#github copilot#copilot fusion#claude code#opus 5.5#costo por tarea#ia para programar

    GitHub Copilot Fusion: the AI that picks the model for you and what it costs

    GitHub Copilot Fusion is on every plan: it picks the model for you and claims half the cost of Claude Code. The math and the fine print.

    GitHub Copilot Fusion isn't a new model: it's an option that decides which models to use and how they work together. On October 28, 2026, at GitHub Universe, it came out of preview and became available on every Copilot plan. GitHub says it lands within two points of Claude Code with Opus 5.5 Max at about half the cost. Here is the math and how to check it with your own tasks.

    What changed with GitHub Copilot Fusion

    Almost everyone described it as "Copilot got another model." It didn't. Before, you opened the list and picked: the cheap one, the expensive one, the one that shipped yesterday. Now you pick Fusion and an effort level, and Copilot decides which models to use. It runs in Visual Studio Code, in the Copilot app, in the command line, and in JetBrains.

    It arrives for a concrete reason: since June 1, Copilot bills by tokens, at each model's price. Picking wrong stopped being a matter of taste: it's a money decision.

    Code on a dark screen

    How it works inside

    Fusion has three effort levels: Standard for everyday work, High for hard tasks, and Max for the most demanding. For each request it picks one of three flows:

    • Single: one model solves it.
    • Cascade: an efficient model writes a draft, and a quality check accepts it or sends it up to a stronger one.
    • Critique: one model drafts, another reviews, and the first one fixes it.

    You don't choose which one gets used. That saves you time, and it's what you'll have to keep an eye on.

    How to test Copilot Fusion step by step

    1. Update your client: Fusion only shows up in newer versions.
    2. Open the model picker in Visual Studio Code or the Copilot app and choose Copilot Fusion.
    3. Start on Standard.
    4. Type a real task, for example "fix the contact form, which doesn't send the email when the phone field is empty."
    5. Check your billing page for how many credits it used and compare with the same task on your usual model.

    If the change doesn't solve it, don't keep pushing on Standard: move up to High and save Max for what truly calls for it. Write down the cost of five tasks: without that table, everything is just an opinion.

    The math: how much GitHub Copilot Fusion costs

    An assumption to keep the numbers round: a task with Claude Code and Opus 5.5 Max costs you $1. GitHub says Fusion Max costs about half ($0.50), and there's a 30% discount, for a limited time, on paid plans ($0.35).

    | Option | Cost per task | Tasks with $10 of credits (Copilot Pro) | Team of 5, 1,000 tasks a month | | --- | --- | --- | --- | | Claude Code + Opus 5.5 Max | $1.00 | 10 | $1,000 | | Fusion Max | $0.50 | 20 | $500 | | Fusion Max with discount | $0.35 | almost 29 | $350 |

    That's $650 less a month, $7,800 a year. Your numbers will differ, but the formula doesn't: cost per task times tasks per month. And the math stops paying off if Fusion costs you more than $1.43 per task.

    Calculator and receipts on a desk

    Three business cases (my own analysis)

    A freelancer maintaining twelve client sites. Fusion on Standard for everything, and only what fails moves up to High. The cost no longer depends on guessing the model, so they can charge a fixed monthly maintenance fee without their margin swinging with every request.

    A small software company on Copilot Business. At $19 per user with $19 in credits, the admin sets Fusion as the default on Standard and turns off, just for Fusion, the models they don't want to pay for, without taking them away from everyone else. Spending becomes predictable. Detail: if the credits run out there's no cheaper fallback model, so usage is part of the budget. Fusion only uses the models your organization's policy allows.

    An online store with a single developer. Fusion on Max only for what touches payments and billing, and Standard for the rest. It's the hardest to copy because it requires knowing which parts of your code are expensive to get wrong; Fusion doesn't decide that, you do.

    Who it's not for

    • If by contract or rule you need a specific, documented model on every task: Fusion decides for you.
    • If you only use autocomplete: code suggestions don't spend credits on any plan.
    • If you code now and then: at $3 a month, the difference doesn't pay for the time it takes to compare.

    The objection and what they don't tell you

    GitHub reports the figure about its own product, and the cost per task depends on how it was measured. But GitHub isn't selling that Fusion is better; it's saying it's nearly the same, and cheaper, and you can check that with five of your tasks in one afternoon. If your work is very particular (an unusual language, a legacy system), your own test counts more than any index.

    On "Fusion beats Claude Code at half the cost": it doesn't beat it. Fusion Max scored 60 on the Artificial Analysis coding agent index, and Claude Code with Opus 5.5 Max scored two more. GitHub reports that measurement, with no independent confirmation. In the preview it measured against Claude Opus 5 (not 5.5), offline: it gained 4.9 points on Terminal-Bench 2.1 with 67% lower estimated cost, landed a point and a half behind on DeepSWE with 36% lower cost, and a tenth behind on CheckpointBench, an internal benchmark. Treat the 50% as a hypothesis to test.

    What breaks in practice

    • You don't control the model: if a result comes out odd, you can't tell whether it was Single, Cascade, or Critique.
    • Critique uses two models for one task (my inference): on an easy change it could cost more than going straight to the cheap model.
    • No fallback model: when the credits run out you have to buy more.
    • Not in every Copilot client yet; GitHub says more are coming.

    How to apply it in your business?

    The five-step method: 1) update and pick Copilot Fusion; 2) start on Standard, move up to High only if it fails, and reserve Max for where a mistake costs money; 3) write down five real tasks with cost and time and repeat them on your usual model the same day; 4) do the math per task, not per token; 5) if you're an admin, set Fusion as the default on Standard and remove the models you don't want to pay for. The 30% discount has no published end date: when it ends, measure again, because the $0.50 math is without the discount.

    To see how to measure cost per task with other models, read the full breakdown of Kimi K3.1 Pro vs Opus 5.5, and for another Copilot case, the Grok 4.7 in GitHub Copilot breakdown.

    Frequently asked questions

    What is GitHub Copilot Fusion?

    It isn't a model: it's an option that chooses which models to use and how they work together, with three effort levels (Standard, High, and Max).

    Which plans have it?

    Every Copilot plan, since October 28, 2026.

    Does it beat Claude Code with Opus 5.5?

    No: on the Artificial Analysis index, Fusion Max scores 60 and Claude Code with Opus 5.5 Max scores 62, as GitHub reports.

    Where does it work?

    In Visual Studio Code, the Copilot app, the command line, and JetBrains.

    Conclusion

    The question about GitHub Copilot Fusion isn't whether it's better than Opus: GitHub itself says it isn't. It's how much of your AI spending is just paying for the priciest model out of habit. Test it with five of your own tasks before deciding.