Two model deadlines for your calendar:
Nov 16: GitHub Copilot drops Claude Opus 4.8 and Opus 5 (use Opus 5.5) and GPT-5.6 Terra (use GPT-6.1 Sol or Luna).
Nov 30: Claude Code removes Opus 4.8 fast mode. Calls will error, no fallback.
Model names hardcoded? Grep now.
Codex Ultrafast math:
GPT-6 Astra Ultrafast: tokens up to 8x faster than Standard.
Cost: plan limits burn at 8x, purchased credits at 6x.
About 8x speed for 8x usage. Use it when you're blocked waiting, not for overnight runs.
Pro $500 and some Enterprise/Edu only.
Codex custom agent gotcha: sandbox_mode or approval_policy in an agent file is ignored. Agents inherit the parent's permissions.
To shrink a reviewer, in .codex/agents/reviewer.toml:
features.shell_tool = false
"Don't edit files" in instructions is not a read-only sandbox.
Running claude -p in CI? Check your Claude Code version.
Before 2.1.314, a startup failure could exit 0 with no output, so the pipeline went green on nothing. Now it prints the error and exits 1.
Same release: PreCompact hook output is cut at 10,000 chars.
GitHub Actions workflow lockfiles are now GA.
gh actions-lock writes .github/workflows/actions.lock: exact commits for each action and its transitive deps. Tags stay readable; runs use locked commits.
Dependabot updates it in the same PR.
Pinned is not the same as safe.
SemiAnalysis' new GDPx agent benchmark: 206 tasks, 11 domains. The best model scores 12%.
Opus 5.5 12.0, Gemini 4 Carbon 10.9, GPT-6 Astra 9.6. Finance and legal: basically all 0.
Catch: safety classifiers drag some scores down. For cyber, Muse Wonder Full leads clearly.
Agents write more of the code every month. Which skill will matter most for developers in 2027?
A. System design
B. Reviewing code fast
C. Product sense
D. Orchestrating agents
GitHub CLI tokens can now expire and auto-refresh. Opt-in today, planned default later.
gh auth login --expiring
or: gh config set expiring_credentials enabled
A leaked token dies within hours.
Catch: scripts that cache `gh auth token` break. Read it fresh each time.
DeepSeek's harness team ran a test every agent builder should see.
~300-message workload, 5 runs each:
One session, neutral prompt: p95 29.5 min. Subagent used in 0 of 5 runs.
Router + isolated workers: p95 15.7 min.
Models don't delegate on their own. Your harness has to.
Codex now has a built-in Google Cloud Vertex AI provider (0.165).
In config.toml:
model_provider = "gcp"
plus your GCP project and location. GPT-6.1 Terra joined the Vertex catalog in 0.166.
If your company buys AI only through its cloud bill, this is the path.
Gemini CLI deletes chat sessions older than 30 days by default.
The setting is general.sessionRetention in settings.json: enabled true, maxAge "30d".
Want more history? Set maxAge to "90d", or use maxCount to keep the N most recent. minRetention (1 day) is the floor.
Two new Codex CLI commands in 0.166:
codex doctor --network-only
Checks only config, connectivity and WebSocket. Good first step when Codex hangs behind a proxy.
codex queue --resume
Queue a message and resume an unloaded thread.
Also fixed: early compaction after resume.
Budgeting Gemini API spend for Q1? Read the dates on the pricing page.
• Gemini 4 Argon: $2/$10 per M tokens through Nov 30, $4/$20 from Dec 1
• Gemini 3.8 Flash: $0.75/$3.75 through Dec 31, then $1.50/$7.50
• Batch and Flex are half price
• Thinking tokens bill as output
Multi-hop RAG retrievers train on perfect histories, then run on their own mistakes.
New paper: by step 3 of 4-hop questions, 70% of states hold a wrong passage. It then finds the next right one 28% of the time vs 67% from a clean start.
Fix: train on states it really reaches.
Your AI agent just got its own mouse.
Vercel Labs open-sourced agent-desktop: computer use for coding agents via the OS accessibility tree, not screenshots. macOS, Windows, Linux. Headless by default, so it doesn't steal your pointer.
#AI #AIAgents #ComputerUse #OpenSource

