Copilot Fusion Scores 59.9 to Claude Code’s 61.7 at Half the Cost. Codex on GPT-6.1 Sol Costs $1.50.

What GitHub shipped

On October 28, at its Universe conference, GitHub made Copilot Fusion generally available. Instead of choosing Claude, GPT or Gemini from a dropdown, you choose “Copilot Fusion” and an effort level: Standard, High or Max. Copilot then decides which models do the work.

GitHub’s changelog lists three ways Fusion can run a task:

• Single: one model solves it.
• Cascade: a cheaper model drafts, a quality check accepts the draft or escalates to a stronger model.
• Critique: one model drafts, a second reviews, the first revises.

Fusion runs in VS Code, the GitHub Copilot app, Copilot CLI and JetBrains IDEs. It only uses models your organization has enabled, and enterprise admins can switch a model off for Fusion alone or make Fusion the default.

The headline claim, from GitHub’s own changelog: Fusion Max scored 60 on the Artificial Analysis Coding Agent Index, “within two points” of Claude Code with Opus 5.5 Max, at about half the API cost and tokens per task and about 40% less time. We opened Artificial Analysis’s leaderboard to check every one of those numbers.

The numbers check out

Artificial Analysis rounds scores to whole numbers on its charts. The leaderboard data behind them carries one decimal, and that’s what we use below. Coding Agent Index v2.0 averages 3 tests with equal weight: FrontierCode v1.1 (100 tasks), Terminal-Bench 4.0 (66 tasks) and SWE-Atlas-QnA-AA (124 tasks). Cost is pay-per-token API spend per task.

Every part of GitHub’s claim holds. Fusion trails by 1.8 points. It costs $5.59 per task to $11.42, 51% less. It uses 7.34 million tokens per task to 14.52 million, 49% fewer. It finishes a task in 36.1 minutes on average to 61.0, 41% faster.

Here’s the top of the board, default configurations with a published cost:

Source: Artificial Analysis Coding Agent Index v2.0, opened November 2, 2026. Muse Code with Muse Wonder Full scores 58.9 but has no public API price, so it’s left out.

Where Fusion wins, and where it doesn’t

A close average can hide a lopsided scorecard, and Fusion’s is lopsided. It posts 71.2 on Terminal-Bench 4.0, the highest of all 34 agent configurations Artificial Analysis lists, 8.1 points above Claude Code with Opus 5.5. Then it gives most of that back on the other two tests.

SWE-Atlas-QnA, from Scale AI, asks an agent to answer questions about production open-source codebases: trace execution paths, explain architecture decisions. Fusion’s 57.0 there is the lowest of the top 10 configurations on the index, below both Codex setups. FrontierCode is the test where every top agent scores lowest, and Fusion trails Claude Code there by 4.6 points.

So the plain reading: if your agent work looks like Terminal-Bench (run commands, edit files, get tests passing in a shell), Fusion is the best measured option right now, and the cheapest of the top 3. If it looks like “understand this big repo and answer precisely,” Claude Code and Devin still lead.

The $1.50 row

Fusion halves Claude Code’s bill. Codex with GPT-6.1 Sol at max effort cuts it by 87%. It scores 54.1, 7.6 points under Claude Code and 5.8 under Fusion, for $1.50 a task and 23.7 minutes.

Put another way: for the $11.42 Claude Code spends on one task, Codex with Sol can run 7.6 tasks. Fusion can run 2. Whether 7.6 points of score are worth 7.6x the price depends on how often you’d retry a failed run anyway.

One more row worth knowing: Claude Code with Sonnet 5.5 at max is the most expensive configuration in the table, $13.30 a task and 79.7 minutes, for a lower score than Opus 5.5. A cheaper per-token model that thinks longer isn’t a cheaper agent.

What it costs you inside Copilot

Artificial Analysis measures API list cost. Inside Copilot, GitHub says it bills each model Fusion uses at that model’s standard rate, with no orchestration fee, in AI credits worth $0.01 each. High and Max can call more than one model per task, so they can use more credits than Standard.

Sources: GitHub Copilot plans page and GitHub Docs, Copilot Fusion, opened November 2, 2026. Business and Enterprise also get all 3 levels.

The plan that matters for the benchmark result is Pro+. The $10 Pro plan only gets Fusion Standard, which uses one model per task and isn’t the configuration Artificial Analysis tested.

A rough rule of thumb, if Copilot’s rates match API list prices: Pro+’s $70 of credits covers about 12 benchmark-sized Fusion Max tasks at $5.59, or about 17 with the 30% discount that runs through November 30. Benchmark tasks are long agent runs, not quick chat questions, so your everyday prompts will cost far less each.

What this means if you code with AI

Pick by cost per task, not price per token. Sonnet 5.5 is cheaper per token than Opus 5.5 and costs more per task. Fusion mixes expensive models and still halves the bill. The only number that tells you what you’ll pay is what a finished task costs.

Routing is now a top-3 strategy. Two routers sit in the top 3 of the index, both cheaper than Claude Code with Opus 5.5 alone. Letting software pick the model per task is no longer a trade of quality for price.

If you already pay for Copilot Pro+, try Fusion Max before November 30. It’s close to the top score, the fastest of the top 3, and 30% cheaper than its list rates until the discount ends.

On a tight budget, Codex with GPT-6.1 Sol is the number to beat. 54.1 for $1.50 is the best balance of score and cost near the top of the board.

For hard repo questions, keep a frontier model in the loop. Fusion’s 57.0 on SWE-Atlas-QnA is its weak spot. If your work is mostly reading and reasoning about a large codebase, Claude Code and Devin Fusion score 8.9 and 11.3 points higher.

Full disclosure: we sell an AI coding agent, and we did not run these benchmarks. MegaBrain’s terminal coding agent puts many models behind one plan: free models on Free, included models such as GLM, Kimi, MiniMax, DeepSeek, Qwen and Grok on the paid plans, Claude, GPT and Gemini from Pro, plus any model through your own API key and local models, and Auto picks one for each task. See the plans. MegaBrain is not affiliated with GitHub, Anthropic, OpenAI or Artificial Analysis.

Sources

All pages opened November 2, 2026.

• GitHub Changelog — Route and orchestrate models with GitHub Copilot Fusion (October 28, 2026)
• GitHub Docs — Copilot Fusion (effort levels by plan, billing, 30% discount through November 30, 2026)
• GitHub — Copilot plans (Free, Pro $10, Pro+ $39, Max $100, AI credits)
• Artificial Analysis — GitHub Copilot Fusion orchestrates Anthropic and OpenAI models to balance performance and cost (October 28, 2026)
• Scale AI — SWE-Atlas-QnA dataset card (124 codebase question-answering tasks)
• Artificial Analysis — Coding Agent Index v2.0 leaderboard (scores, cost, time and tokens per task)