Five models joined the Copilot picker in nineteen days: Claude Opus 5 on 26 July, Kimi K3 on 6 August, MAI-Code-1.1-Flash on 11 August, Gemini 3.7 Flash on 13 August, Grok 4.6 on 14 August.
None of them arrived with a comparison to the others, because none of them was built to compete on the same axis. Read together, they sort into three clear jobs rather than one ranked list.
The five, sorted by what they are for
Not a ranking — a map of intent, which is more useful than a score when the models were not built to answer the same question.
- Claude Opus 5 — the top-tier, expensive-by-design option. Reach for it when the task itself is hard: an ambiguous refactor, an unfamiliar codebase, a decision that needs judgement rather than pattern-matching
- Grok 4.6 — pitched specifically at long, multi-step agentic runs with sustained reasoning. The one to try when a task keeps failing because the agent loses the thread over many steps, not because any single step is hard
- Gemini 3.7 Flash — fast, and gated behind a policy whose name literally includes "Preview". Good for volume work you are willing to re-evaluate if its behaviour shifts under you
- Kimi K3 — the one with unusual, transparent-ish token pricing ($3 / $15 / $0.30 per million input, output and cached input) rather than a premium-request multiplier. Cost-predictable for repeated work over the same repository, because of the cached-input discount
- MAI-Code-1.1-Flash — the cheapest, with native vision and a 73% lower list price than its predecessor. Auto-selected for Free and Student users, which tells you what GitHub thinks the default volume model should be
What they share, and it is the part that costs you time
Every one of the four newer models — Kimi K3, MAI-Code-1.1-Flash, Gemini 3.7 Flash, Grok 4.6 — is off by default for Copilot Business and Enterprise, requiring an administrator to enable a specific policy before anyone can select it. Claude Opus 5 followed the same pattern in July.
That is five separate policy decisions sitting in an admin console right now, most of them probably still unanswered, because nobody scheduled time to answer five questions that arrived nineteen days apart from five different reasons.
A policy that survives the sixth model
Do not evaluate each one individually — that does not scale at this cadence. Separate two decisions instead.
Enabling is cheap: it puts a model in the picker for anyone who wants to try it, and the cost is mostly about data handling and unpredictable spend, not quality. Enable generously unless a specific model fails a stated criterion — a pricing mechanism you cannot forecast, or a data-handling concern.
Adopting — making a model your team's default — deserves the real evaluation: the same three saved tasks (a refactor, a bug with a reproducible test, an explanation of code you already understand), run against your own codebase, twice, before you trust a win.
A starting default, if you want one
For most teams doing ordinary work: MAI-Code-1.1-Flash or Kimi K3 as the default for routine tasks, because both are cheap and their pricing is at least legible. Reach for Claude Opus 5 deliberately for the specific task that is actually hard, not as a general upgrade. Try Grok 4.6 the next time an agent loses the thread on a long task rather than swapping your default to it outright. Treat Gemini 3.7 Flash as evaluation-only until "Preview" leaves its policy name.
That is a starting point, not a verdict — the point of separating enable from adopt is that your own three-task test, not this list, should make the adoption call.