GitHub Copilot Fusion mixes models to achieve near-Opus 5.5 coding performance at half the per-task cost
B · Useful, limited A VentureBeat launch report by Carl Franzen that adds an executive interview and carefully separates Artificial Analysis's independent benchmark figures from GitHub's own claims; useful and precise, but it relays others' measurements rather than producing any.
Summary based on the full text·Original: about 10 minutes to read
In short
A VentureBeat article (28 October 2026) on GitHub Copilot Fusion, launched at GitHub Universe. Fusion is an orchestration layer that routes each coding request to one model or coordinates several. In Artificial Analysis's tests, Fusion Max scored 60 on its Coding Agent Index against about 62 for Claude Code with Opus 5.5 at Max, at $5.59 per task against $11.42, about 51% less. The article keeps GitHub's larger claims (40% faster, up to 77% cheaper) apart from the independent numbers, and covers a new enterprise billing option without per-seat credit allowances.
Key numbers
- $5.59average cost per task for Fusion Max on Artificial Analysis's Coding Agent Index
- $11.42average cost per task for Claude Code with Claude Opus 5.5 at Max, on the same index
- 60Fusion Max overall index score, against approximately 62 for the Claude Code configuration
- 7.2 milliontokens per task for Fusion, about half the Claude Code configuration's use
- 77%maximum cost reduction GitHub reports from its own separate tests, not the independent average
As reported by the source.
English·Web page·opens venturebeat.com
What it reports
- Three effort modes (standard, high, max) and three patterns: Single (one model), Cascade (cheaper model first, escalate if needed) and Critique (one model drafts, another reviews, the first revises).
- The model pool observed by Artificial Analysis includes Claude Opus 5.5, Claude Sonnet 5.5, GPT-6 Astra and GPT-6 Sol; actual choice varies by request and organisation policy.
- Index scores: Fusion 60, Claude Code with Opus 5.5 Max about 62, Devin Fusion 62 at $8.00 per task. Fusion used 7.2 million tokens per task, so savings came partly from doing less model work.
- Uneven by component: Fusion led on Terminal-Bench 4.0 (71) but trailed on FrontierCode (51 against 56 and 54).
- The 40% time saving is GitHub-reported. The 77% figure comes from GitHub's own DeepSWE tests, with up to 8.7 points better task completion, under different conditions.
- Billing: each underlying model call is charged at its standard rate; a temporary 30% discount applies to Fusion use. Eligible Enterprise Cloud customers can pay from an enterprise AI budget with no monthly credit allowance per seat; chat, code review and cloud agents stop when the budget runs out, completions do not.
- Administrators can see and disable models eligible for Fusion and set a default effort level.
How they got it
Launch coverage combining GitHub's announcement and documentation, an interview with GitHub's chief product officer, Artificial Analysis's published test, and another outlet's report on GitHub's internal tests.
Limits
- Benchmark cost per task, not an enterprise bill; no worked example of a team's monthly usage.
- One independent test; the 40% and 77% claims are vendor-reported.
- Benchmark tasks exclude human review, retries, integration and security work.
- Per-request audit detail for which models were used was not established in the interview.
Why it matters for companies
Coding-agent spend is moving from seats to metered usage, and this launch makes routing the main cost lever. A two-point score gap at half the cost is a good trade for routine work, but the component scores show it is not uniform: repository-change tasks did worse. Buyers should test on their own repositories, confirm that model policies and audit logs meet their governance needs, and budget for metered billing before moving off seat licences.
Summary based on the full text. Read the original for the full argument and data.Read the original on venturebeat.com