Claude Haiku 5.5 vs GPT-6 Luna
| Benchmark | Haiku 5.5 | GPT-6 Luna |
|---|---|---|
| AutomationBench-AA | 28.6 | 40.5 |
| GDPval-AA (work tasks) 2.1 | 38.7 | 38.1 |
| AA Analyst Agent | 53.8 · max | 43.8 · max |
| Terminal-Bench 4.0 | 15.2 | 2.5 |
| Terminal-Bench Science 0.1 | 7.1 | 8.6 · max |
| AA-Briefcase 1.1 · Elo | 1,371.5 | 1,098.8 |
| AA-LCR (long context) 1.1 | 77.3 | 78.3 |
| FrontierCode 1.1 | 40.4 | 41.9 · max |
| Harvey LAB-AA 1.1 | 0.6 | 3.3 · max |
| Furniture Assembly | 47.5 · max | 44.2 · max |
| DeepSWE 1.1 | — | 67 · max |
| SWE-bench Pro (v2 full) | 87.2 · high | — |
| SWE-bench Pro (v2 hard) | 52.9 · high | — |
| EBR Bench | 43.8 · max | — |
| SpireClimb | — | 0 · max |
| SpireClimb StS1 A20 Heart | — | 0 · max |
| SpireClimb StS2 A10 | — | 0 · max |
Specs
| Spec | Haiku 5.5 | GPT-6 Luna |
|---|---|---|
| Input, per 1M | $0.10 | $0.10 |
| Output, per 1M | $0.50 | $0.50 |
| Cost per task? | $0.047 | $0.017 |
| Speed, tok/s | 179 | — |
| Context | 1M | 1.05M |
| Weights | Closed | Closed |
| Released | Oct 2026 | Sep 2026 |
| See images | ✓ | ✓ |
| Read PDFs | Partial | Partial |
| Hear audio | — | — |
| Watch video | — | — |
| Call tools | ✓ | ✓ |
| Force a tool call | Partial | Partial |
| Return strict JSON | ✓ | ✓ |
| Search the web | ✓ | ✓ |
| Run code | ✓ | ✓ |
| Use a computer | ? | ✓ |
What breaks when you switch
Function calling only with none · Chat Completions›
Chat Completions function calling requires reasoning_effort none; use Responses for reasoning with tools.
More than 272K input tokens›
Prompts above 272K input tokens charge 2× input and cache rates and 1.5× output rates for the entire request.
Remove unsupported sampling parameters›
When reasoning effort is not none, remove temperature, top_p and top_logprobs, plus Chat Completions logprobs or Responses output logprobs inclusion.
Cache writes billed at 1.25x›
Cache writes cost 1.25× the uncached input token rate.
Context: 1.05M → 1M
Remove sampling parameters›
Unsupported sampling values, any top_k or combined temperature and top_p return HTTP 400; omit all three and use prompting.
Replace manual thinking budgets›
Manual thinking with budget_tokens returns HTTP 400; replace the old thinking configuration.
No assistant prefill›
A final assistant prefill returns HTTP 400 even with thinking off; end messages with a user turn.
Replace old computer tool · Claude API / Google Cloud›
computer_20250124 returns HTTP 400 on the Claude API and Google Cloud; use computer_toolset_20260801 there.
Keep conversations append-only›
Replaying thinking after earlier history changes returns HTTP 400, with older accounts checked only when the binding field is set; keep history append-only.
Bedrock needs another computer version · Bedrock›
Bedrock does not accept computer_20250124; use computer_20251124 with the computer-use-2025-11-24 beta header.