GitHub made Copilot Fusion generally available on October 28, the first day of GitHub Universe, in VS Code, the Copilot app, Copilot CLI and JetBrains IDEs. Instead of picking a model per task, you pick an effort level and Fusion decides whether a prompt goes to one model, a cascade or a draft-and-critique loop, then bills each model it used at that model’s standard rate. The same day GitHub announced that GitHub Enterprise now includes Copilot for every licensed user, with no monthly credits per seat and AI usage metered against the enterprise budget from the first request, for enterprises working with a GitHub or Microsoft account team. Fusion changelog · Enterprise changelog · Universe post
S5 Labs has not run Fusion. The benchmark figures below are GitHub’s or Artificial Analysis’s, and the prices and plan rules come from GitHub’s docs as read on October 28. Where the changelog and the docs disagree, we say so.
How Fusion routes a prompt, and who gets which level
Fusion combines Project HydraFusion’s multi-model orchestration with Copilot’s auto model selection. For each prompt it picks one of three patterns. Single sends the prompt to one model. Cascade has a model draft a response, then a judge either accepts the draft or escalates to a stronger model. Critique has one model draft, one or more reviewing models critique it, and the drafting model revise. GitHub’s descriptions of critique differ: the changelog says “another model reviews it”, the docs say “multiple models independently critique it in parallel”, and the September HydraFusion post describes a single read-only critic from a different model family. Cascade and critique take longer because they add review calls, and a session can mix patterns, since the choice is made per prompt. Fusion changelog · Fusion docs · HydraFusion post
The effort level sets the pool. Standard “balances cost, quality, and latency” and picks one model per task. High is “balanced orchestration with moderate token savings” for well-scoped work such as a complex bug or a multi-file change. Max is “frontier capability, with token efficiency still considered.” A fourth level, Fast, exists only in VS Code inline chat and cannot be selected. Fusion docs
The changelog says Fusion is “available on all Copilot plans.” The docs table is narrower:
| Effort level | Free, Student, Pro | Pro+, Max | Business, Enterprise |
|---|---|---|---|
| Standard | Included | Included | Included |
| High | Not included | Included | Included |
| Max | Not included | Included | Included |
Free, Student and Pro get Standard only; High and Max need Pro+, Max, Business or Enterprise. Fusion docs
What a Fusion request costs
There is no charge for Fusion itself. You pay for each model it uses, at that model’s standard rate, so a High or Max task that runs a draft, a critique and a revision consumes more AI credits than the same prompt at Standard. GitHub says the main conversation stays on one model where possible to preserve cached tokens, and publishes no multipliers. Through November 30, 2026, paid plans get a 30% discount on model costs when using Fusion; the changelog says only “for a limited time”, and the end date appears in the docs alone. Auto model selection, which Fusion replaces wherever it shows up in the picker, carries a 10% discount. GitHub does not say how either discount interacts with budgets or whether the 30% will be extended. Fusion changelog · Fusion docs · Auto model selection docs
GitHub does not publish the pool, only that it mixes “versatile models” with “models with stronger reasoning” and changes over time. Artificial Analysis observed Claude Opus 5.5, Claude Sonnet 5.5, GPT-6 Astra and GPT-6 Sol in its runs, and after any response you can hover over the effort level name to see the models used. GitHub’s rates per million tokens for those models, where 1 AI credit is $0.01:
| Model | Input | Cached input | Output |
|---|---|---|---|
| Claude Opus 5.5 | $4.00 | $0.20 | $20.00 |
| Claude Sonnet 5.5 | $2.00 | $0.10 | $10.00 |
| GPT-6 Astra (up to 272K) | $10.00 | $1.00 | $50.00 |
| GPT-6 Sol (up to 272K) | $2.00 | $0.20 | $10.00 |
| MAI-Code-1.1-Flash | $0.20 | $0.02 | $1.20 |
Above 272K input tokens Astra moves to $20 / $2 / $75 and Sol to $4 / $0.40 / $15; the Anthropic models list no long-context tier. The Opus 5.5 row matches Anthropic’s list price, so Copilot passes it through unchanged. Requests processed under FedRAMP enforcement or data residency carry a 10% increase on AI credit consumption. Models and pricing · Artificial Analysis · FedRAMP models docs · Data residency docs
Fusion uses only models enabled in enterprise and organization policies. Administrators can see which models it picks from and disable a model for Fusion without disabling it for everyone; if none of Fusion’s models is available to you, Fusion does not appear in the picker. A discarded draft does not undo workspace edits, so review changes before committing. Fusion changelog · Fusion docs
GitHub’s benchmark claim, and what Artificial Analysis backs
GitHub’s changelog says: “In the Artificial Analysis Coding Agent Index, Copilot Fusion Max scored 60, within two points of Claude Code with Opus 5.5 Max. It delivered that result at about half the average API cost per task, using about half as many tokens and taking about 40% less time per task.” Artificial Analysis published its own article on the run. The two agree on some numbers and not on others.
| Claim | GitHub | Artificial Analysis |
|---|---|---|
| Index score | 60 | Not stated; “2 points below” Claude Code with Opus 5.5 |
| Cost per task | “about half” | $5.59 against $11.42, about 49% |
| Tokens per task | “about half as many” | 7.2M, about half of Opus 5.5 in Claude Code |
| Time per task | “about 40% less” | Not stated |
| Setting tested | Max | Max, in Copilot CLI |
Artificial Analysis’s breakdown: Terminal-Bench 4.0 71, first place; FrontierCode 51, third behind Claude Code with Opus 5.5 max at 56 and Devin Fusion at 54; SWE-Atlas-QnA-AA 57. Devin Fusion, which pairs Opus 5.5 max with GPT-6 Luna high, scored 62 at $8.00 per task. Index v2.0 weights the three benchmarks equally. Only Max was benchmarked, so none of this describes Standard, the one level every plan gets. The only evidence on Standard is a vendor-supplied quote from AT&T’s Hari Pokhrel, who says it improved token efficiency by picking the right model for each task, so developers no longer have to remember to switch models. Fusion changelog · Artificial Analysis
Local models: the routing stays in the cloud
A second entry adds local models to Fusion as an experimental preview (the docs heading says public preview). On supported Windows devices a managed MAI Code 1.1 Flash download runs through the Windows ML provider, enabled per client (/experimental on then /model in the CLI, a “Local model auto routing” experiment in the app, chat.copilot.localModel.enabled in VS Code). Bring-your-own-key routing in the CLI works on Windows, macOS and Linux with MAI Code 1.1 Flash or Qwen3.8 Flash Next served from a local OpenAI-compatible server. Local models are used only at Standard. Local models changelog · Fusion docs
Two lines in the local-models changelog matter more than the setup: “Local inference doesn’t make your session offline,” and routing decisions are still made in the cloud. Neither the changelog nor the docs say what the router sees or how local calls are billed, and the changelog warns that local models may carry different safety protections. Local models changelog · Fusion docs
The managed model is not small. Microsoft’s October 7 post describes MAI Code 1.1 Flash as a 137B-total, 6.8B-active mixture of experts quantized to 53 GB at about 3.3 bits per weight, tested on a Surface Laptop Ultra with up to 128 GB of unified memory, where peak memory hit 75.5 GB at 256K context. Microsoft’s reported figures put the on-device build at 70.80% on SWE-Bench Verified against 72.6% for the cloud model, and 66.29% on Terminal-Bench 2.1 against 62.9%. No license or minimum hardware is stated. Qwen3.8 Flash Next ships under the Qwen Community License, which needs a separate license for a model-as-a-service or “AI work-assistant” business; internal tools are carved out. Microsoft post
Copilot with GitHub Enterprise: metered from request one
Copilot has billed by GitHub AI Credits, consumed by tokens at each model’s published API rate, since June 1. Universe adds a billing model with no credit pool at all. Enterprises on GitHub Enterprise Cloud working with an account team can move to “Copilot access”: a $0 seat for every licensed user, $10 per user per month Standalone for users without a GHEC license, the same features and models as Copilot Enterprise, and “no included usage allowance to track or top up.” Self-serve is “currently unavailable”, with dates “in the coming weeks.” Moving is voluntary, but an enterprise picks one model: Copilot access cannot coexist with Business or Enterprise seats in the same enterprise. April billing post · Enterprise changelog · Plans docs · What’s changing
| Plan | Price | Monthly AI credits |
|---|---|---|
| Pro | $10 | 1,500 |
| Pro+ | $39 | 7,000 |
| Max | $100 | 20,000 |
| Business | $19 per user | 1,900 per user |
| Enterprise | $39 per user | 3,900 per user |
| Copilot access | $0 seat with GHEC, or $10 Standalone | None |
Pools reset at 00:00 UTC on the 1st and unused credits are forfeited; Free and Student amounts are not on the plans page. Plans docs · Billing docs
What stops at zero, in GitHub’s words: “Code completion and next edit suggestions keep working even with no budget available. Copilot Chat, Copilot code review, and Copilot in the cloud require available budget to use. An exhausted budget never removes access.” The billing docs list Chat, CLI, cloud agent, Spaces, Spark and third-party coding agents as credit consumers and say a $0 user-level budget blocks them immediately. Spending limits apply from the first metered use under Copilot access, and only after the pool runs out on Business and Enterprise. The “AI credits paid usage” policy is on by default and, under Copilot access, must stay on: disabling it blocks every credit-consuming feature. Enterprise changelog · Billing docs
The rest of Universe, briefly
- Agentic Workflows are GA on Free, Team and Enterprise. The default Copilot CLI engine needs a Copilot license, workflows run in a sandboxed container with read-only permissions by default, and
gh aw forecastnow estimates AI credit spend. No pricing. Agentic Workflows Our longer read covers the per-run credit caps and server-to-server auth. - Copilot app: Projects in public preview, spanning several repos with as many agents as the work needs; computer use in public preview on macOS and Windows, with approval prompts and an org-level off switch. Copilot app
- Agent Merge: “Watch and fix” in the merge box repairs failing Actions runs, merge conflicts or both in a cloud environment, on paid plans. No preview or GA label, no billing statement. Agent Merge
- Automations: 24 new triggers, 20 new tools, custom MCP servers, a REST trigger API and AI credit limits per run, day, week or month. Automations
- Code review Max in public preview runs several review agents from different providers plus an aggregator and uses more credits than Lite or Balanced; approvals are GA, off by default. Code review
- Cloud sandboxes are GA on paid plans, with sandbox billing starting November 2026. Connected cloud
- SDK: GitHub Copilot SDK becomes Microsoft Copilot SDK in phases, with no immediate changes and no timeline; server-to-server authentication (October 27, public preview) bills an app’s usage to the owning org. SDK · Server-to-server auth
- Lockfiles: Actions workflow lockfiles are generally available. Our preview article now has an October 28 update, including a correction on reusable workflows.
What to check
Standard against your old default. Standard is the only level Free, Student and Pro get, and the only level the AT&T quote covers. Run a week of ordinary prompts on each and compare credits per finished task from the usage report, not per request.
The hover, and the date. Every Fusion response shows which models ran. Sample a day of High or Max responses and tally how often a critique or cascade fired; that is your real multiplier, since GitHub publishes none. Any cost measured before November 30 is 30% below the undiscounted rate on paid plans, so divide by 0.7 before budgeting December.
The zero-budget moment. Before moving to Copilot access, set a $0 budget on a test user and confirm that completions keep working while Chat and code review stop. Then confirm the “AI credits paid usage” policy is on.
Model policy. Fusion draws only from models your policy enables, and GitHub retires models on its own schedule. Review the list Fusion shows for your enterprise and decide whether any model should be disabled for Fusion alone.
Sources
GitHub changelog, October 27 and 28
- Route and orchestrate models with GitHub Copilot Fusion
- GitHub Copilot Fusion now supports selection of local models
- GitHub Enterprise now includes access to GitHub Copilot
- GitHub Agentic Workflows generally available, GitHub Copilot app, Universe releases, Agent Merge watches and fixes pull requests in the cloud, New tools, triggers, and cost controls for Copilot Automations, Copilot code review: Merge with confidence with Max and approvals, Work from anywhere with a new, connected Copilot cloud experience, GitHub Copilot SDK is becoming Microsoft Copilot SDK, GitHub Copilot supports server-to-server authentication
GitHub documentation
- Copilot Fusion, Auto model selection, Models and pricing, Plans, Billing for organizations and enterprises, What’s changing with GitHub Enterprise and Copilot, FedRAMP models, Copilot with data residency
GitHub and Microsoft posts
- GitHub Universe 2026: Evolving every layer of the stack — GitHub Blog
- Project HydraFusion: Frontier quality via multi-model orchestration — GitHub Blog, September 4, 2026
- GitHub Copilot is moving to usage-based billing — GitHub Blog, April 27, 2026
- Bringing local models and sandboxed tools to Windows and GitHub Copilot — Microsoft, October 7, 2026
Independent benchmark