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

Anthropic released Claude Haiku 5.5 on October 7, 2026, and for prompts up to 100,000 tokens it lists at exactly the price of OpenAI's GPT-6 Luna: $0.10 per million input tokens and $0.50 per million output tokens, with $0.01 cache reads and $0.125 cache writes on both. The sticker is the same, so the choice turns on three other things: what each model gets right per task, what happens to the price on long prompts, and the platform and API rules around each one. This guide uses only the vendors' published figures.

The short answer

• Choose Claude Haiku 5.5 for computer-use and browser agents, for knowledge-work and visual tasks where Anthropic's published results favor it, for subagents that run under Claude Sonnet 5.5 or Opus 5.5, and when your stack already runs on the Claude API, Amazon Bedrock, Google Cloud or Microsoft Foundry.
• Choose GPT-6 Luna when individual prompts regularly exceed 100,000 tokens, when you need OpenAI's Responses API tools, Flex processing or EU data residency, or when your agents already run on GPT-6 Sol in the OpenAI stack.
• For short classification, routing and extraction, the list price is identical, so decide on accuracy measured on a labeled sample of your own data.

Price and limits side by side

Anthropic's long-prompt price applies to the whole request once its prompt passes 100,000 tokens, and OpenAI's multipliers apply to the whole request once it passes 272,000 input tokens. Both price steps are per request, not per extra token.

What long prompts do to the bill

The examples below use list prices without caching. They are illustrations, not measurements.

Below 100,000 tokens the two cost the same per token. Between 100,000 and 272,000 tokens, Haiku 5.5 charges five times its base rate while Luna stays at its base rate, so a 150,000-token document costs five times as much to read on Haiku 5.5. Above 272,000 tokens Luna's rate rises too, but less, and Haiku 5.5 still costs about 2.5 times as much in the example. If you want Haiku 5.5 for long documents, split them into chunks under 100,000 tokens.

Tokenizers also differ. Anthropic says Haiku 5.5 splits text into roughly 30% more tokens than Haiku 4.5 did, and Luna uses OpenAI's separate tokenizer, so one document can produce different token totals on the two models. Run a sample of real inputs through both APIs and compare the reported usage before committing a budget.

The benchmarks Anthropic published

Anthropic's launch table compares Haiku 5.5 directly with GPT-6 Luna. These are Anthropic's own runs, described in the Haiku 5.5 system card, and the OSWorld figures use its offline subset. Sonnet 5.5 is included only for reference.

Haiku 5.5 is ahead on each of the six rows it shares with Luna in this table. The gaps are widest on computer use and Terminal-Bench 4.0 and narrowest on FrontierCode, the mergeable-code test. OpenAI has not published Luna results on these benchmark versions; its own launch post reports Luna at 66.6% on DeepSWE v1.1 at max effort, a test Anthropic's table does not include. Treat the table as a reason to shortlist Haiku 5.5 for agent-style work, not as a verdict on your workload. Anthropic itself says Sonnet 5.5 and Opus 5.5 remain the better choice for complex agentic coding.

Which model fits which job

Platform and API differences

Claude Haiku 5.5 runs on the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS. It is also generally available in GitHub Copilot on Pro, Pro+, Max, Business and Enterprise, billed at Anthropic's list price. In Claude Code, the haiku alias selects it on the Anthropic API from version 2.1.293. Moving from Haiku 4.5 needs code changes: non-default temperature, top_p or top_k values, assistant prefill and manual budget_tokens all return errors, and responses can start with thinking blocks.

GPT-6 Luna runs in the OpenAI API through the Responses, Chat Completions and Batch endpoints, with web search, file search, computer use and MCP tools on the Responses API. EU data residency is available on Standard, Flex and Batch processing, and regional processing adds 10%. Chat Completions supports function calling only with reasoning effort set to none, and Luna cannot be fine-tuned.

Bottom line

At the same $0.10/$0.50 list price, GPT-6 Luna is the cheaper model whenever single prompts pass 100,000 tokens, and Claude Haiku 5.5 is the stronger candidate in Anthropic's published agent and knowledge-work results. For short, high-volume work, run a labeled sample through both and keep the one with more accepted outputs per dollar on your own data.