GitHub Copilot Free tested: every check passed, but Fusion did not start

GitHub Copilot has a free plan. On 28 October 2026 GitHub also launched Copilot Fusion, which picks and mixes AI models for you, "on all Copilot plans". We tested the free plan with the Copilot CLI: GitHub's own tool that you use by typing commands in a terminal window. It got all 15 of our coding runs right: 310 of 310 hidden checks. But Fusion would not start on our free account: the tool gave an error.

What is Copilot Fusion?

Fusion is a new choice in Copilot's model list. Instead of picking one AI model yourself, you pick Fusion, and Copilot decides which models to use. GitHub's changelog names three ways it can work:

• Single: one model does the job.
• Cascade: a cheaper model writes a first answer. A check either accepts it or hands it to a stronger model.
• Critique: one model writes, another reviews it, and the first one fixes it.

Fusion has effort levels: Standard, High and Max. GitHub's Fusion docs say Standard "selects one model for each task". High and Max may use more models, and so more AI credits. GitHub's table shows Fusion Standard on every plan, including Copilot Free. High and Max start at Copilot Pro+.

GitHub also says Fusion Max scored 60 on the Artificial Analysis Coding Agent Index, an outside ranking of AI coding helpers. It says Fusion Max did that at about half the average API cost per task. That is GitHub's claim. We could not test Max on the free plan.

What does Copilot Free include?

GitHub's plan page lists what Copilot Free gives you:

• Price: free.
• Code completions: up to 2,000 a month. Those are the small suggestions while you type.
• AI credits: "an allowance", with no number. AI credits pay for chat and agent work. An agent is an AI helper that can change files and run commands for you. One credit is worth $0.01, GitHub's price page says.
• Models: "Auto model selection only". Auto picks the model for you.
• Tools: "All plans include Copilot CLI and the Copilot app."

The paid plans for one person are Pro ($10 a month), Pro+ ($39) and Max ($100). They come with 1,000, 3,900 and 10,000 base credits a month. On top of that comes a "flex allotment": extra monthly credits (500, 3,100 and 10,000) that GitHub says may change over time. Verified students can get Copilot Student for free.

Our test

We gave Copilot our secret coding test, the same one we use for other AI coding tools. It has three jobs: fix the money maths in a small program, write a bill splitter, and write a CSV reader. A CSV reader turns a text file with commas into a table. Hidden checks grade the result. The AI never sees those checks.

We used the official Copilot CLI, version 1.0.105, with the site owner's own Copilot Free login. It ran in a clean setup: no personal settings and no add-ons. Each job started in its own new folder, and the CLI was allowed to run commands on its own.

On 29 October 2026 we ran each of the three jobs three times. On 1 November we ran each job twice more. So each job ran five times: 15 runs in total. In our other tests, each tool got three tries per job. For the last three, we used the CLI's normal text output, because that shows the AI credits each job used.

Fusion on the free plan: not in our test

First we asked the CLI for Fusion. GitHub's docs on picking a model give the setting for that: model "fusion". On our Copilot Free account, the CLI stopped with an error: "The Copilot Fusion model alias requires dynamic Auto tiers." The message did not say how to fix it. We tried again on 1 November, with the newest CLI version, and got the same error.

That does not match GitHub's own words. The changelog says Fusion is available "on all Copilot plans" in the Copilot CLI, and the Fusion table shows Fusion Standard for Copilot Free. The docs also say Fusion does not show up when none of the models it uses are open to you. That may be the reason here, but GitHub does not say so. We could not check if Fusion works on Free in Visual Studio Code.

So we tested what Free did give us: Auto.

Which model did Auto pick?

Before each job, the CLI tells you which model Auto chose, and why. For all 15 of our coding runs, it said: "Auto selected GPT-6 Luna because it was the best choice out of the models available in your plan."

For a one-line prompt ("Reply with the word ok only."), Auto picked a different model: MAI-Code-1.1-Flash. In the detailed logs of 12 of the 15 runs, only one model was listed as "available" each time. The other 3 runs used plain text output, which only shows the choice. We cannot see how Auto decides inside. For all 15 coding runs, the CLI reported GPT-6 Luna as Auto's chosen model.

GPT-6 Luna is OpenAI's small, cheap model. We tested it before through OpenAI's API, the paid connection for developers. There, with one message per job and no tools, it passed 184 of 186 checks.

How well did it code?

Very well. Copilot Free passed all 310 of 310 hidden checks in 15 runs. A typical job took about 52 seconds, from start to finish.

We compare it with our Cursor vs Claude Code test on the same jobs, from 10 October:

Typical time is the middle value. Copilot was slower than Claude Code, but it passed every hidden check. Unlike our API test, the CLI could run the code and fix it before it was done. Our jobs are small and clear, so this does not show how Copilot does on a big project.

What does one coding job use?

In plain text mode, the CLI shows after each job how many AI credits it used and how many tokens. A token is a short piece of a word. We used that mode for our last three runs, on 1 November. They used 0.44, 0.53, 0.37 AI credits: about 0.45 per job. At $0.01 per credit, that is about half a US cent per job.

That fits GitHub's price list for GPT-6 Luna. Most of what the model read came from the cache: about 85,000 to 123,000 tokens per job. The cache is text the model has already seen, which costs much less to read again.

How many jobs like ours fit in Free's allowance? We cannot say, because GitHub gives no number for it. Bigger jobs can use more credits.

Should you use Copilot Free?

• You want to try an AI coding agent for free? Copilot Free did all our small jobs right. Good to start with.
• You want to pick the model yourself? Free only has Auto. Copilot Pro ($10 a month) gives you "a selection of models".
• You want Fusion High or Max? GitHub's table puts those on Pro+ ($39) and higher. We did not test them.
• You code a lot every day? GitHub gives no number for Free's allowance, so you may hit a limit. A paid plan tells you how many credits you get.

Want to pay per use through an API instead? See our GPT-6.1 Terra test.

How we checked

On 1 November 2026 we read GitHub's plan page, Fusion docs, price page and Fusion changelog. We installed the official Copilot CLI 1.0.105 from npm in a test folder, with a clean settings folder. The site owner logged in to their own Copilot Free account; we never saw a password.

We ran our secret coding test, explained on our methods page: 15 runs between 29 October and 1 November 2026. Hidden checks graded each job. The model Auto chose comes from the CLI's own messages, and the AI credits from its summary after each job. The Claude Code and Cursor results come from our test of 10 October 2026.

The screenshots are our own captures of GitHub's pages, only cropped. The charts are our own, and none of the images was made by AI. Prices and plans can change, so check GitHub's pages before you choose.