Compare · updated 2026-11-01
Claude Haiku vs GPT-6 Luna
Claude Haiku is better at maths; GPT-6 Luna is better at quick answers, writing, coding, reasoning, structured data, long documents, translation. GPT-6 Luna is the cheaper of the two on output.
Measured quality by task
| Task | Claude Haiku | GPT-6 Luna | Better |
|---|---|---|---|
| Quick answers | 85/100 | 93/100 | GPT-6 Luna |
| Writing | 89/100 | 90/100 | GPT-6 Luna |
| Coding | 82/100 | 84/100 | GPT-6 Luna |
| Reasoning | 74/100 | 77/100 | GPT-6 Luna |
| Maths | 79/100 | 70/100 | Claude Haiku |
| Structured data | 86/100 | 93/100 | GPT-6 Luna |
| Long documents | 87/100 | 88/100 | GPT-6 Luna |
| Translation | 83/100 | 90/100 | GPT-6 Luna |
Quality is Nivana’s measured score for each kind of task, from weekly pairwise evals judged by models from several labs. Prices are live AI Gateway list prices per million tokens. Updated 2026-11-01. How this is measured.
Price and context
| Claude Haiku | GPT-6 Luna | |
|---|---|---|
| Lab | Anthropic | OpenAI |
| Input per 1M tokens | $0.1 | $0.1 |
| Output per 1M tokens | $0.5 | $0.5 |
| Typical chat answer | $0.00050 | $0.00050 |
| Context window | 1M | 1.05M |
| Router picks it for | Fallback only | 16 of 24 request types |
Which should you use?
Use the one that wins the kind of work you do most, unless the price gap outweighs the quality gap for easy requests. Or don’t choose: Nivana scores both, and the rest of the pool, on every request and sends it to whichever is best for that request.
Questions
Is Claude Haiku better than GPT-6 Luna?
On Nivana’s weekly evals Claude Haiku scores higher on 1 of 8 task types (maths) and GPT-6 Luna on 7 (quick answers, writing, coding, reasoning, structured data, long documents, translation). Neither is better at everything.
Which is better for coding, Claude Haiku or GPT-6 Luna?
GPT-6 Luna: 84/100 against 82/100 for Claude Haiku on Nivana’s coding evals.
Which is better for writing, Claude Haiku or GPT-6 Luna?
GPT-6 Luna: 90/100 against 89/100 for Claude Haiku on Nivana’s writing evals.
Or let the router pick, per request
Nivana sends every message to the model that scores best for it, sets how hard it thinks, and tells you which one answered. 2,000 free credits to start, no card.