Frontier AI
Requests to gemini-3.7-flash now go to gemini-3.8-flash, and requests to gemini-3.5-flash now go to gemini-3.6-flash.
Why it matters. Outputs can shift with no change on your side, so update the model string and rerun your evals.
Source: Gemini API deprecations
Organization owners can bill reviews from licensed members to the organization, which needs AI Credits paid usage, instead of each member's quota.
Why it matters. Under the default setting a review fails once that member's quota runs out, and the new option is under Copilot policies.
Source: GitHub changelog
Set autoCompactWindow in subagent frontmatter or --agents definitions so a subagent auto-compacts before it reaches the main conversation's window.
Why it matters. Use it to keep context heavy subagents, such as file readers or reviewers, inside a smaller window than the main session.
Source: Claude Code v2.1.296 release
Set service_tier to ultrafast with gpt-6.1-sol to cut the time between output tokens, open to all API users subject to rate limits.
Why it matters. OpenAI recommends a persistent WebSocket connection, since per request network overhead can eat into the latency gain.
Source: Ultrafast mode guide
Every package now requires Python 3.11 or newer, and installs on Python 3.10 resolve to 2.54.0 or earlier.
Why it matters. Projects still on Python 3.10 stop receiving new releases without any error, so move to 3.11 to keep upgrading.
Source: Pydantic AI v2.55.0 release
In beta, enable workflows in the agent's multiagent field, and the agent writes a program that runs many agents in phases in the background.
Why it matters. Jobs like reviewing hundreds of documents no longer need to fit in one conversation, and you follow each run through workflow_run events.
Source: Workflow runs