BRENNEN

GitHub Copilot Fusion

Effort, not model

GitHub Copilot Fusion picks the AI models for each request, and in a test GitHub paid for, its top setting scored 2 points below a rival at about half the cost.

Agents · Checked 13 Nov 2026 · 9 sources · 9 min

A programmer working with an AI assistant types requests into a box beside the code. They are ordinary sentences, such as asking why an invoice total comes out wrong, and the assistant reads the project's files, makes a fix and reports back. Next to the box is a small menu that decides what does the work. It lists AI models, the programs that do the reading and writing.

The models differ the way workers do. Some are fast, cheap and good enough for routine jobs. Others handle harder problems, take longer and cost more. A model is billed for the text it reads and writes, counted in tokens, each a word or part of a word, and the rate depends on the model. So whoever is at the menu has to judge how hard a job is before anyone has tried it. Aim too high and routine work is paid for at the top rate. Aim too low and the hard job may come back wrong.

On 28 October 2026, at its GitHub Universe event, GitHub made GitHub Copilot Fusion generally available, meaning released and not a preview. Copilot is GitHub's coding assistant. With Fusion, the person at its menu chooses an effort level, Standard, High or Max, and Copilot chooses the models for each prompt, as a single request is called, and arranges how they work together.

With the launch, GitHub reported that its top setting, Copilot Fusion Max, scored 60 on a test called the Artificial Analysis Coding Agent Index, "within two points of Claude Code with Opus 5.5 Max". It got there, GitHub said, at about half the average API cost per task, the pay-per-token charge for using the models, and in about 40% less time per task. Claude Code is the rival here, a competing coding assistant set up with a model called Opus 5.5. Whether nearly the same score can cost about half as much depends on where the saving comes from.

Suppose ten jobs come in over a morning. The case is made up, and so is every price in it. Seven are routine, like rewording an error message. Three are hard, like the invoice total that is wrong for no visible reason. Suppose, generously, that the strongest model gets all ten right and costs $3 a job. Send everything to it and the morning costs ten times $3, or $30. On seven of those jobs the $3 bought more ability than the work needed. The whole saving lies in those seven, and it can be collected only by something that can tell the seven from the three. Both kinds arrive looking the same, as a sentence or two of ordinary words.

A cheap draft and a judge

GitHub's documentation calls Fusion "an intelligent runtime model router and orchestrator". A router decides which model a request goes to. An orchestrator sets more than one model to work on the same request. For each prompt, the documentation says, Fusion uses what it calls capability signals, for reasoning, code generation, debugging and tool use, to pick one of three ways of working.

In the first, the request goes to one model and that model does the job. The model is still chosen before any work is done, only now by software. The second waits to see some work before judging. An efficient model, one of the cheaper ones, drafts the job. Then a judge looks at the draft and either accepts it or passes the job up to a stronger model. GitHub calls this Cascade. In the third, a model writes a draft, one other model or several review it, depending on which GitHub page is read, and the first model revises it.

Run the ten jobs through a cascade. Suppose a draft and the judging of it cost $1 a job together. All ten jobs pay that, which comes to $10. The judge accepts the seven routine drafts and passes the three hard jobs up at $3 each, which adds $9. The morning costs $19 where it cost $30.

The same sums show how a cascade fails. A judge that trusts nothing passes all ten jobs up, and the morning costs $10 plus $30, or $40, more than the strong model alone, because every job was paid for twice. A judge that trusts too much accepts a wrong draft of one hard job. The bill drops to $10 plus two jobs at $3, or $16, the lowest yet, and one job in ten is wrong. Nothing on the bill tells a good morning from a bad one; that shows only when someone checks the finished work.

So what a cascade delivers depends on its judge, and what a router delivers depends on its rule for sizing up a request. GitHub has published neither, and it publishes no list of the models Fusion draws on. Artificial Analysis, which ran the test, names four of the models Fusion chooses from, and one of them, Claude Opus 5.5, is the model the rival was set up with.

With the workings out of view, the system can be judged only the way the ten jobs were: give it work that can be graded, then look at the bill and the finished work together. A fixed set of such work with fixed rules for grading is called a benchmark.

Ahead on one test, behind on two

The Coding Agent Index combines three benchmarks into one score. It covers 290 tasks, with three attempts at each. In 100 of them the assistant changes a real open-source project, one whose code is public, and is graded on whether the people who look after it would accept the change; another 66 are jobs done in a terminal, the bare text window where programmers type commands. The last 124 are questions about a project's code, judged by an AI model. Artificial Analysis tested Fusion's max setting in GitHub Copilot CLI, Copilot's terminal version, and no other setting has third-party figures.

Artificial Analysis ran the test, but GitHub commissioned and funded it, which it disclosed in a reply on X. GitHub's changelog, where it announced the launch, and the Artificial Analysis pages, as served on 13 November 2026, carry no such statement. A test the seller paid for is weaker evidence than one a stranger chose to run, because a seller chooses which tests to pay for and which results to put in its announcements. The figures may still be right, but no repeat of the test by a party without a commercial link to GitHub appears in launch coverage, and the 100 open-source tasks are not publicly released, so a third of the index cannot be rerun from outside.

On the leaderboard, the ranked table of results, as served on 13 November 2026, Copilot Fusion (max) scored 59.86 and the row labelled Claude Code with Claude Opus 5.5 (Max, Default Fallback) scored 61.65. Subtract one from the other and the gap is 1.79 points, with Fusion the lower; GitHub's 60 is Fusion's score rounded.

"Within two points" is true and doesn't say on which side. Artificial Analysis does, placing Fusion 2 points below Opus 5.5 in Claude Code, and on the leaderboard Fusion ranked third of 36 rows. The New Stack attributes each figure to GitHub, though its headline says Fusion "beats Claude Code".

The headline score is itself three numbers. The index is the equal-weight average of the three benchmark scores, so each counts for a third however many tasks it holds. On the terminal jobs Fusion leads the Claude Code row, 71.21 to 63.13. On the open-source changes it trails, 51.37 to 55.97, and on the code questions it trails again, 56.99 to 65.86. Add Fusion's three scores, divide by three, and the result is 59.86. The Claude Code row's three average 61.65.

Artificial Analysis Coding Agent Index, score by benchmark

Ahead on one benchmark, behind on two

Artificial Analysis

FusionThe Claude Code row
Terminal jobs71.2163.13
Open-source changes51.3755.97
Code questions56.9965.86
The average of the three59.8661.65

An average like that describes a mixed pile of work and no single kind in it. For work that is all terminal jobs, the part of the test that applies shows Fusion ahead. For work that is all questions about a project's code, it shows Fusion behind. Artificial Analysis's own advice is that the index "should be read alongside the per-eval breakdowns", an eval being a test.

Whether 1.79 points is a difference at all has not been established. A model doesn't do the same thing every time it is asked, so the same test run twice can give two slightly different scores. The Artificial Analysis pages give no confidence interval, the range that shows how far chance alone could move a score, and without one the gap cannot be compared with that variation. Nor was the rival one model throughout. Three attempts at each of 290 tasks make 870, and on 76 of 870 attempts, 8.7%, the Claude Code row fell back to an earlier Opus model after a safety refusal, where the model declines to continue. No source states the effect on its score.

For time, the leaderboard measures average agent wall time per task, the clock time an assistant spends on a task, not counting setup or grading. Fusion averaged 2,167 seconds, or 36.1 minutes, against 3,661 seconds, just over an hour, for the Claude Code row, which is 40.8% less. Yet 22 of the 36 rows have a shorter average than Fusion's.

A token bill, not a subscription

Artificial Analysis gives the cost as $5.59 for Fusion against $11.42 per task, slightly under half, and describes it as pay-per-token API cost, not consumer plan pricing. This is the real counterpart of the made-up morning's bill: token spend on 290 benchmark tasks against one competitor configuration. No source claims that a Copilot subscriber's bill halves.

For each task, as Artificial Analysis measures it

Less clock time, and under half the token cost

The cost is pay-per-token API cost, not consumer plan pricing. 22 of the 36 rows on the leaderboard have a shorter average time than Fusion's.

Artificial Analysis articleLeaderboard

FusionThe Claude Code row
Average time on a task2,167 seconds3,661 seconds
Cost of a task$5.59$11.42

A customer's bill works differently. There is no separate charge for Fusion: each model it uses is billed at its standard rate, less a 30% discount for paid plans through 30 November 2026. Because each model has its own rate, the cost of a task depends on which models Fusion picks. At the Standard effort level, the documentation says, Fusion selects one model for each task, and a task at High or Max may consume more AI credits than one at Standard.

Picking a model from the menu names what will do the work. Picking Fusion names a setting, and behind the setting are a pool of models and an unpublished rule for choosing among them. GitHub's documentation says the pool may change. If it does, the score announced at the launch describes an earlier configuration.

Check it yourself

  1. Open the GitHub Changelog entry and read its third paragraph to confirm the words "scored 60", "within two points" and "about 40% less time per task".
  2. Open GitHub's reply on X to confirm that GitHub commissioned and funded the evaluation.
  3. Open GitHub's Models and pricing table to confirm that each model is billed at its own rate per million tokens.
  4. Read About GitHub Copilot Fusion, including its limitations, to confirm the three ways of working, the 30 November 2026 end of the discount and that a task at High or Max may consume more AI credits.
  5. Read the Artificial Analysis article to confirm "2 points below", $5.59 against $11.42 and the score on each of the three benchmarks.
  6. Find the Copilot Fusion (max) row and the Claude Code with Claude Opus 5.5 row on the Coding Agent Benchmarks leaderboard, then recompute the gap in score and the difference in time from its charts.
  7. Read the component table on the methodology page to confirm the 290 tasks, the scoring rules and that the 100 open-source tasks are not publicly released.
  8. Inspect the tasks and the judge configuration in the dataset of code questions to confirm the 124 tasks behind a third of the index.

Sources