Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Playbooks · Coding

GPT-6.2 Astra and Sol in GitHub Copilot

Written November 20, 2026

GPT-6.2 Astra and Sol are generally available in GitHub Copilot for agentic coding and code review, billed at provider list pricing, with a gradual rollout.

Two new models, generally available

GitHub's November 17 changelog post announces that OpenAI's GPT-6.2 Astra and GPT-6.2 Sol are now generally available in GitHub Copilot. Both models are, in GitHub's description, designed for agentic coding and code review, with stronger coding judgment than earlier GPT-6 models. That positioning matters for how a Copilot user should think about the switch: these are not pitched as general chat upgrades, but as models for the two jobs where an agent reads a codebase, acts on it, and then has its work judged — writing code through an agent, and reviewing code before it lands. A user deciding whether to move needs three separate things kept apart: what GitHub claims the models do better, who can actually select them today, and what they cost. The first is an early-testing claim, the second is an availability fact, and the third is a billing rule, and the post treats them differently.

What GitHub claims from early testing

The performance case in the changelog post is framed as early testing, and it should be read that way — as GitHub's early-testing claims, not as measured results from your own repositories. GitHub reports that in early testing Astra caught significantly more real issues in code review than GPT-6 Astra, and did so while using noticeably fewer tokens. For Sol, GitHub reports that in early testing it caught nearly twice as many real issues as GPT-6.1 Sol while keeping its findings accurate. Those are strong claims, and the honest way to use them is as a reason to run your own comparison, not as the comparison itself. Code review is exactly the kind of task where a model's hit rate depends on the codebase, the review style, and what counts as a real issue, so a claim earned on GitHub's test set may not transfer at the same size to yours. The practical move is the one used for any new Copilot model: pick a small set of representative reviews and agentic coding tasks, run them on Astra or Sol against the model you use now, and score the finished work — issues caught, false alarms, and total cost per completed task.

Who gets them, and where

Availability is the firmer ground. The post states that GPT-6.2 Astra and GPT-6.2 Sol are available to Copilot Pro+, Max, Business, and Enterprise users, selectable in the model picker across VS Code, Visual Studio, Copilot CLI, the coding agent, the Copilot app, github.com, GitHub Mobile, JetBrains, Xcode, and Eclipse. The rollout is gradual, so a model missing from your picker today is not evidence that your plan is excluded — the post's own advice is to check back. That breadth of surfaces also means a switch is rarely one setting: a developer may meet Astra first in an editor picker, while the same organization's coding agent and code review flows draw on the same model policy, so a team evaluating the models should test them on the surface where they will actually spend, not just the one where the picker is easiest to reach. The model-testing shape in Test Gemini 4 Carbon in Copilot before Dec 1 — score per completed task rather than per token — applies here unchanged, minus that page's December price deadline.

What they cost, and what an admin controls

On billing, the post states that both models are billed at provider list pricing under usage-based billing. There is no promotional window in this announcement and no rate card in the post itself, so the cost question for a team is not a December-style deadline but a standing one: usage-based billing means every agentic run and every review draws metered spend, and a model that uses fewer tokens on the same task — as GitHub claims Astra did in early testing — changes the per-task total even at the same list price. Control sits with administrators on the organization plans: Business and Enterprise admins manage access to the GPT-6.2 models through the model policy in Copilot settings. Under default model enablement, new models are automatically enabled unless an administrator has turned off the global default or explicitly disables these models — so on a Business or Enterprise plan, the models may already be selectable by your developers before anyone has made a deliberate decision about them. An admin who wants an evaluation before open use should set the policy first; a developer who cannot see the models should check the policy before assuming the gradual rollout is the reason.

What to do

Check the model picker on the surface you use most; if Astra or Sol is absent, check back rather than concluding anything, and on a Business or Enterprise plan check the model policy in Copilot settings as well. Then run a small, scored comparison on your own reviews and agentic tasks before switching defaults — GitHub's early-testing claims say the models caught more real issues with fewer tokens, and your own tasks are the only test that can confirm whether that holds on your code. Admins: decide the model policy deliberately, because under default enablement the decision has, in effect, already been made for you.

Related playbooks

Sources

Back to: Coding playbooks hub · Also: Test Gemini 4 Carbon in Copilot before Dec 1