![Copilot Fusion Hits GA: 3 Tiers, 30% Off [2026]](https://shattered.io/wp-content/uploads/2026/10/github-copilot-fusion-ga-three-effort-tiers-30-percent-discount-2026-gen.webp)
GitHub made Copilot Fusion generally available on October 28, 2026, and The New Stack covered the debut as a model selection and orchestration system. Instead of asking developers to pick a model for every task, Fusion decides whether one model or several should handle the work. It runs on all Copilot plans in Visual Studio Code, the GitHub Copilot app, GitHub Copilot CLI and JetBrains IDEs.
The launch sits on top of a credit-based billing model, which this breakdown of GitHub Copilot pricing and AI credits explains in detail. Paid plans get a 30% discount on Fusion model costs through November 30, 2026. That discount, the three effort levels and a benchmark claim of roughly half the cost of Claude Code explain why the release drew so much attention within two days.
Two days in, the facts are clearer than the hype. GitHub Copilot Fusion is a router, not a model. It went from research preview to general availability in about four weeks. Its local-model routing is still experimental, and its headline benchmark comes from GitHub’s own announcement.
Don't miss new tech stories on Google
Add Shattered once in the Google app and our stories show up in your news suggestions.
What GitHub Shipped on October 28 #
GitHub announced Fusion at GitHub Universe 2026 and described it as multi-model orchestration across reasoning, code generation, debugging and tool use. The GitHub Universe 2026 recap frames it as one layer of a wider push to turn Copilot from an assistant into a delivery system for agents.
Fusion replaced the earlier model-selection workflow with automatic routing and orchestration. It appears in the model picker with three effort levels, and the rollout is gradual, so some accounts will not see it yet. GitHub’s documentation says users must update their client to get access.
Several other changes landed the same day. Experimental routing between local and cloud models arrived in VS Code, the Copilot app and Copilot CLI. GitHub announced that a cloud-agent session can move between the app, VS Code, the CLI and GitHub.com. The company also said the GitHub Copilot SDK is becoming the Microsoft Copilot SDK, though the announcement gave no migration timing.
One more model release landed two days earlier. GPT-6.1 Terra became generally available in GitHub Copilot on October 26, and GitHub describes it as OpenAI’s newest model. GitHub has not said whether Terra sits inside the Fusion pool, which is an open question covered below.
How GitHub Copilot Fusion Works: A Router, Not a Model #
GitHub’s documentation is explicit that Fusion is not a foundation model. Users pick an effort level, and Copilot evaluates the task, then chooses a single model or a combination. GitHub says the system weighs reasoning, code generation, debugging and tool-use needs. It combines the multi-model orchestration from Project HydraFusion with Copilot’s fine-tuned automatic model selection.
Three Effort Levels
The levels are Fusion Standard, Fusion High and Fusion Max, and they represent increasing effort and model capability. High prioritizes capable models for difficult tasks, and Max prioritizes the most capable models for demanding reasoning. GitHub’s docs add that locally installed models are more likely to be used on Standard when they fit the job.
The tiers replace manual coordination. A developer who once switched between a fast model for edits and a heavier model for debugging now states how hard the task is and lets the router decide. That shift matters most for teams that burned time debating which model to standardize on.
The Model Pool and Admin Controls
According to the AICoder write-up of GitHub’s announcement, the pool includes Claude Opus 5.5, Claude Sonnet 5.5, GPT-6 Astra and GPT-6 Sol. Administrators can disable individual models for Fusion only and set Fusion as the default. That control matters for compliance teams that approve vendors one at a time, because a blocked model stays blocked inside the router too.
Model churn is the reason a router exists. When Copilot retired six models and shifted toward GPT-5.6, developers who had hard-coded a favorite had to rework their habits. A router that absorbs those swaps behind a stable name removes that chore.
Where Copilot Fusion Is Available Right Now #
As of October 30, 2026, GitHub lists four clients on all Copilot plans, with more promised as coming soon and no date attached. Each surface sits at a slightly different stage.
| Client | Fusion status | Local-model routing | Key detail |
|---|---|---|---|
| Visual Studio Code | Generally available Oct. 28 | Experimental preview | Ships with VS Code 1.144 |
| GitHub Copilot app | Generally available Oct. 28 | Experimental preview | Intelligent local routing limited to managed Windows ML models |
| GitHub Copilot CLI | Generally available Oct. 28 | Experimental preview | v1.0.105 fixes Fusion High and Max in long-context sessions |
| JetBrains IDEs | Available, gradual rollout | Not listed in the announcement | Replaces the Auto option |
| GitHub.com (web) | Not listed as a Fusion client | Not applicable | Cloud-agent sessions can continue here |
The CLI deserves a closer look. GitHub documents installing Copilot CLI through the VS Code installer on Windows, by adding VS Code to the path on macOS, or by running the copilot command on Linux. Release v1.0.105, dated October 28, fixes the selection of Fusion High or Max in long-context sessions, which suggests the heaviest tiers hit real-world friction in the final days before launch.
Copilot in JetBrains: Auto Out, Fusion In #
For JetBrains users the change is a swap. Fusion replaces the Auto model option, and the rollout reaches Copilot users gradually. Anyone who still sees Auto in their model picker has probably not received the update yet. Auto is now the former option, not the current one. The same JetBrains release reportedly added local sandboxes and next edit suggestions. The announcement does not list JetBrains among the clients for local-model routing, so Copilot JetBrains users get the cloud-side orchestration first and the on-device preview later, if at all. GitHub gave no timeline.
JetBrains is running its own agent strategy in parallel. Its Air environment supports Codex, Gemini, GitHub Copilot, Claude Agent and Junie, with other agents connectable through the Agent Client Protocol. In a JetBrains blog post on why faster developers do not make a faster team, the company said Air added GitHub Copilot, OpenCode, Pi and Cline support during October. Air ships as a Marketplace plugin and is built into the JetBrains IDE 2026.3 EAP builds.
That dual track shows how JetBrains is hedging. Its own Java and Kotlin tooling now reaching other editors signals a company that wants its intelligence layer to travel, whichever agent the developer prefers. Fusion inside JetBrains IDEs fits that picture, because the router becomes one more agent choice rather than a replacement for JetBrains’ own.
Local Model Routing: What the Experimental Preview Covers #
Local routing is the least mature part of the launch, and GitHub labels it an experimental preview. Fusion can choose between supported on-device models and cloud models in Copilot CLI, the Copilot app and VS Code. The managed download is Windows-only and uses MAI Code 1.1 Flash through Windows ML. On macOS and Linux, developers can configure local providers manually.
GitHub named two examples. MAI Code 1.1 Flash and Qwen3.8 Flash Next can both run through a local OpenAI-compatible server. Developers who manage their own providers must handle model downloads and inference runtimes themselves. A separate CLI change announced October 9 already let developers discover models from a running Ollama instance through the model command.
The limit that matters is that local inference does not make a session offline. Fusion’s routing decisions still happen in the cloud, so a laptop on a plane gains nothing from the preview. What developers gain is cost control, since cheap local models can absorb simple tasks while the router saves expensive cloud calls for the hard ones.
The Benchmark Claim: Fusion Max at 60 vs Claude Code at 62 #
GitHub’s launch material leans on one comparison. On the Artificial Analysis Coding Agent Index v2.0, as relayed by AICoder’s launch report, Fusion Max scored 60. Claude Code with Opus 5.5 Max scored 62, and the Devin Fusion pairing also scored 62. The gap is two points, while the cost gap is far larger.
| Setup | Index score | Cost per task | Efficiency note |
|---|---|---|---|
| Copilot Fusion Max | 60 | $5.59 | About 7.2M tokens per task |
| Claude Code with Opus 5.5 Max | 62 | $11.42 | Roughly twice the tokens of Fusion |
| Devin Fusion pairing | 62 | $8.00 | Fusion Max uses about 30% fewer tokens |
The sub-scores show a mixed picture. Fusion Max took a slight lead on Terminal-Bench 4.0 at 71. It placed third on FrontierCode at 51, behind Claude Code at 56 and Devin Fusion at 54. It scored a weaker 57 on SWE-Atlas-QnA-AA. GitHub also reports about 40% less time per task.
Basic arithmetic puts the saving at $5.83 per task, or about 51%, against the Claude Code setup. The per-token price looks nearly identical in both setups, at roughly $0.78 per million tokens for Fusion. The savings come from using about half the tokens, not from cheaper tokens. That is a believable mechanism for a router that skips heavy models on easy steps.
Treat the numbers with care. They are vendor-promoted, and the source is GitHub’s own announcement. No independent replication exists yet. For a deeper head-to-head on seat prices and workflows, see this comparison of GitHub Copilot vs Claude Code.
What Fusion Costs: Credits, the 30% Discount and Billing Rules #
Fusion bills each model it invokes at that model’s standard rate. The announcements describe no separate Fusion surcharge. Per Visual Studio Magazine’s report on VS Code 1.144, Microsoft said paid Copilot plans receive a 30% discount on Fusion usage through November 30, 2026. GitHub’s own documentation describes the discount for JetBrains as covering the first 30 days. The two windows nearly coincide, but the November 30 date is the one printed on the official page.
Copilot’s AI credit system sets the context. One credit equals $0.01, and every plan includes a monthly allowance. GitHub’s plans page lists $15 of monthly credits for Pro, $70 for Pro+ and $200 for Max. Users can set a dollar budget for extra usage, and alerts fire at 75%, 90% and 100% of a configured budget. Code completions and next edit suggestions do not consume credits. The same page reportedly lists individual monthly prices of $0 for Free, $10 for Pro, $39 for Pro+ and $100 for Max.
Rough math shows why the allowances matter. A benchmark-style Fusion Max task at $5.59 equals 559 credits. If Fusion usage draws on the monthly allowance, which GitHub’s description of credits as the way to pay for AI usage implies, a Pro plan’s 1,500 credits cover under three such tasks. Pro+ covers about 12 and Max about 35. Real tasks run lighter than the benchmark, so this is a ceiling on heaviness, not a forecast. It still shows that heavy agent work moves developers into paid budgets quickly.
Enterprise billing is shifting too. Reports from the Universe announcements say GitHub Enterprise customers can use Copilot features for every licensed user and pay for AI usage instead of managing separate Copilot seats. Treat that as reported rather than confirmed pricing. For buyers weighing alternatives, the GitHub Copilot vs Cursor comparison of four models in eight days shows how fast the model lineup around these prices changes.
Competitive Comparison: Fusion, Claude Code, Junie and Air #
Fusion competes on three fronts. Against Claude Code, the pitch is near-parity quality at roughly half the price per task. Against JetBrains Junie and Air, the pitch is reach, since Copilot is already installed across editors. Against orchestration layers from other vendors, the pitch is a model pool that spans Anthropic and OpenAI under one subscription.
JetBrains is pushing hard on price and locality. Junie Lite launched on October 2 as a free Junie tier for every JetBrains account, with Qwen 3.6 27B behind it at launch. JetBrains says the offer runs through December 31, 2026, and warns that terms and models may change. Junie Local reached NVIDIA RTX Spark hardware on October 16, with stable Windows support, and JetBrains describes it as free and without token quotas.
JetBrains also published a measurement worth a look. Its tests found that JetBrains IDE intelligence made coding agents typically 20% to 30% more efficient, and up to about 65% in the best case. That claim comes from the vendor, but it points the same direction as GitHub’s. Both companies argue that routing and context, not raw model strength, now decide cost.
Developers who track the terminal-agent field will recognize the pattern. The Claude Code, Codex CLI and OpenCode updates within four days show rivals shipping on the same weekly cadence. Teams choosing between Copilot and Cursor can start from this Cursor vs GitHub Copilot benchmark walkthrough and rerun it with Fusion selected.
From HydraFusion Preview to GA in Four Weeks #
Fusion’s path to launch was short. On October 2, GitHub’s weekly release notes still described HydraFusion as a research preview in VS Code, the Copilot app and Copilot CLI. By October 28 the product had a public name, three effort levels and general availability. The dated events below shaped that stretch.
| Date (2026) | Event |
|---|---|
| Oct. 2 | HydraFusion still a research preview in VS Code, the Copilot app and CLI | | Oct. 2 | JetBrains launches Junie Lite, a free tier for every JetBrains account | | Oct. 9 | Copilot CLI gains local model discovery through the model command, including Ollama | | Oct. 16 | Junie Local arrives on NVIDIA RTX Spark, Windows support called stable | | Oct. 22 | CLI v1.0.103 retires legacy remote-launch flags | | Oct. 26 | GPT-6.1 Terra becomes generally available in GitHub Copilot | | Oct. 28 | Fusion goes generally available, VS Code 1.144 ships, CLI v1.0.105 lands, local-model preview opens | | Oct. 29 | User-reported VS Code issue about Fusion edits without a confirmation (unconfirmed by GitHub) |
GitHub has not formally retired the HydraFusion name, and the available sources do not give a separate retirement date. A Max variant, HydraFusion Max, reportedly carried the same preview label as recently as mid-October. The public product name today is Copilot Fusion, and anyone still reading HydraFusion documentation should treat it as the predecessor.
Session Continuity, Agent Host and VS Code 1.144 #
Routing is only half the story. GitHub also announced continuity for cloud-agent sessions across the Copilot app, VS Code, Copilot CLI and GitHub.com. A developer can start a task in one surface and continue it in another while keeping the conversation, code changes and working context. For teams that jump between a terminal and an editor, that removes a real source of lost context.
Visual Studio Magazine’s report on VS Code 1.144 lists the other headline changes in the editor release. VS Code 1.144 previews image generation and editing in Copilot Chat using GPT Image 2.5 Flare. It adds experimental desktop computer control and a marketplace-style Browse interface for plugins, skills and other agent tools. Microsoft also published version 1.0 of the Agent Host Protocol, an open protocol for talking to agent sessions across local and remote environments.
The protocol is the sleeper item. An open standard for agent sessions lets third-party clients attach to the same sessions that Copilot’s own apps use. If other editors adopt it, Fusion’s session continuity could extend beyond GitHub’s four named clients. GitHub said additional clients are coming soon but named no date.
Microsoft Takes the Branding, GitHub Keeps the Product #
Two naming changes landed on October 28 and both point toward Microsoft. GitHub said the GitHub Copilot SDK is becoming the Microsoft Copilot SDK. Separately, Microsoft’s Copilot Studio team announced that the GitHub Copilot harness is now the Microsoft Copilot harness. Reports say existing agents keep working without capability or billing changes.
Neither change affects Copilot’s core editor products. The sources confirm no rename of VS Code, Copilot CLI or the JetBrains plugin, and no acquisition or leadership change at GitHub or Microsoft in the week of October 23 to 30. The pattern is still worth watching. Shared SDK and harness names suggest Microsoft wants one agent stack across GitHub, Copilot Studio and its own Copilot apps.
For developers, the practical advice is simple. Check import paths and package names before assuming SDK documentation matches what you installed last quarter. The announcement gave no migration timing or completion status, so any rename could roll out in stages.
Known Limits and Open Questions #
Several gaps remain between the launch story and what developers can verify today. These matter most for teams deciding whether to switch their default model.
- Terra’s place in the pool: GPT-6.1 Terra went generally available in Copilot on October 26, yet the sources do not say if Fusion can route to it.
- Independent evidence: The 60 vs 62 comparison originates with GitHub’s announcement. No third-party rerun has appeared.
- Safety of edits: A user-reported VS Code issue on October 29 alleges Fusion can apply edits without a confirmation . GitHub has not confirmed that behavior, and it should not be read as policy.
- Local routing scope: The preview is experimental, JetBrains is not listed, and routing decisions still happen in the cloud.
- Discount cliff: The 30% discount ends November 30, 2026, so December bills will be the real test of the cost claim.
- Plan pricing: The official Fusion announcements do not list plan names or subscription prices, so cost comparisons rest on per-model rates.
Teams should run their own pilot rather than adopt the benchmark by faith. A fair test uses the same ten or twenty real tickets across Fusion Standard, High and Max, then compares credits burned, review time and rework. Anything less measures the marketing instead of the tool.
Market Impact for Developers and Tool Vendors #
GitHub holds a distribution advantage that rivals cannot copy quickly. Its own Copilot page says more than 90% of the Fortune 100 use it, and it publishes customer stories such as Mercado Libre cutting coding time by 50% and Grupo Boticário raising developer productivity by 94%. Those are vendor-published figures, not audited results. They show where GitHub aims the pitch: enterprise buyers who want one contract and one admin console.
A headline in AICoder’s coverage of GitHub Octoverse 2026 says agents now start over 40% of pull requests and that Copilot users tapped 47 models. If that holds, model choice is already too sprawling for humans to manage by hand. A router turns that sprawl into a product feature, and it moves bargaining power from the model vendors toward the platform that decides which model gets the call.
Standalone agent vendors face the sharpest pressure. If a Copilot subscription already bundles Claude and GPT routing, a separate agent has to beat Fusion on quality or add something the router cannot. Review-tool vendors feel it as well, which is why the CodeRabbit, Greptile, Qodo and Bugbot pricing comparison now has to treat Copilot as a bundled rival. The wider coding and developer tools coverage tracks how these bundles reshape the market.
Five Predictions for AI Coding Tools After Fusion #
- Rivals copy the effort-tier model. Expect at least one major competitor to ship a named routing layer with low, medium and high tiers within a quarter.
- Credit transparency becomes a selling point. Once the 30% discount expires, vendors that show per-task credit costs in the editor will win trust from finance teams.
- Local routing reaches JetBrains. The IDE vendor already ships local agents, so pressure to match the VS Code and CLI preview will build.
- Admin controls tighten. Enterprises will demand per-model audit logs for Fusion, since the router makes model choices a human no longer sees.
- Independent benchmarks appear. A third party will rerun the Coding Agent Index comparison, and the gap between 60 and 62 will either hold or shift.
These are analyst judgments, not reported plans. The first four follow from incentives visible in the October 28 announcements. The fifth follows from the simple fact that a vendor-supplied score invites a rerun.
FAQ: GitHub Copilot Fusion #
What is GitHub Copilot Fusion?
It is a model-routing and orchestration system, not a single model. Users choose an effort level, and Copilot decides whether to use one model or several for reasoning, coding, debugging and tool use. It became generally available on October 28, 2026.
Where can I use Copilot Fusion?
Fusion works on all Copilot plans in Visual Studio Code, the GitHub Copilot app, GitHub Copilot CLI and JetBrains IDEs. The rollout is gradual, and you must update your client to see it.
How much does Fusion cost?
GitHub bills each model Fusion invokes at its standard rate. Paid Copilot plans get a 30% discount on Fusion model costs through November 30, 2026. The announcements do not list a separate Fusion price.
Does Fusion replace the Auto option?
In JetBrains IDEs, yes. Fusion replaces the Auto model option there, with a gradual rollout starting October 28. GitHub has not announced a rename of the wider Copilot product.
Can Copilot Fusion run local models?
Only in an experimental preview in VS Code, the Copilot app and Copilot CLI. MAI Code 1.1 Flash runs through Windows ML, and manual setups can use MAI Code 1.1 Flash or Qwen3.8 Flash Next on Windows, macOS and Linux. Routing decisions still happen in the cloud, so sessions are not offline.
Is Fusion better than Claude Code?
GitHub’s reported index score is 60 for Fusion Max against 62 for Claude Code with Opus 5.5 Max, at $5.59 versus $11.42 per task. The comparison comes from GitHub’s announcement and lacks independent replication, so run your own tests before switching.
What happened to HydraFusion?
HydraFusion was the research-preview name for the multi-model orchestration work. GitHub folded it into Copilot Fusion, combining it with Copilot’s automatic model selection. Treat HydraFusion as the predecessor.