Claude Haiku 5.5 vs GPT-6 Luna: Same Price, Very Different Results

Claude Haiku 5.5 and GPT-6 Luna list at the same price, $0.10 per million input tokens and $0.50 per million output tokens, yet Artificial Analysis measured $0.21 per task for Haiku and $0.07 for Luna at Max effort. Haiku also scored higher on its Intelligence Index v4.3.2 (43 vs 38). The price is identical, but the tokens each model burns to get there are not.

All figures below were captured on 13 October 2026. Each is marked as lab-reported or independent, and benchmark versions are named because several of these tests have changed versions recently.

Do Claude Haiku 5.5 and GPT-6 Luna really cost the same?

On list price, yes. Anthropic lists Haiku 5.5 at $0.10 input and $0.50 output per million tokens for prompts up to 100k tokens. OpenAI's model page lists GPT-6 Luna at $0.10 input and $0.50 output. Cached input is $0.01 for both.

The two diverge on long prompts and on batch handling, as of 13 October 2026:

Haiku's long-context step is 5x on both input and output. Luna's is 2x on input and 1.5x on output, so at the threshold Luna would be $0.20 / $0.75 per million tokens, the figure Vals AI also states. If your workload regularly sends prompts between 100k and 272k tokens, the effective price gap is large. The Haiku 5.5 figures are from Anthropic's announcement and docs.

Anthropic also says Haiku 5.5 costs around 75% less to run than Haiku 4.5 on average, after accounting for a new tokenizer that uses slightly more tokens per task. That is a company claim about a different comparison, not about Luna.

Which model scores higher on independent benchmarks?

Haiku 5.5 leads on the headline index in both independent sources reviewed. Artificial Analysis scores it 43 against 38 for Luna on Intelligence Index v4.3.2, with both models at Max. On Vals AI, Haiku 5.5 holds a Vals Index of 54.31% (±1.41) against 51.22% (±1.07) for Luna, both run at max effort.

The picture is mixed below the headline. Artificial Analysis, independent, Max effort for both:

Haiku wins the large gaps (Terminal-Bench 4.0, AA-Briefcase, GDPval-AA). Luna wins AutomationBench-AA by 18 points. Four benchmarks are ties or within two points. "Very different results" is accurate for agentic and office-task tests, and not accurate for SciCode, long-context reasoning or CritPt.

Does the lab-reported table match the independent runs?

Directionally yes, numerically no. Anthropic's announcement includes a GPT-6 Luna column, so this is a lab-reported comparison against a competitor. Anthropic names no effort setting for the Haiku or Luna figures and no harness.

The two Artificial Analysis benchmarks reproduce almost exactly, which is unsurprising since Anthropic's page describes GDPval-AA as Artificial Analysis's benchmark. The gap shows up on Terminal-Bench 4.0, where Anthropic's 39.2% for Haiku is 4 to 6 points above the independent runs. Here the lab number sits above the independent ones. Luna's score is low in all three, so the ranking holds even where the absolute value moves.

Why does Haiku cost three times more per task at the same price?

Because it generates roughly three times as many tokens. At Max effort, Artificial Analysis recorded about 162k output tokens per task for Haiku 5.5 and about 50k for Luna. Reasoning tokens account for most of the difference: 129k per task for Haiku against 39k for Luna.

Vals measures a different task mix, so its ratio is larger. The direction is the same in both: for the same per-token price, Haiku at max effort is the more expensive model to run. If you were choosing on price per token alone, you would miss this. One caveat applies to the cost per task. These runs used Max effort. Haiku 5.5 defaults to medium effort in the API, and Luna defaults to medium as well. Neither independent source reported a medium-effort comparison on this page, so the 3x figure should be read as a Max-effort result.

How do speed and latency compare?

Haiku 5.5 streams faster, but Luna reaches its answer sooner. Artificial Analysis measured output speed at 238 tokens per second for Haiku and 134 for Luna. Time to first token, which at Max includes reasoning, was 309.80 seconds for Haiku and 113.29 seconds for Luna. End-to-end time for a 500-token response was 311.90 seconds against 117.02 seconds.

The long first-token times are a product of reasoning at Max effort, not network latency. Time per task, which Artificial Analysis computes as output tokens divided by output speed and so excludes first-token wait, was closer: 421.38 seconds for Haiku and 373.02 seconds for Luna. On Vals, whole-run latency was 47 minutes 17 seconds for Haiku and 30 minutes 34 seconds for Luna. Artificial Analysis's own FAQ on the page concludes that Haiku is faster while Luna has lower latency. No p50 or p95 distribution is published for either model.

Where does each model win?

The independent data suggests a split by task type:

• Agentic coding and terminal work: Haiku. Terminal-Bench 4.0 is 33% vs 13% at Artificial Analysis and 35.35% vs 13.64% at Vals.
• Code generation: Haiku on Vals' Vibe Code Bench v1.1 (90.44% vs 81.65%). SciCode is tied at 55%.
• Reasoning and proofs: Haiku on ProofBench v1.1 at Vals (89.00% vs 64.00%) and on Humanity's Last Exam at Artificial Analysis (44% vs 39%).
• Workflow automation: Luna on AutomationBench-AA (53% vs 35%).
• Security: a near tie on CyberBench v1.1 at Vals (75.48% for Haiku, 76.25% for Luna).
• Cost-sensitive volume work: Luna, at roughly a third of the per-task cost on Artificial Analysis.

Refusals differ too. Vals reports a refusal rate of 0.22% for Haiku 5.5 and 0.04% for Luna, and counts refusals as failed tasks. Both are low, but Haiku refuses about five times as often on Vals' set.

What's still unclear?

Several points limit how far these numbers go.

• Effort settings. Anthropic's table names no effort for Haiku or Luna. Both independent runs used Max. Results at default effort, where most production traffic runs, are not reported in either independent source.
• Snapshots move. Vals' launch-day post for Luna showed a 58.45% index at rank #20 of 65, and its current header shows 51.22% at rank #28 of 47. Vals does not explain the change on the page. Luna's Terminal-Bench 4.0 on Vals also differs between the launch post (9.60%) and the current header (13.64%). Treat any leaderboard position as dated.
• Sample size and intervals. Artificial Analysis does not publish confidence intervals on the comparison page. Vals gives ±1.41 and ±1.07 on the index, so the 3.1-point index gap is outside both intervals, but individual sub-benchmark gaps of a few points are not established.
• Cost per task is provisional in spirit. It depends on effort, task mix and, for Haiku, on whether prompts cross 100k tokens. Neither source models the long-context surcharge in a task-level figure on the pages reviewed.
• Different benchmark versions. Vals' launch post for Luna also lists Terminal-Bench 2.1, and its live table lists Terminal-Bench Science. Only Terminal-Bench 4.0 is compared in this post, to avoid mixing versions.
• Weakest point of the argument. Haiku 5.5 launched on 7 October 2026, and its independent results are about a week old. A re-run with different harness versions could shift the gaps.

Should you choose Haiku 5.5 or GPT-6 Luna?

Choose on workload, not list price. If your tasks look like terminal or coding agents, proof-style reasoning or office-task generation, the independent data favours Haiku 5.5 enough to justify roughly three times the cost per task at Max effort. If your tasks look like workflow automation or high-volume classification where cost per task dominates, Luna's 53% on AutomationBench-AA and $0.07 per task favour it.

Run your own evaluation at the effort level you will ship, with prompts at your real length. The long-context price steps differ by a wide margin, and that alone can reverse the ranking.