A single daily edition of announcements from the AI coding tools you use, read from a curated source list, summarised, and linked back to the original. No infinite feed, no account, no advertising.
Re-reading context on Claude's top coding model dropped to a quarter of its old price
Claude Fable 5.1 and Mythos 5.1 are generally available, and API cache reads fall from $1.00 to $0.25 per million tokens while input and output pricing holds at $10 and $50. Anthropic puts a typical workload about 25% cheaper than Fable 5 and highly agentic work up to roughly 45% cheaper, because a long-running loop re-reads far more than it writes. Terminal-Bench 4.0 rises to 55.8% from 42.0% and Terminal-Bench-Science 0.1 to 52.6% from 24.7%, with CursorBench 3.2.0 at 73.4% under max effort. Mythos 5.1 reaches 60.9% on Terminal-Bench 4.0 but is restricted to vetted US organisations.
Auto mode gained a rule against your agent fetching cloud credentials
Claude Code 2.1.257 adds a Containment Escape rule to auto mode covering cloud metadata-credential fetches and cross-tenant reach, the two moves that turn a sandboxed agent into one holding your infrastructure keys. The same release shows a one-time prompt before the first file read outside the working directories. It also makes claude-fable-5-1 the default Fable model with a 1M-token context window, and adds timeFormat and timeZone settings plus a CLAUDE_CODE_SUBAGENT_MODEL_FORCE environment variable.
An agent's approval can satisfy your repository's merge requirement
GitHub Copilot code review can submit approvals that count toward repository merge requirements, though approvals are disabled by default and an administrator has to turn them on. Control is gated at enterprise, organisation and repository level, and a repository can also restrict which file paths Copilot reviews. New commits dismiss a Copilot approval exactly as they dismiss a human one. The capability is in public preview for Copilot Pro, Pro+, Max, Business and Enterprise.
Fable 5.1's low and medium effort settings do no reasoning at all
Simon Willison ran one prompt across all five of Claude Fable 5.1's effort tiers and found low and medium emit no reasoning tokens, both landing near 2,000 tokens in about 23 seconds for roughly 10 cents. High barely engages at about 2,600 tokens, 30 seconds and 13 cents. Extra-high jumps to 36,767 tokens over 7 minutes 51 seconds for $1.83, and max reaches 65,927 tokens in nearly 14 minutes for $3.30 — a twenty-five-fold cost gap between high and max for the same request.
Six models you may have pinned in Copilot stopped answering on 1 September
GitHub deprecated Gemini 3.1 Pro, Claude Opus 4.5 and 4.6, Claude Sonnet 4.5 and 4.6, and Raptor Mini across Copilot experiences on 1 September. The suggested replacements are Gemini 3.7 Flash, Claude Opus 4.7, 4.8 or 5, Claude Sonnet 5, and MAI-Code-1.1-Flash. Claude Sonnet 4.6 stays available to individual subscribers on annual plans. Enterprise administrators may need to enable the replacements in Copilot settings before anyone can select them.
A graphics API was reimplemented by an agent, and its author has not fully read it
Rick Brewster, the author of Paint.NET, credits Claude with a clean-room reverse-engineered rewrite of Direct2D that lets the editor run under WINE, saying it was written by our good friend Claude and would not have been possible otherwise. He describes the 180,000-line result as largely vibe coded, meaning he has not thoroughly reviewed it. Brewster praises Claude's clever and tireless reverse engineering of Direct2D's effects library, while noting he still had to supervise its resource management and architectural decisions.
A 1.7GB toolchain sits in your Codex cache, complete with LibreOffice
Simon Willison found that OpenAI's ChatGPT and Codex desktop app unpacks 1.7GB of runtimes into a local cache directory: Python at 440.6MB, Node.js at 446.4MB, a headless LibreOffice at 429.7MB, Poppler at 187.9MB, git at 148.1MB, and smaller libraries including libheif and jxrlib. A plugins directory alongside them carries skills telling Codex how to invoke each binary, which makes the bundled toolchain part of the agent's documented capability surface rather than an incidental dependency.
Specs and delegated agent work left preview and run in a browser
Kiro Web reached general availability on 1 September, letting developers build, delegate and steer development work from a browser with no local installation required. The announcement describes exploring ideas collaboratively, shaping work with Specs, and handing off tasks end to end. A separate same-day change lets cloud configuration sync into local sessions, so Steering files and custom agents defined in the cloud are available when an IDE or CLI session starts.
Query results from an MCP server arrive with their column names attached
datasette-mcp 0.2 is the first non-alpha release of the plugin that adds an MCP server to a Datasette instance. The rows payload returned by execute_sql changes from an array of arrays to objects, mapping each value to its column name instead of leaving the model to track positions, which Willison reports helps weaker models keep context. The minimum dependency moves to mcp>=2.1.1.
