Thursday 29 October 2026
Copilot Fusion is GA and picks the model per prompt
Five bullets
The one thing
GitHub made Copilot Fusion generally available on all Copilot plans. You pick Fusion and an effort level instead of a model. On the Artificial Analysis Coding Agent Index, Fusion Max scored 60. That is within two points of Claude Code with Opus 5.5 Max.
GitHub reports Fusion got there at about half the average API cost per task. It also used about half the tokens and took about 40% less time. You pay each model's standard rate. Paid plans get a 30% discount on Fusion model costs for a limited time.
You choose Standard, High or Max effort. Fusion then picks a workflow per prompt: Single, Cascade or Critique. Cascade lets an efficient model draft, then escalates if a quality check fails. Critique has a second model review the draft before the first revises.
Picking models by hand is a small tax on every task. AT&T says its developers often forgot to switch back from a stronger model. Routing only helps if you can tell when it chose badly. The skill that grows is judging output per task, not by model name.
Run Fusion on tasks you already know well, then compare quality and cost with your usual model.
Try this · 15 minutes
See which models and workflows Copilot Fusion picks for your tasks
- Update your Copilot client in VS Code, the Copilot app, Copilot CLI or JetBrains.
- Open the model picker, select Copilot Fusion, and choose Fusion Standard.
- Give it one routine task and one hard task. Note which workflow it uses for each.
- Rerun the hard task on Fusion High. Compare the result with your usual hand-picked model.
Stories
- WATCHClaude Code background agents resend full transcripts every 30 secondsClaude Code sends each background agent's full transcript to its model about every 30 seconds, according to a Flamerix AI report. Its only output is a 3-5 word status line that no setting can disable. In one team's telemetry, that cost 16.4% of Claude Code spend, more than the main conversation. Anyone running parallel background agents should check their own OpenTelemetry data.
- TRYopencode's /goal keeps a session working until a reviewer agreesopencode v2.2.1 added /goal <objective> to the TUI and Desktop. The session keeps going until a reviewer confirms the objective is met. You steer it with /goal pause, /goal resume and /goal clear. Try it on a long task where you want a check before the agent stops.
- WATCHGemini 4 Carbon tops Artificial Analysis v5, with limited accessGoogle released Gemini 4 Carbon which ranks first on the Artificial Analysis Intelligence Index v5.0 with a score of 59. Google reports successful cheating on Terminal-Bench 4.0 fell to 0.0%. For now, only a subset of cyber defenders can use it. Teams picking models for long agent runs should wait for developer access and real harness results.
- WATCHViBench 2.0: top models fully pass only 49-65% of product buildsAgents build 17 products from an empty repo, one feature at a time. Nearly three-quarters of failed builds broke only where features intersect.
- TRYEffort-economics paper: start agents at low effort, then escalateHigher effort does not guarantee better results per task. Separately, h1 matched specialized theorem-proving workflows with a simple feedback loop.
- TRYSimon Willison says Opus 5.5 can now vibe code fun gamesHe built four browser games in HTML and JavaScript. One came from a single prompt in regular Claude chat, without Claude Code.
- WATCHAmp's macOS Satellites let the agent build and test Apple appsA fresh cloud Mac with Xcode boots in about 30 seconds, from $2 an hour. It is in beta behind a waitlist.
- WATCHVercel CI beta cut push-to-green wait 2-6x in Vercel's reposIt runs Turborepo tasks tagged ci on every push. It only runs the CI affected by your changes.
Safe to ignore
53 sources read. 91 items. 13 shortlisted. 5 bullets.