GitHub Copilot Fusion Leads Terminal-Bench 4.0 at 71, Costs $5.59 Per Task Versus $11.42 for Claude Code
GitHub has released Copilot Fusion inside GitHub Copilot — a runtime orchestration system that selects among Claude Opus 5.5, Claude Sonnet 5.5, GPT-6 Astra, and GPT-6 Sol for each coding task, optimising for performance, cost, and latency simultaneously. Artificial Analysis benchmarked it at max effort in GitHub Copilot CLI as part of the Coding Agent Index v2.0, published October 28.
The Numbers
The Coding Agent Index v2.0 averages three benchmarks equally: Terminal-Bench 4.0, FrontierCode, and SWE-Atlas-QnA-AA.
| System | TB 4.0 | FrontierCode | SWE-Atlas | Index Score | Cost/Task |
|---|---|---|---|---|---|
| Claude Code (Opus 5.5 max) | — | 56 | — | 62 | $11.42 |
| Devin Fusion (Opus 5.5 + GPT-6 Luna) | — | 54 | — | 62 | $8.00 |
| Copilot Fusion (GitHub Copilot) | 71 | 51 | 57 | ~60 | $5.59 |
| GPT-6 Astra (Codex) | — | — | — | — | — |
| Gemini 4 Argon (Antigravity CLI) | — | — | — | — | — |
Copilot Fusion leads the Terminal-Bench 4.0 component with a score of 71 — above Carbon’s 65.2% on the same benchmark and above Sonnet 5.5’s standalone score of 70.6%. It ranks third on FrontierCode at 51, behind Claude Code at 56 and Devin Fusion at 54. SWE-Atlas-QnA-AA comes in at 57, falling between GPT-6 Astra in Codex and Gemini 4 Argon in Antigravity CLI.
Overall Coding Agent Index: approximately 60, two points behind Claude Code and Devin Fusion (both at 62).
Token Efficiency Is the Headline
At 7.2 million tokens per task, Copilot Fusion uses roughly half the tokens of Claude Code with Opus 5.5 — AA estimates Claude Code at around 14.4M tokens per task — and 30% fewer than Devin Fusion. The lower token count is a direct output of model routing: routine subtasks go to Sonnet 5.5 or GPT-6 Sol; only the hardest reasoning steps escalate to Opus 5.5 or Astra.
Token costs are charged at the underlying model’s rate, so cost per task varies with the actual model mix. The $5.59 benchmark figure reflects AA’s standard task distribution.
What It Means
The emerging pattern in enterprise coding agents is a two-axis competition: raw capability versus cost-adjusted throughput. Claude Code remains the ceiling on FrontierCode and overall index score. But Copilot Fusion’s Terminal-Bench 4.0 lead at 71 suggests that for bash-heavy, multi-step agent tasks, model routing can outperform a single frontier model — at least when the routing system has access to complementary model strengths across providers.
GitHub describes Copilot Fusion as a “runtime orchestration system.” The architecture is not a fine-tuned router; it is a policy layer above the models. That distinction matters for enterprise buyers evaluating whether vendor lock-in is traded for orchestration dependency.
Devin Fusion, with its own multi-model approach (Opus 5.5 plus GPT-6 Luna), scores identically to Claude Code at 62 but at $8.00 per task — 30% cheaper than Claude Code, 43% more expensive than Copilot Fusion. The market is now clearly segmented: frontier single-model (Claude Code, $11.42), multi-model near-frontier (Devin Fusion, $8.00), and cost-efficient multi-model (Copilot Fusion, $5.59).
Copilot Fusion ships with no confirmed pricing changes to existing GitHub Copilot plans.