GitHub Copilot retires 4 more models Oct 2; Kimi K3 costs 3x K2.7
Four models leave every Copilot surface, and their replacements cost up to 3x. Plus a gh CLI signing key expires Sept 5 — apt/rpm installs break without it.
Copy markdown
The Oct 2 list: four models, every surface
Gemini 3.5 and 3.6 Flash both fold into 3.8 Flash, Kimi K2.7 Code moves to Kimi K3, and Claude Opus 4.7 maps to Opus 5 — pulled from Copilot Chat, inline edits, ask and agent modes, and code completions. If you've pinned any of these in a workflow or CI step, it stops resolving on October 2.
The migration hides a 3x bill
A replacement name isn't a free swap. Kimi K3 lists at $3.00/$15.00 per million tokens against K2.7 Code's $0.95/$4.00 — roughly triple — so any agent loop pinned to Kimi gets materially pricier the moment you cut over. Audit pinned models and re-check billing before the deadline, not after.
Third retirement wave in five weeks
Oct 2 caps a fast purge: Sept 1 already retired six models (Gemini 3.1 Pro, Claude Opus 4.5 and 4.6, Sonnet 4.5, most Sonnet 4.6 access, Raptor mini), and MAI-Code-1-Flash drops Sept 10. Copilot's model menu is turning over fast enough that anything hard-coded needs a standing review, not a one-time fix.
Act by Sept 5: gh's Linux signing key expires
Unrelated but urgent — the PGP key signing GitHub CLI's APT/RPM packages expires September 5. Anyone who installed gh before April 8 must refresh the keyring or apt/dnf will refuse to install or update it. Re-fetch githubcli-archive-keyring.gpg (or re-add the rpm repofile); Homebrew, macOS, Windows and source builds are unaffected.
Gemini 3.8 Flash joins Copilot; admins get a default-model lock
The same changelog batch adds Gemini 3.8 Flash to Copilot — GitHub flags it as strong on complex terminal-based coding — and lets enterprise-managed settings pin any model as the org-wide default for new conversations. Handy if you're standardizing a team on one model ahead of the deprecations.
Elsewhere: Anthropic cuts Fable 5.1 cache reads 75%
Claude Fable 5.1's cache-read price dropped from $1.00 to $0.25 per million tokens — a reported 25-45% total saving on cache-heavy agentic workloads. If you run long, stateful Claude sessions (Claude Code included), the economics just moved in your favor.