/ultrareview uploading uncommitted credential files) and a token-limit bug (model switching mid-session); update nownpm install -g jevgrep
jevgrep "retry logic for API calls"
jevgrep "authentication middleware" | head -100
| Benchmark | GPT-6.1 Sol | GPT-6 Astra | GPT-5.6 Sol | Claude Opus 5.5 |
|---|---|---|---|---|
| DeepSWE v1.1 | matches Astra | — | −6.4 pts vs Sol 6.1 | — |
| AutomationBench | — | — | — | Sol 6.1 +2.2 pts |
| OSWorld 2.0 | within 2.1 pts of Astra | — | — | — |
ultrafast: true parameter in the API (exact parameter name to be confirmed in platform docs). GPT-6.1 Sol Ultrafast is "coming in the coming days" per OpenAI's announcement — only GPT-6 Astra Ultrafast was confirmed as available now in the API.Nothing cleared the quality bar this period. arXiv.org is egress-blocked — cs.AI, cs.CL, and cs.LG new submissions for September 29 could not be directly fetched and scored. The closest candidate, arXiv:2609.31587 ("Compact Documentation for Coding Agents: A Benchmark, an Optimizer, and Why It Does Not Transfer"), has a strong result (compact documentation does not help agents resolve real repo issues when source is present) but was published September 25 — outside the 24-hour scan window. Moved to near-misses.
@ characters were exposed in redacted logs); one correctness fix (model-switching mid-session left incorrect output-token limits).CLAUDE_CODE_DISABLE_WEB_FETCH, claude --desktop, claude plugin configure, and allowedProviders; fixes a security bug where /ultrareview uploads could include uncommitted credential files, and a log redaction bug exposing URL passwords./ultrareview credential exposure: if you use /ultrareview and have uncommitted .env, credential, or secret files in your working directory, those files could have been included in review uploads before this fix. Audit any /ultrareview runs you've done with sensitive files present. Update to v2.1.285 immediately. (2) Security — URL password redaction bug: if your MCP server configs or worktree remotes use URLs with passwords (format: user:password@host), those passwords were being logged in redacted logs. This is fixed in v2.1.285. (3) New allowedProviders managed setting: gateway deployments can now restrict which API providers Claude Code is allowed to call — useful for organizations that need to enforce Claude-only usage or specific provider allowlists. Also: CLAUDE_CODE_DISABLE_WEB_FETCH=1 disables the WebFetch tool globally (useful for air-gapped or restricted environments); claude --desktop opens the Claude desktop app on the current directory./ultrareview in repositories containing sensitive files (update immediately); you run Claude Code with URL-password-formatted MCP configs; you administer gateway deployments and need provider restriction.EffortQuick — Claude Code auto-updates; audit prior /ultrareview runs for credential exposure.npm install -g jevgrep), then use jevgrep "retry logic for API calls" — it queries Jev's relevance model across your directory structure and returns file paths plus verbatim excerpts. The key integration: it outputs to stdout as clean file paths + excerpts, making it directly usable as a shell tool in Claude Code (--allowedTools bash), Codex, Aider, or any coding agent that can execute shell commands. The tool calls Jev (TypeSafe's decision model) for relevance scoring — this is not fully local, so requires a Jev account.npm install -g golive-skill) or add as an MCP skill to Claude Code. The workflow: golive detect scans the project and identifies what it needs, golive plan proposes the infrastructure, golive approve lets you review before anything is created, golive apply provisions it on your accounts, golive verify confirms everything is running. The explicit approve step means no resources are created without review. Works with Claude Code, Codex, Cursor, and any agent that can execute shell commands or call MCP tools.No confirmed new leaderboard entries within the 24-hour scan window. Current SWE-bench Pro standings (as of September 28–29, 2026): Claude Opus 5.5 leads at 89.9%, followed by Claude Sonnet 5.5 (81.3%) and Claude Fable 5.1 (81.2%). On SWE-bench Verified, Claude Fable 5 leads at 0.950 (116 evaluated models). Notable prior entry: Qwen3.8 Max at 67.7% SWE-bench Pro ($2/$6 per MTok) ahead of GPT-5.6 Sol (64.6%) — model released August 3, 2026; leaderboard position confirmed this week but entry predates scan window. lmarena.ai egress-blocked; ELO movements not confirmed for September 29.
Nothing cleared the quality bar this period (concrete data, primary source, within window). HN threads on DevDay announcements were accumulating at time of scan but not yet at score >200. simonwillison.net egress-blocked.
us.xai.grok-4.7 (and global.xai.grok-4.7); 500K token context window; 4 configurable reasoning effort levels; supports Converse, Responses, Chat Completions, and InvokeModel APIs. Announced September 28. aws.amazon.com/blogs/machine-learning/grok-4-7-is-now-available-on-amazon-bedrock/Standard GPT-6.1 Sol is available now; the Ultrafast variant (8× speed, 6× price) was announced today but is not yet accessible in the API. Watch platform.openai.com/docs/changelog for the release. [openai.com/index/devday-2026-recap/]
Announced alongside Opus 5.5 on September 22 as "coming in the coming weeks." No model card, pricing, or date announced as of September 29. [Anthropic Opus 5.5 announcement, September 22, 2026]
Limited preview today; broad release expected shortly. [openai.com/index/devday-2026-recap/]
Standard pricing begins; free billing period ends October 4. [platform.openai.com/docs/changelog]
Migrate to antigravity-preview-09-2026, rename snake_case 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.