Claude Haiku vs ChatGPT Luna: Scores, Prices, and Speed (2026)

Claude Haiku 5.5 and GPT-6 Luna are Anthropic and OpenAI’s respective affordable small models. The name “ChatGPT Luna” in the title refers to OpenAI’s GPT-6 Luna: it is available through the API and, starting October 8, is also gradually rolling out in ChatGPT Free and Go chats.

The short answer

As of October 8, 2026, both models have the same standard API list price for short prompts: US$0.10 per million input tokens and US$0.50 per million output tokens. With the max setting on Artificial Analysis (AA), Haiku 5.5 has an Intelligence Index score of 43, compared with 38 for Luna. Luna’s preliminary cost per task is US$0.07, versus US$0.21 for Haiku 5.5. With both set to max on AA, Haiku 5.5 generates more tokens per second, but Luna finishes a 500-token answer sooner. This article covers Claude Haiku 5.5 and GPT-6 Luna as of October 2026.

Specs and pricing at a glance

First, a quick definition: a token is the unit of text used for model billing. The content you send counts as input, and the model’s reply counts as output. The long-prompt surcharge threshold applies to a single request, not your total monthly usage. The table below shows standard API list prices from official sources, as of October 8, 2026. Cache creation, Batch, and other options are billed separately.

Effort is a control for asking a model to think more. Higher settings usually improve scores, and usually increase token usage and cost as well. Settings with the same name do not mean the models use the same amount of reasoning. Both default to medium, so the comparisons below specify the setting each time.

Luna’s long-prompt price follows OpenAI’s stated rules: for requests with more than 272K input tokens, input and cache prices are multiplied by 2, and output by 1.5. The multipliers apply to the entire request.

Here’s an example where the two differ: a single prompt with 150K tokens. Haiku is past its 100K threshold, so the input costs 150K × US$0.50 ÷ 1 million = US$0.075. Luna is still below 272K, so 150K × US$0.10 ÷ 1 million = US$0.015, one-fifth of Haiku’s cost. This counts input only, not output.

Capability: which scores higher?

AA’s Intelligence Index combines several tests into one score. Each row below compares the same setting against the same setting. “Cost per task” is the weighted average cost across AA’s index tests, and is preliminary. It includes input, cache, reasoning, and answer usage, so it is different from list prices per million tokens. The Haiku 5.5 figure does not yet include the surcharge for prompts above 100K.

At matching settings, Haiku 5.5 leads by 4 to 5 points in every row, and its cost per task is also higher. At high, for example, the scores are 38 versus 33, and the cost per task is US$0.08 versus US$0.03. With both set to max on AA, Haiku 5.5 produces about 162K total output tokens per task, including reasoning and answers, while Luna produces about 50K, so Haiku 5.5 uses roughly 3.2 times as many.

The Vals Index is a separate evaluation. With both models set to max, Haiku 5.5 scores 54.31% and Luna scores 51.22%. The cost per Vals test is US$2.991 for Haiku 5.5 and US$0.431 for Luna.

On October 8, 2026, the Arena text leaderboard listed Luna (gpt-6-luna-max) at 1443 points, but did not yet list Haiku 5.5, so this source cannot be used for a direct comparison.

Anthropic’s published results also show Haiku 5.5 leading GPT-6 Luna in some areas, but the testers differ: for the OSWorld 2.1 offline subset, both used max and Luna was tested by Anthropic through the OpenAI API; Cognition tested FrontierCode; and AA independently tested GDPval-AA and AA-Briefcase. OpenAI’s article announcing Sol and Luna did not compare them with Haiku.

Speed

Speed has more than one measure. For tokens generated per second, Haiku 5.5 produces 178 tokens at AA’s high setting, compared with 112 for Luna. But how long does a complete answer take? AA measures end-to-end time for a 500-token answer: at high, Haiku 5.5 takes 29.49 seconds and Luna takes 23.75 seconds; with both at max, it is 371.33 seconds versus 112.54 seconds. In other words, Haiku 5.5 generates tokens quickly, but spends longer thinking first. Luna also has a shorter time to first token: at low, it is 3.23 seconds versus 11.19 seconds. For a 500-token answer, Luna finishes sooner when both use high or max. These are AA measurements; real-world results vary by provider and load.

How to choose

Both companies position their models for high-volume work: OpenAI’s model page calls Luna an efficient model “best for focused, high-volume tasks”; an AWS blog post introducing Haiku 5.5 describes it working alongside Opus 5.5, which handles planning and judgment. None of the suggestions below is a universal winner.

• For short, high-volume tasks where the AA Intelligence Index matters: test Haiku 5.5 first. With both set to max, it scores 43 versus 38; a specific task may still produce a different result.
• If you have a very tight budget per task or care about the wait for a complete answer: test Luna first. With both set to max on AA, its preliminary cost per task is US$0.07 versus US$0.21.
• If a single uncached input is over 100K tokens and you want to reduce input costs: test Luna first. Input for 150K tokens costs one-fifth as much as Haiku; add output for the total cost.
• If you already use one ecosystem: stay with that provider. Claude Code users can try Haiku 5.5 first, and Codex users can try Luna first, avoiding switching costs.
• For other cases: run 20 of your own real tasks on each model and compare quality and cost before deciding.

Start tests at the default medium setting, and increase effort if quality is not good enough. Higher effort usually increases usage, but not by a fixed amount for every task. Look at the total cost to complete a task, not only the list price.

Can you use them for free?

Haiku 5.5: Anthropic’s Haiku page says Free, Pro, Max, Team, and Enterprise users can select it on claude.ai, on the web, iOS, and Android.

GPT-6 Luna: On October 7, 2026, OpenAI announced that GPT-6 Luna would power Free and Go chats in ChatGPT, with a gradual rollout starting October 8. Your account may not see it right away. The chat version is tuned for everyday conversations and is not exactly the same as the API version of Luna. The desktop app’s Work and Codex are not part of this update. The plan page lists GPT-6 Luna at Standard speed for Free (US$0) and Go (US$8 per month), subject to rollout availability.

As of October 8, 2026, the official pages above do not list a fixed numeric usage limit for either model’s free plan. Check your account for remaining usage and reset times.

If your free account can already select the relevant model, you can send both the same set of questions and compare answer quality and wait time. This compares the chat experience and cannot reproduce the API evaluations in this article.

Penchan’s take

The list prices are the same, but the trade-offs differ: with AA’s max setting, Haiku 5.5 scores higher, while Luna finishes a 500-token answer sooner and has a lower preliminary cost per task. My simple rule: if scores are the bottleneck, test Haiku 5.5 first; if the bill or wait time is the bottleneck, test Luna first. The real difference depends on how many tokens your tasks use, so measure your own workload before deciding. For an introduction to Haiku 5.5, see this article; for ChatGPT, see the ChatGPT overview.

Further reading

• Claude Haiku 5.5 introduction
• Claude vs ChatGPT comparison
• Claude Opus vs Sonnet
• AI model comparison overview

This article compares AI tool features and subscription plans for consumer information. It is not securities or investment advice. Actual pricing follows each platform’s latest official announcements, and the information in this article may be outdated. Information checked on October 8, 2026.