Gemini 4 Carbon vs Hy4
Gemini 4 Carbon leads by 5.8 capability points (≈95% interval +3.1 to +8.6): a clear lead over Hy4 · P(higher) ≈ 100%
Wins 17 of 17 shared benchmarks (0 lost)
Capability vs. relative cost
Gemini 4 CarbonHy4Closed weightsOpen weights
Hy4 has no setting with a relative cost yet.
Benchmark wins17 – 0 · 17 shared
Comparing matching setups only (155 benchmarks excluded) · suites counted once
Area← Gemini 4 CarbonHy4 →
Data: Surge AI, Mercor, Collinear AI and 4 othersatlas.village.dev
Margins
Each benchmark scaled 0–100% (its declared range, else the field’s range); Δ is Hy4 minus Gemini 4 Carbon.
| Chartography (November 2026 rescoring) | Perception & media › Documents & charts | 85.5 | 49.2 | −36.3 |
| RiemannBench (November 2026 rescoring) | Math | 80.0 | 51.2 | −28.8 |
| DAYJOB: Healthcare (October 2026 rescoring) | Health & medicine | 35.2 | 7.2 | −28.0 |
| HANDBOOK.md Agents v1.1 | Tool & computer use › Customer service & operations | 49.8 | 23.7 | −26.1 |
| GDP.xlsx | Professional › General knowledge work | 38.9 | 17.7 | −21.2 |
| DAYJOB: Finance (October 2026 rescoring) | Professional › Finance, accounting & tax | 23.8 | 9.2 | −14.6 |
| GDP.pdf (November 2026 rescoring) | Professional › General knowledge work | 56.4 | 43.6 | −12.8 |
| FoodTruck Bench | Business simulations | 77218 | 67598 | −12.3 |
| EnterpriseBench: CoreCraft | Tool & computer use › Customer service & operations | 80.5 | 68.2 | −12.3 |
| Harvey LAB (Mercor run, undisclosed ~213-task cohort) | Professional › Legal | 23.0 | 13.6 | −9.4 |
| τ³-Banking (Mercor run, bm25_grep retrieval) | Tool & computer use › Customer service & operations | 55.7 | 48.1 | −7.6 |
| DAXBench v2 | Software › Data & ML | 93.6 | 87.2 | −6.4 |
| APEX-Agents 1.1 | Professional › General knowledge work | 90.4 | 85.1 | −5.3 |
| Agents' Last Exam v1.1 (ALE v1.1) | Professional › General knowledge work | 45.9 | 41.0 | −4.9 |
| LLM Debate Benchmark | Language & writing › Language & wordplay | 1676 | 1647 | −4.2 |
| ComplexConstraints (November 2026 rescoring) | Language & writing › Instruction following | 70.9 | 66.8 | −4.1 |
| CWE-bench v2 Pass@1 | Software › Cybersecurity | 67.3 | 64.1 | −3.2 |
Data: Surge AI, Mercor, Collinear AI and 4 othersatlas.village.dev