — model ID swap is Quick, but the three breaking changes require a code audit before switching.

API & SDK Changes

1
Medium

Anthropic SDKs v1.8.0 (Python) / v0.128.0 (TypeScript) — inline tools beta + MCP tool-list pinning

What changed
Both SDKs add Claude Opus 5.5 support, inline tool definitions via the inline-tools-2026-09-15 beta header, and MCP tool-list pinning (beta). Python also fixes a Python 3.13 crash when a stream is left open at exit, and both SDKs fix add_tools() to take effect immediately.
TL;DR
Inline tool definitions (inline-tools-2026-09-15 beta) let you introduce or modify a tool mid-conversation via a tool_addition block in a system message, without touching the original tools array — so the prompt cache still hits on the unchanged earlier context.
Developer signal
(1) Inline tools solve the "I want to add a tool mid-session without a cache miss" problem. Pass the anthropic-beta: inline-tools-2026-09-15 header and add a role "system" message with a tool_addition block containing the full tool definition. The earlier messages and tools array stay exactly as originally sent, so the prompt cache is preserved. (2) MCP tool-list pinning (beta) freezes the MCP tool catalog mid-session — useful when your MCP server updates its tool list and you need consistent behavior for a long-running agent session. Combine with mcp-client-2026-09-15 header when adding MCP servers inline. (3) Python 3.13 fix is important: if you run SDK code on Python 3.13 and leave a stream open at exit (common in async agent loops), upgrade to v1.8.0 immediately to avoid interpreter crash. (4) add_tools() now takes effect immediately — previous behavior deferred the update; if you call add_tools() mid-session and then invoke a tool in the next turn, it now works as expected without a session restart.

Affects you ifYou use anthropic Python ≥ any version below v1.8.0 or TypeScript SDK below v0.128.0; you run multi-turn agent sessions where tool availability changes; you use Python 3.13 with streaming.EffortQuick — pip install "anthropic>=1.8.0" or npm install @anthropic-ai/sdk@^0.128.0; inline tools require beta header and request structure changes.

Research

1
High

Devin factors RSA-260 — agent builds GPU GNFS siever in 9 hours; 1,024-bit RSA now estimable at ~$30M

What changed
Cognition engineer Eric Lu directed Devin to build a GPU-accelerated lattice siever for CADO-NFS (the open-source General Number Field Sieve implementation), then used it to factor RSA-260 — a 260-digit, 862-bit number from the RSA Factoring Challenge that had stood since February 2020, making it the largest number ever publicly factored from that challenge.
TL;DR
Devin produced a working GPU siever (named glas) that beat the CPU baseline in ~9 hours from a single prompt on August 13, 2026; total computation was ~4,900 GPU-days; Eric Lu estimates a hyperscaler could factor 1,024-bit RSA for ~$30M at current market GPU prices.
Developer signal
(1) 1,024-bit RSA is now in-scope for well-funded actors. The $30M estimate is for current market GPU pricing with the optimizations Devin built; with further optimization or bulk GPU pricing, the effective cost is likely lower. If any service you build or maintain uses 1,024-bit RSA keys for anything longer-lived than a session (TLS certs, signing keys, stored secrets), plan a migration to 2,048-bit RSA or an ECC equivalent. (2) RSA-2048 is not affected. The GNFS algorithm's sub-exponential complexity still makes 2,048-bit RSA infeasible at any current-generation GPU scale. (3) The developer angle is the agent story, not the math. Devin was given a single GPU and a prompt ("build me a drop-in GPU replacement for the CPU lattice siever in CADO-NFS") and produced a working implementation 9 hours later that beat the CPU baseline. Lu then directed optimization over the following week. This is the most significant public demonstration of an agent solving an open research problem from a standing start rather than executing a pre-specified implementation. (4) The GNFS pipeline Devin built is open. CADO-NFS is MIT-licensed and the GPU siever (glas) will be contributed upstream per Lu's post — developers interested in applied cryptography or GPU numerical methods can examine the agent-generated code directly.

Affects you ifYou maintain any system using 1,024-bit RSA keys; you build cryptographic tooling; you're evaluating what long-horizon agentic coding can accomplish on novel technical problems.EffortQuick (key migration) — update key length in key generation tooling; Significant if you manage long-lived signed data (certificates, tokens, stored secrets with HSM-backed keys).

Tooling

1
High

OpenAI Codex 0.156.0 — first stable release with voice-by-default, fullscreen TUI, and worktree support

What changed
Codex 0.156.0 is the first stable (non-alpha) release of the 0.156.x line, shipping features that graduated from the alpha series: a fullscreen terminal UI (/tui), voice conversations enabled by default, usage analytics, worktrees as a default workflow, and enhanced sandbox security.
TL;DR
Codex 0.156.0 ships a fullscreen TUI with transcript search and mouse copy, F8-toggle voice with bundled audio runtimes for Linux/Windows, /usage analytics, worktree support on by default, and Mermaid diagram rendering — the most UX-significant update to Codex since its initial release.
Developer signal
(1) Voice is now on by default — toggle with F8, configure via /voice settings. Bundled audio runtimes are included for Linux and Windows, so no separate install needed. If you run Codex in headless environments (CI, containers), confirm the voice runtime doesn't add unintended dependencies to your image; disable with --no-voice if needed. (2) /tui fullscreen mode adds transcript search and mouse-based selection and copy — useful for reviewing long multi-file sessions. Works on all platforms, including SSH sessions (clipboard forwarding fixed for tmux). (3) Worktrees are now default: Codex creates a git worktree per session, so concurrent sessions on the same repo don't interfere. If your workflow relied on in-place edits, adjust your session management. (4) /usage dashboard shows per-session and cumulative token/cost stats inline — no external dashboard needed for cost monitoring. (5) Sandbox security: Windows, Linux, and macOS vulnerabilities addressed — if you run Codex with any user input being executed in the sandbox, update now. (6) /daemon command allows background server updates without interrupting an active session.

Affects you ifYou run OpenAI Codex as a CLI coding agent; you run Codex in CI/CD or containerized environments; you use concurrent sessions on the same repository.EffortQuick — npm install -g @openai/codex@0.156.0; worktree default may require workflow adjustment for in-place edit workflows.

Rising Dev Tools

1
Notable

superdesigndev/treg — "OpenRouter for agent tools": 3,000+ API endpoints via a single token

Developer signal
Install the CLI and point your agent's tool calls at the treg endpoint. For Claude Code, the plugin handles the integration automatically. You can also bring your own credentials (connect existing API keys to access endpoints unmetered) or use treg's pooled keys for per-call billing. The treg call <provider.endpoint> --query key=value pattern is a one-liner for testing; in agents, the endpoints are exposed as standard MCP tools.

Affects you ifYou build agents that need access to specialized external data sources; you manage shared API credentials across a development team; you want to experiment with 3rd-party data APIs before committing to individual subscriptions.
GitHub Trending (Python, weekly) | Stars: ~2,140 (↑ +197 today) | Link: github.com/superdesigndev/treg

Benchmarks & Leaderboards

Terminal-Bench 4.0 — new SOTA with Opus 5.5: Claude Opus 5.5 scores 66.4% on Terminal-Bench 4.0, up from Opus 5's 52.3% — a 14-point improvement and the largest single-generation jump Anthropic has published on this benchmark. Fable 5.1 was the previous reference point (Anthropic does not publish its Terminal-Bench 4.0 score for Fable 5.1, citing the benchmarks as "less reliable guides to real-world quality" as margins narrow). Source: anthropic.com/claude-opus-5-5

GDPval-AA v2.1 (44-occupation professional competency benchmark): Opus 5.5 at 1846 Elo, ahead of Fable 5.1 (1735) and Opus 5 (1708). This is Anthropic's internal benchmark covering knowledge work spanning 44 professional occupations. Source: anthropic.com/claude-opus-5-5

Other leaderboards: lmarena.ai (LMSYS Chatbot Arena) and swebench.com were egress-blocked this scan; no new ELO movements or SWE-bench position changes confirmed. Last known standings carry forward: SWE-bench Pro — Claude Fable 5.1 leads at 81.2%.

Technical Discussions

Nothing cleared the quality bar this period. Simon Willison's Sep 21 post on Jev provides secondary coverage of the TypeSafe AI launch (moved to Trends); no HN threads with score >200 and concrete technical depth on new items within the 24-hour window were confirmed; Nathan Lambert, Eugene Yan, and Sebastian Raschka had no Sep 22 posts confirmed.

Quick Hits

  • Claude Code v2.1.280 (Sep 22) — Opus 5.5 as default Opus model; mouse support in fullscreen (wheel scrolling, clickable options); CLAUDE_CODE_MAX_MCP_DESCRIPTION_LENGTH env var for MCP tool descriptions; VS Code: /skills command, /plan mode, Status/Sandbox/Chrome integration dialogs; multiple keyboard input and resume fixes. [github.com/anthropics/claude-code/releases/tag/v2.1.280]
  • OpenHands v1.22.0 (Sep 22) — Remote (SSE/HTTP) MCP servers can now be health-checked from the Settings page via new POST /api/v1/mcp/test endpoint; Git Sync opened to org admins/owners on cloud backends (previously local-backend-only). [github.com/OpenHands/OpenHands/releases/tag/v1.22.0]
  • OpenHands v1.21.0 (Sep 22) — Docker execution workspace request support; GitHub tree URL acceptance for app sources; MCP server sibling preservation fix. [github.com/OpenHands/OpenHands/releases/tag/v1.21.0]
  • Cline SDK v0.0.85 (Sep 22) — Improved retry logic for token-limit failures (3 retries with model-aware output token scaling); model catalog updated 6,188 → 6,237 models across 209 providers; new Cline Pass tier models added; defaults changed for 10 providers. [github.com/cline/cline/releases/tag/v0.0.85]
  • LangGraph 1.2.12 + SDK 0.4.5 (Sep 21) — response_schema parameter added to interrupt() for typed interrupt responses; type declaration fixes for v3 stream projections; dependency bumps (anyio, websockets). [github.com/langchain-ai/langgraph/releases/tag/langgraph==1.2.12]
  • pydantic-ai v2.47.0 (Sep 21) — TypeSafeModel improvements for Jev integration: None route support, expanded field type coverage (3 new types), rubric validation strictness, route naming in requests; ag-ui-protocol pinned below v1. [github.com/pydantic/pydantic-ai/releases/tag/v2.47.0]
  • Gemini CLI v0.62.0-nightly.20260922 (Sep 22) — Proxy-agent compatibility normalization for environment proxy handling; tool call updates now emitted before permission requests in ACP mode. [github.com/google-gemini/gemini-cli]

Worth Watching (Announced, Not Yet Shipped)

4

Claude Sonnet 5.5 and Haiku 5.5 — within weeks

Anthropic confirmed both follow Opus 5.5 within weeks. No pricing or specific dates announced. Expect the Claude 5.5 family to complete before end of October 2026. [anthropic.com/claude-opus-5-5]

Google Antigravity `antigravity-preview-05-2026` shutdown — **October 5, 2026 (13 days)**

Migrate to antigravity-preview-09-2026, rename all snake_case built-in tool parameters to PascalCase, and update file-edit handling from full-rewrite to line-range model. See the September 18 digest Breaking Changes for the full migration guide. [ai.google.dev/gemini-api/docs/changelog]

GPT-6 Astra billing starts October 5, 2026 (13 days)

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]

DeepSeek V4.1 Pro — no launch date confirmed

V4.1 Pro confirmed in development as V4 Pro successor. No date announced; V4 Pro continues at existing pricing until V4.1 Pro ships. [api-docs.deepseek.com/updates/]

Filtered from 30+ primary sources against a published quality rubric. No press releases, no fluff — only what changes what you build.