GitHub transitioned all Copilot plans to usage-based billing on June 1, 2026. This change replaced premium request units with GitHub AI Credits where one credit equals one cent. Code completions and next edit suggestions remain included without consuming these credits. Agentic workflows consume tokens through input, output, and cached processing. Organizations using intensive agentic sessions will see higher costs because those sessions re-read repositories and carry large context windows. A single ambitious agent session that consumes 4 million input tokens and 400,000 output tokens costs $18, which is equivalent to 1,800 AI credits for any enterprise organization. Enterprise seats cost $39 per month and include 3,900 credits, while Business seats cost $19 per month and provide 1,900 credits. You should monitor your token consumption if your team adopted agentic workflows during the summer. Most Enterprise customers used promotional allotments of 7,000 credits during June, July, and August. These allotments shrink to the standard 3,900 credits on September 1, 2026. Earlier deployments like Accenture’s 50,000 developers saw an 84% increase in successful builds when using Copilot. This distinction is separate from Microsoft 365 Copilot, which is a different product for Office users.
Poolside offers on-premises weight delivery
Nvidia signed a $6 billion non-exclusive license for Poolside’s Model Factory software in August 2026. This deal includes a $1 billion investment at a $12 billion pre-money valuation and brings 100 Poolside engineers to Nvidia. Poolside builds foundation models like malibu, which handles code generation and refactoring, and point, which provides code completion. The company markets full model weights for on-premises and air-gapped government deployments. These contracted deployments allow companies to avoid the per-token cloud fees that drive up GitHub Copilot costs. A test of the laguna-xs-2.1 model showed it successfully planned and executed a two-scenario test suite in five and a half minutes. The laguna-s-2.1 model fails to follow instructions and produced twenty-one invalid JSON replies across three different schemas. One engineer found that while laguna-xs-2.1 works for reading pages, it is significantly slower than Groq, which answers in a third of the time.
Evaluating model ownership and orchestration
Enterprise teams must decide between purchasing a model or orchestrating an existing application layer. Poolside provides full weights for users who require vendor accountability and hardware optimization. Eigent provides an Apache-2.0 multi-agent workspace that connects to various compatible endpoints. You can use Eigent for cross-functional multi-agent work or pair it with OpenHands for asynchronous software-agent infrastructure.
| Feature | GitHub Copilot Enterprise | Poolside (Contracted) |
|---|---|---|
| Monthly Seat Price | $39 | Contact sales |
| Included AI Credits | 3,900 | None (No per-token fee) |
| Deployment Type | Cloud-based | On-premises / Air-gapped |
| Model Weights | Managed by GitHub | Full weight delivery |
Procurement teams should evaluate whether they need downloadable weights or merely deployable models. They must also consider if they need to own the upgrades, incident response, and evaluation processes. Poolside is the correct choice for those requiring contracted full-weight delivery and enterprise hardware integration. What happens to the remaining budget when models achieve parity? Select an orchestration layer like Eigent if your team prefers to change model endpoints without replacing the entire workspace.
