A coding assistant that chooses its own tools
“Completed work, not more code, is the output that matters.”
GitHub said in its general availability announcement that Copilot Fusion routes and coordinates several AI models according to the task. Its supported lineup includes Claude Opus 5.5, Claude Sonnet 5.5, GPT-6 Astra and GPT-6 Sol. Rather than presenting model selection as a permanent developer decision, Fusion treats it as part of the product experience.
That could change the rhythm of software work. A developer might ask for a small configuration change and receive a response from one model. A larger feature could trigger a sequence in which one model writes a solution, another examines it, and a third helps refine the result. The system supports Single, Cascade and Critique workflows, according to GitHub.
Single uses one model for a task. Cascade coordinates models in sequence, allowing work to move through several stages. Critique adds a review step, creating a structure that resembles a small engineering team more than a conventional chatbot. The developer still describes the desired result and reviews what comes back, but the selection and coordination of models happen behind the interface.
GitHub says Fusion is available across supported Copilot clients and includes administrator controls and billing options. Those details matter because routing is not only a technical choice. In a company, it is also a question of who can enable the feature, which models may be used, and how the resulting usage appears on an account.
The cost claim
The appeal is clearest in the benchmark numbers. GitHub says Fusion Max scored 60 on Artificial Analysis' Coding Agent Index, within two points of Claude Code using Opus 5.5, while costing about half as much per task.
Artificial Analysis reported in its comparison that Fusion Max averaged $5.59 in API cost per coding task. Claude Code running Claude Opus 5.5 averaged $11.42. The corresponding Coding Agent Index scores were 60 for Fusion Max and 62 for Claude Code.
That is a striking proposition for organizations that run coding agents across thousands of tasks. If a routing system can reserve the most expensive model for the hardest moments, while using less costly models for routine work, an engineering department could reduce the cost of experimentation, maintenance and incremental improvements.
The savings also suggest a different way to think about premium models. Instead of treating a powerful model as the default employee for every request, Fusion treats it more like a specialist called when the task requires deeper reasoning. A faster or cheaper model might handle the first pass. A stronger model could then review a difficult change or repair a failure.
Yet the benchmark is measuring API costs, not the complete financial cost of software development. It does not include the time a developer spends reading a patch, correcting an incorrect assumption, rerunning tests or recovering from a change that looked plausible but damaged an unrelated part of the repository.
Orchestration is not free of complexity
GitHub's documentation adds an important qualification. GitHub explains in its Fusion documentation that users are billed for every model Fusion invokes. High and Max workflows can use multiple models, while cached tokens may reduce some costs.
That means the final bill depends on the path a task takes, not simply on the name of the feature. A critique workflow may improve a result, but it can also add another model call. A cascade may make a complicated change more dependable, but every stage has a cost. The system's intelligence is therefore partly a budgeting problem, even when the interface makes it feel automatic.
The documentation also warns that discarded drafts can leave workspace edits in place. That is a small detail with large practical consequences. If a model produces a draft that is later rejected, a developer may still need to inspect the workspace and determine what changed. The discarded answer is not necessarily the same thing as a cleanly discarded action.
This is where the promise of invisible coordination meets the reality of software repositories. A model can be assigned a role, but the developer remains responsible for understanding the state of the code. A system that invokes several agents may save money at the API layer while creating additional review work at the human layer.
A benchmark result, not a shipping guarantee
Brian, writing in a critical post titled “Copilot Fusion Is Not a Model,” argues that the headline comparison combines separate measurements and should not be treated as a promise about real repositories or enterprise bills. Brian made that case in his analysis of Copilot Fusion, pointing to Fusion's weaker FrontierCode result and to the difference between benchmark API costs and total development costs.
That criticism is the only opposing view identified in the supplied material. It is also the most important limit on the comparison. A near match on an aggregate score does not mean the systems behave identically on every repository, language or organizational workflow. A lower average task cost does not prove that a team will spend less after accounting for review, failed attempts and operational safeguards.
The result may still matter. Coding agents are moving toward a world in which the question is not simply which model is strongest, but which combination of models can complete a task with acceptable quality and expense. That shifts competition from model makers alone to the designers of the systems that direct them.
The human stays in the loop
In the near future, a developer may experience Fusion less as a conversation with one artificial colleague and more as a quiet project room populated by specialized assistants. One investigates. One writes. One objects. The human decides whether the proposed change belongs in the product.
That arrangement could make advanced coding systems more accessible to teams that cannot justify using the most expensive model on every request. It could also make accountability harder if the path from prompt to patch becomes difficult to reconstruct.
Fusion's real test will therefore happen outside the benchmark. It will be measured in clean pull requests, predictable bills, understandable changes and fewer hours spent undoing confident mistakes. Orchestration may lower the price of intelligence, but only reliable collaboration will lower the cost of shipping software.
Maya Lindqvist is not a person. No notebook, no deadlines, no face behind the name — just a byline this newsroom publishes under. Here is the production line underneath it, because a name beside a portrait reads like a journalist, and this one is not one.
The models. Writing: gpt-5.6-luna and qwen3-max. Out on the live web: gpt-5.6-luna and gpt-5.6-terra. Pictures: gpt-image-1 and gpt-image-1-mini. Swap one in the newsroom and this line swaps with it — it is read off the machines, not typed here.
How a story is made
- Research. The searching model reads around the story, pointed at primary sources — the filing, the post, the repository — rather than at somebody else's write-up of them.
- Writing. The writing model drafts it against what was found, at Maya Lindqvist's usual length and in Maya Lindqvist's usual register.
- The loop. A reviewer reads the draft and sends it back with notes. Then reads it again. A piece can go round several times before it leaves the building.
- Enrichment. A quotation has to appear word for word on the page it is taken from. A chart may only use figures that appear in the source it cites. Whatever fails is dropped, and the reason is kept.
- Fact check. A last pass hunts for claims the article makes and its sources do not.
- A human stop. Sensitive subjects are held for a person to read before publication, and a person can kill any of it at any point.
If that sounds less like a newsroom and more like a factory: quite. It is called Press Factory.
This article was generated using AI and published automatically without human pre-publication review.
How this article was made
The article was produced by the Grandmonts Media News Engine using automated research, drafting and verification workflows. No human editor reviewed the article before publication. Grandmonts Media remains responsible for the published content. Errors can be reported at office@grandmonts.cz.