The full-duplex model that has powered ChatGPT Voice since July became available to all API users on September 10. Benchmarks cited by Unite.AI put it at 86.2% on Tau3 Voice Intelligence against 45.7% for GPT-Realtime-2.1, turn-taking latency at 0.798 seconds against 1.41, and Full Duplex Bench Interactivity at 80.10% against 45.4%. The design that matters more than any of those: GPT-Live-1 holds the conversation and hands reasoning to GPT-6 Astra or a third-party backend without dropping state.

Why it matters. The $0.05 buys the front-end voice layer only, per Unite.AI's pricing breakdown. An hour of conversation is $3.00 before a single backend token, so budget against cost per conversation hour with the reasoning model priced in, not the per-minute headline.

Run the per-hour math →

Dario Amodei published We Must Pace the Frontier on September 12 and committed Anthropic unilaterally to embedded third-party evaluators, naming METR as the model: desks in Anthropic offices, access badges, company laptops, workspace permissions. The contract term is the sharp part. Reviewers can publish findings on risk levels, incidents, and the access they did or did not receive, with no editorial control by Anthropic, which keeps only a narrow redaction right for security-sensitive, legally privileged, commercially sensitive, or third-party confidential material. Reviewers can also say publicly when a redaction cut something material. Sam Altman endorsed the essay on X within hours and said OpenAI would adopt one of the proposed safeguards, per NBC News.

Why it matters. Every safety claim you currently evaluate a vendor on is self-reported. If this holds its shape through the actual contract, third-party findings become something a procurement team can ask for by name.

Read the evaluator terms Anthropic wrote down →

Manifold Security disclosed GitSpawn on September 12, reported by The Hacker News. A poisoned core.fsmonitor entry in a repo's .git/config names a command Git runs on any index refresh, so git status or git diff fires it at agent startup, with your full privileges, before the agent shows a trust prompt. It needs an intact .git directory to travel, which means ZIP archives, USB sticks, and synced folders; a normal git clone strips it. The fix is one line in any terminal: git config --global core.fsmonitor false.

Patch state is uneven, so run the flag regardless of what you use. Codex CLI is fixed at 0.131.0 (CVE-2026-19592) and Goose at 1.44.0 (CVE-2026-72718). Claude Code closed the core.fsmonitor path at 2.1.196, but the researchers found a second route through claude ultrareview still live at 2.1.252. Hermes Agent, Qwen Code, and Grok Build were unpatched as of September 1. Cursor users are the least exposed of the group: it shipped a fix roughly three weeks before disclosure, so a current install is already covered.

Check your agent against the patch table →

Per GitHub's September 11 changelog, Lite reviews now combine findings from several agents rather than a single pass. In GitHub's own experimentation that raised the average number of addressed comments per review by 47% for high-severity findings, 31% for medium, and 11% for low, while cutting review cost by about 8%. Copilot now also resolves its own comment threads once a pushed commit addresses the feedback.

Why it matters. Lite was the tier teams downgraded to when the full reviewer got expensive. More severe findings for less money removes that tradeoff, and the auto-resolve ends the manual thread cleanup that made Copilot reviews unreadable on a 40-file pull request.

See what changed in Lite →