Claude Haiku 5.5 vs GPT-5.6 Luna
Claude Haiku 5.5 leads GPT-5.6 Luna by 6.1 points on the Shogo Agent Index, but their ranges overlap, so treat it as close. Claude Haiku 5.5 wins 15 and GPT-5.6 Luna wins 5 of 20 shared benchmarks. Claude Haiku 5.5 is cheaper for agent workloads.
Add more models to this comparison →Pillars
| Pillar | Claude Haiku 5.5 | GPT-5.6 Luna |
|---|---|---|
| Agentic coding | 67 | 59 |
| Long-horizon autonomy | 62 | 61 |
| Knowledge-work agents | 46 | 39 |
| Computer use | 47 | 40 |
| Security & context learning | 55 | 51 |
Shared benchmarks
| Benchmark | Claude Haiku 5.5 | GPT-5.6 Luna |
|---|---|---|
| APEX-Agentsindex | 56.1% | 43.0% |
| Vending-Bench 2index | $3,602 | $4,095 |
| CursorBenchindex | 48.4% | 35.9% |
| ARC-AGI | 89.0% | 88.0% |
| ARC-AGI-2 | 61.3% | 59.5% |
| Chess Puzzles | 55.0% | 43.0% |
| CritPt | 22.6% | 20.6% |
| DTBench | 95.0% | 89.1% |
| FrontierCode | 46.4% | 39.8% |
| FrontierMath Tier 4 | 46.3% | 61.0% |
| FrontierMath Tiers 1-3 | 75.1% | 82.1% |
| GPQA Diamond | 89.6% | 91.6% |
| LMCA | 51.1% | 48.5% |
| MirrorCode | 78.9% | 21.1% |
| OTIS Mock AIME | 98.9% | 98.3% |
| ProofBench | 89.0% | 60.0% |
| SciCode | 55.0% | 53.6% |
| SimpleBench | 59.2% | 46.8% |
| SimpleQA Verified | 23.8% | 41.0% |
| WeirdML v3 | 11.7% | 5.2% |
Specs and price
| Claude Haiku 5.5 | GPT-5.6 Luna | |
|---|---|---|
| Agent Index | 57.5 (#18) | 51.4 (#34) |
| Agent $/M | $0.050 | $0.10 |
| Input / output $/M | $0.10 / $0.50 | $0.20 / $1.20 |
| Context | 1M | 1.1M |
| Released | Oct 7, 2026 | Jul 9, 2026 |
| Weights | Closed | Closed |
| API hosts | 5 | 5 |
Don’t pick one. Route.
Shogo can run Claude Haiku 5.5 and GPT-5.6 Luna side by side and send each step of an agent to the one that’s cheaper for it.
Try both in Shogo