AI models
GPT-6 Luna vs Claude Haiku 5.5 vs DeepSeek V4.1 Flash
GPT-6 Luna, Claude Haiku 5.5 and DeepSeek V4.1 Flash on 5 benchmarks and price per 1M tokens.
GPT-6 Luna
US$ 0.10 / US$ 0.50 prompts up to 272,000 tokensInput US$/1M / Output US$/1M · Official pricing
Claude Haiku 5.5
US$ 0.10 / US$ 0.50 prompts up to 100,000 tokens · US only ×1.1Input US$/1M / Output US$/1M · Official pricing
DeepSeek V4.1 Flash
US$ 0.30 / US$ 1.20 peak hoursInput US$/1M / Output US$/1M · Official pricing
GPQA Diamond
90.5% · 89.6% · 89.8%
GPT-6 Luna · Claude Haiku 5.5 · DeepSeek V4.1 Flash
01
Direct answer
see more
02
On GPQA Diamond, Claude Haiku 5.5 is in a statistical tie with DeepSeek V4.1 Flash, Claude Haiku 5.5 is in a statistical tie with GPT-6 Luna and DeepSeek V4.1 Flash is in a statistical tie with GPT-6 Luna. On FrontierMath, Claude Haiku 5.5 leads DeepSeek V4.1 Flash and GPT-6 Luna leads DeepSeek V4.1 Flash. On SimpleQA Verified, DeepSeek V4.1 Flash leads Claude Haiku 5.5 and GPT-6 Luna leads Claude Haiku 5.5. On τ-bench Verified, Claude Haiku 5.5 is in a statistical tie with DeepSeek V4.1 Flash, Claude Haiku 5.5 is in a statistical tie with GPT-6 Luna and DeepSeek V4.1 Flash is in a statistical tie with GPT-6 Luna.
GPT-6 Luna
- FrontierMath +3.8 p.p.
statistical tie
- GPQA Diamond statistical tie
- τ-bench Verified statistical tie
03
| Metric | GPT-6 Luna | Claude Haiku 5.5 | DeepSeek V4.1 Flash | Source |
|---|---|---|---|---|
| Input US$/1M | US$ 0.10 / US$ 0.50 Official pricing prompts up to 272,000 tokens | US$ 0.10 / US$ 0.50 Official pricing prompts up to 100,000 tokens · US only ×1.1 | US$ 0.30 / US$ 1.20 Official pricing peak hours | |
| Output US$/1M | US$ 0.50 | US$ 0.50 | US$ 1.20 | |
| Context | 1,050,000 | 1,000,000 | 1,048,576 | |
| BENCHMARKS | ||||
| GPQA Diamond statistical tie | 90.5%between 37th and 64th of 250 statistical tie | 89.6%between 48th and 72nd of 250 statistical tie | 89.8%between 46th and 71st of 250 statistical tie | Epoch AI |
| OTIS Mock AIME saturated | 98.9% statistical tie with the top score | 98.9% statistical tie with the top score | 98.3% statistical tie with the top score | Epoch AI |
| FrontierMath (tiers 1 to 3) | 78.9%between 22nd and 24th of 85 | 75.1%between 25th and 26th of 85 | 67.4%between 33rd and 35th of 85 | Epoch AI |
| SimpleQA Verified | 41.4%between 48th and 53rd of 87 | 23.8%73rd of 87 | 42.9%between 46th and 48th of 87 | Epoch AI |
| τ-bench Verified (airline) | 68.4%between 90th and 92nd of 130 | 74%between 52nd and 71st of 130 | 77.8%between 10th and 30th of 130 | OpenRouter Benchmarks |
| OTIS Mock AIME saturated | 98.9% statistical tie with the top score | 98.9% statistical tie with the top score | 98.3% statistical tie with the top score | Epoch AI |
Compare with
How to read
- Benchmark scores are percentages, and higher is better.
- In each row, the highlight marks the model ahead by more than the benchmark's margin; within the margin, the row shows «statistical tie».
- A saturated benchmark stays in the table, with its badge, and does not count toward the verdict.
- Prices are in US$ per 1M tokens and carry the source label: «Official lab price», «Third-party provider» or «No current price».
- Every number comes with its source and date. «No data» means no measurement exists in the open sources this section uses.
Sources
- Epoch AI, CC BY 4.0, latest version we checked.
- OpenRouter Benchmarks, CC BY 4.0, latest version we checked.
- OpenRouter, dados públicos da API; exports de benchmarks e rankings em CC BY 4.0, latest version we checked.
Dates come from the sources: OpenRouter listing or Hugging Face repo creation, which may differ from the official launch. Family and version were grouped by rule from the name.
How to cite
If you use this data, cite the page and the date you accessed it.
Adriano Meirinho. "GPT-6 Luna vs Claude Haiku 5.5 vs DeepSeek V4.1 Flash". meirinho.com.br. https://meirinho.com.br/en/ia/modelos/comparar/claude-haiku-5-5-vs-deepseek-v4-1-flash-vs-gpt-6-luna/. Accessed November 21, 2026.
Third-party data follows the license of each source. The full table is available as CSV and JSON.