GitHub announced GPT-6 Sol and GPT-6 Luna for GitHub Copilot on September 22, 2026. The two models give developers another selection to evaluate: GitHub positions Sol for interactive and agentic coding that benefits from careful, multistep validation, while Luna is intended for smaller, faster, and more cost-efficient tasks.
The rollout is gradual, so an eligible user may not see either model immediately. The change also comes with an important documentation caveat: GitHub’s supported-models page lists GPT-6 Sol and GPT-6 Luna as generally available, while its model-comparison page still labels both as “coming soon.” Developers and administrators should therefore confirm availability in the specific Copilot client and organization where they plan to use the models.
What changed in GitHub Copilot
GPT-6 Sol joins GPT-6 Astra in the GPT-6 family available through GitHub Copilot. According to GitHub’s announcement, Sol is designed as an all-round option for interactive and agentic coding workflows. That positioning makes it a candidate for tasks that require multiple steps, iterative reasoning, or validation across a larger change.
GPT-6 Luna is positioned differently. GitHub describes it as a lightweight model for smaller and faster coding tasks, with lower listed token rates than GPT-6 Sol in both the default and long-context pricing tiers. The pricing difference may make Luna useful for routine interactions or workloads where response speed and consumption efficiency matter more than choosing the model aimed at more involved workflows.
These descriptions are product positioning rather than independent performance benchmarks. The supplied documentation does not establish that Sol writes better code, debugs more effectively, or produces better agentic outcomes than Luna in real-world Copilot use.
Context and reasoning options
The supported-models documentation lists both GPT-6 models with a 1-million-token context option and configurable reasoning. Those capabilities are not documented as universal across every Copilot integration.
- The 1-million-token context option is documented for Visual Studio Code and Copilot CLI.
- Configurable reasoning is documented for Visual Studio Code, Copilot CLI, and the Copilot cloud agent.
- Availability and behavior can still depend on the client, plan, rollout status, and model policy.
Developers should treat these as client-specific capabilities rather than assuming that selecting the same model in every editor or Copilot surface will expose identical controls. The supported-models page also contains tentative or “TBD” entries for some client and version combinations.
Plans, clients, and organizational controls
GPT-6 Sol is listed for Copilot Pro+, Max, Business, and Enterprise plans. GPT-6 Luna is listed for Copilot Pro, Pro+, Max, Business, and Enterprise plans. A qualifying plan does not guarantee immediate access because GitHub is rolling the models out gradually.
The announcement identifies model-picker availability across Visual Studio Code, Visual Studio, Copilot CLI, the Copilot cloud agent, the Copilot app, github.com, GitHub Mobile for iOS and Android, JetBrains, Xcode, and Eclipse. The broader client documentation lists both models across multiple Copilot surfaces, but some minimum-version details remain tentative or undocumented. Checking the model picker in the intended client is the most direct way to determine whether access has reached a particular account.
Business and Enterprise administrators can manage GPT-6 model access through Copilot model policies. Under the documented default model enablement behavior, new models are enabled automatically unless the global default or the specific model has been disabled. Organization policy can therefore affect visibility even when a user’s plan is otherwise eligible.
Pricing and usage considerations
Copilot usage is billed according to the selected model and the amount of input, output, cached, and cache-write token usage. The supplied Copilot billing reference defines one AI credit as $0.01 USD and lists GPT-6 Luna at lower token rates than GPT-6 Sol in the default and long-context tiers.
Those rates do not translate into a universal per-task or monthly cost. Actual consumption depends on the token mix, context size, reasoning settings, and workload. The extended context and reasoning options may increase AI-credit consumption. Teams evaluating the models should compare their own usage patterns rather than assuming that the lower listed rate will produce a fixed percentage of savings.
The supporting pricing reference is hosted on docs-internal.github.com rather than the standard public docs.github.com domain. It is included here as the supplied first-party documentation reference, but the domain difference is worth noting when teams validate billing information.
What developers and administrators should do
- Check the model picker. Look in the Copilot client where the model will be used. Gradual rollout means access may differ between accounts, clients, or integrations.
- Match the model to the task. Start with Sol for more involved interactive or agentic work that benefits from multistep validation, and consider Luna for smaller or faster tasks. Treat this as a documented usage distinction, not a performance ranking.
- Review organization policies. Business and Enterprise administrators should check Copilot model policies if users cannot see either model.
- Verify client and extension requirements. Consult the supported-models documentation for the relevant IDE or Copilot surface, especially where minimum-version entries are tentative or marked TBD.
- Review billing before expanding use. Compare token categories, context length, reasoning settings, and expected workload before estimating the effect on AI-credit consumption.
Availability and safety caveats
The model-status discrepancy in GitHub’s documentation should not be ignored. The supported-models page lists GPT-6 Sol and GPT-6 Luna as generally available, but the model-comparison page still describes them as coming soon. Combined with the gradual rollout and incomplete client/version details, this means “available” should be understood as dependent on the relevant account, plan, client, and policy rather than as simultaneous access for every Copilot user.
GitHub’s supported-models documentation states that default AI-model prompts and completions pass through Copilot content filters for harmful, offensive, or off-topic content, as well as public-code matching when enabled. OpenAI’s supplied GPT-6 Astra system-card material also includes model-level safety evaluations for GPT-6 Sol and GPT-6 Luna. Those evaluations should not be treated as GitHub Copilot coding benchmarks, a guarantee of production behavior, or a complete description of safeguards in every Copilot integration.
For now, the practical change is a broader model choice inside GitHub Copilot, not a documented promise of superior coding results. Teams can evaluate Sol and Luna against their own workflows once the models appear in the relevant client, while keeping rollout status, policy controls, and usage-based billing in view.
