GitHub adds model choice to Copilot Enterprise, opening support for OpenAI, Anthropic, and Meta models

GitHub said Tuesday it will let Copilot Enterprise customers choose among multiple large language models for code completion and chat, including OpenAI’s GPT-4o via Azure OpenAI Service, Anthropic’s Claude 3.5 Sonnet on Amazon Bedrock, and Meta’s Llama 3 family through Azure AI Foundry. The feature, available starting today in public preview, adds per-repository and per-organization model policies with logging and cost controls, while keeping Copilot Enterprise pricing at 39 dollars per user per month.

GitHub framed the change as a response to enterprise demand for cost control, data residency, and regulator transparency as the EU AI Act’s general purpose AI obligations take effect through 2025 and into 2026. The company said model choice helps customers align workloads to compliance requirements and performance needs without changing developer workflows.

Model routing includes baseline usage caps, regional pinning, and audit trails in GitHub Enterprise Cloud. Administrators can set defaults, then allow teams to override for specific repositories or languages, for example selecting Claude for long-context refactoring or Llama for on-prem inference.

Pricing passes through the model provider. GitHub said Azure OpenAI GPT-4o will be billed at the service’s list price, which Microsoft lists at 5 dollars per one million input tokens and 15 dollars per one million output tokens. For Anthropic on Bedrock, AWS lists Claude 3.5 Sonnet at 3 dollars per one million input tokens and 15 dollars per one million output tokens. Open source Llama deployments will reflect the chosen managed host or customer infrastructure, and GitHub will surface estimated throughput and spend before enabling a route.

GitHub will publish evaluation dashboards that show pass rates on HumanEval, MBPP, and SWE-bench for each supported model across five languages, with repository-level acceptance metrics available to admins. The company said internal tests show variation by codebase and tool use, and it encouraged customers to A or B test models before enforcing policies. Results will be updated monthly and attributed to upstream model version changes.

“Enterprises want model choice without fragmenting developer experience,” GitHub CEO Thomas Dohmke said in a statement. He added that prompts and code sent to providers are not used to train foundation models, consistent with GitHub’s existing data handling commitments. GitHub said customer content routing honors each provider’s private on-by-default settings on Azure and Bedrock, with optional zero-retention where available.

Competition is moving in a similar direction. JetBrains added multi-model options to AI Assistant in 2024 using OpenAI and local models. AWS CodeWhisperer integrates with Amazon Bedrock for enterprise model governance. Microsoft’s broader Copilot stack in Microsoft 365 has introduced admin controls for model and modality by department, according to Microsoft.

For the broader industry, model choice embedded in daily developer tools reduces lock-in, makes price and performance a procurement variable, and gives legal teams clearer accountability under emerging AI governance rules. It also pushes model providers to compete on measurable coding accuracy, latency, and total cost, not only on marquee benchmarks.