FOR THE BUILDERS, THE MAKERS & THE SMALL TEAMS WITH BIG IDEAS ✳
AI Coding · News Analysis

GitHub Copilot Adds GPT-6.2 Astra and Sol: Evaluate Coding and Review Separately

GitHub announced GPT-6.2 Astra and GPT-6.2 Sol for Copilot on November 17, 2026, adding two candidates for agentic coding and code review. For small teams already using Copilot, the practical next step is a limited comparison on familiar work. Evaluate implementation and review separately: a model that produces an acceptable patch does not automatically provide the most useful critique of another developer’s changes.

ai-coding: Official update → Team impact → Adoption decision. Editorial workflow illustration.
Source-based News Analysis. Announcement published November 17, 2026; official sources accessed November 18, 2026. No hands-on model, coding, review, or performance tests were performed. Vendor performance claims are attributed to GitHub. Evaluation examples and adoption criteria are editorial proposals. Supporting documentation is mutable.

1. Two new models, with gradual availability

The GitHub announcement describes both models as generally available. It positions GPT-6.2 Astra for agentic coding and terminal work, and GPT-6.2 Sol for complex agentic coding and in-depth code review. These are vendor descriptions of their intended uses, rather than independently established differences in capability.

GitHub lists availability for Copilot Pro+, Max, Business, and Enterprise users. Its supported interfaces include Visual Studio Code, Visual Studio, Copilot CLI, the Copilot coding agent, the Copilot app, github.com, GitHub Mobile, JetBrains IDEs, Xcode, and Eclipse. The announcement explicitly says rollout will be gradual, so general availability does not establish that every eligible account already has access.

For an existing Copilot team, this creates an opportunity to compare models inside a familiar interface. Confirm that the intended model appears where you actually work before scheduling an evaluation. This article verifies published information; it does not report hands-on access or model test results.

2. The review claims need a local baseline

GitHub says early testing found that Astra caught significantly more code review issues than GPT-6 Astra and completed agentic coding tasks using noticeably fewer tokens. For Sol, GitHub reports nearly twice as many real review issues found as GPT-6.1 Sol while keeping findings accurate.

Those comparisons use different predecessor models. They do not establish whether Astra or Sol performs better against the other. The announcement also does not supply enough methodological detail to reproduce the reported gains: it gives no task collection, sample size, or numerical accuracy breakdown.

For a small team, more findings help only when they identify consequential defects without creating excessive investigation work. A longer review can contain repeated observations or suggestions that do not affect correctness. Our editorial recommendation is to count distinct, verified defects and record the effort required to assess every finding. Treat GitHub’s claims as reasons to investigate, with your current workflow providing the baseline.

3. Keep the comparison tied to one interface

GitHub’s supported-model documentation separates model support by client and plan. It also lists minimum IDE versions for recent models, with GPT-6.2 Astra and Sol entries marked TBD when accessed. The documentation cautions that missing minimum-version information does not mean all older versions are supported.

Our proposed pilot therefore starts in one interface that the team already uses. Record the client and extension versions, selected model, repository commit, prompt, and available tools. Give each candidate the same starting files and task instructions. Use separate sessions so one candidate does not inherit another’s answer.

If settings or available tools differ, preserve that distinction in the result. You may still learn which complete workflow suits your team, but you cannot attribute every difference to the underlying model. The criteria in our coding-assistant evaluation guide can help keep the comparison focused on work you regularly perform.

4. Test implementation with an observable requirement

For coding, choose a bounded maintenance task with a clear acceptance condition. One proposed example is fixing a pagination bug where the last page repeats an item. Supply the expected behavior and relevant repository context, then ask each candidate to implement the change and explain its verification.

Inspect whether the patch addresses the requirement, changes unrelated behavior, or needs manual repair. Run the relevant checks yourself and retain their output. An agent’s statement that tests passed is a claim to verify; an unexecuted test command is a proposed next step.

Record completion time, interventions, and any failed attempts alongside correctness. If token usage is observable in your workflow, record it consistently, without treating a lower count as proof of a better result. Our guide to regression testing AI-generated code explains how to turn the original failure into an acceptance check. These are suggested evaluation steps, not experiments performed for this article.

5. Evaluate review without rewarding comment volume

For review, use a separate set of small historical changes whose defects your team understands. Present the changes without the later fix or explanatory discussion. Include a change you have carefully checked and believe is correct, so the exercise also exposes unsupported objections.

Have a maintainer classify each finding as a confirmed defect, a useful question, a duplicate, or an unsupported concern. Record missed known defects separately. Ask for the affected code and a concrete explanation of the failure; reproduce the behavior where practical before crediting a finding.

A small sample can reveal useful differences, but it cannot establish a general performance ranking. Repeat promising results across different tasks before changing a default. Keep final acceptance with a human reviewer, using the evidence requirements in our AI code review workflow. A model’s confident explanation should not substitute for examining the actual change.

6. Change defaults only after repeatable results

Try Astra and Sol when your existing workflow has a specific weakness you can observe: implementation tasks need repeated corrections, or reviews miss defects that a maintainer later catches. Define the desired improvement before running the comparison, and retain your current model as a baseline.

Adopt a candidate for the task where it repeatedly produces acceptable results with manageable verification effort. If coding improves but review does not, use that narrower conclusion. If results remain mixed, continue with the current workflow and revisit after collecting more representative examples. GitHub’s release expands the available choices; the evidence needed to change your team’s defaults still comes from your own work.

Sources & further reading

Primary references used for product and technical information. Features and terms can change; check the provider before committing.

No affiliate links are included in this article. Read our disclosure.