Claude Haiku 5.5 vs GPT-6 Luna: Which Cheap AI Model Should You Build On?
Last verified: October 13, 2026. All figures come from the vendors' own pages linked below; we have not run independent benchmarks.
The small-model tier just got a real fight. OpenAI shipped GPT-6 Luna on September 22, 2026 (Simon Willison), and on October 7 Anthropic answered with Claude Haiku 5.5, "the cheapest, fastest, and most capable small model we've ever released." The Haiku launch reached 1,047 points on Hacker News.
On paper the two cost the same for short prompts. The real difference is in where each one is strong, and what happens to price when your prompts grow.
| Criterion | Claude Haiku 5.5 | GPT-6 Luna |
|---|---|---|
| Input / output price (per 1M tokens) | $0.10 / $0.50 (prompts ≤100k) | $0.10 / $0.50 (prompts ≤272K) |
| Cache reads | $0.01 (≤100k) | $0.01 |
| Long-prompt price | $0.50 / $2.50 over 100k | 2x input, 1.5x output over 272K |
| Context window | Not stated on launch page | 1,050,000 tokens |
| Max output | Not stated on launch page | 128,000 tokens |
| Adjustable effort | Yes (first Haiku with it) | Yes: none → max |
| API model ID | claude-haiku-5-5 | gpt-6-luna |
Practical rule: at under 100k tokens per request they cost the same, so pick on capability. Above that, run the long-context math first.
Why "cheapest model" is the wrong question
Both vendors hit the same headline price. Anthropic's pricing table lists Haiku 5.5 at $0.10 input, $0.50 output and $0.01 cache reads per million tokens for prompts up to 100k. OpenAI's model page lists GPT-6 Luna at $0.10 input, $0.01 cached input, $0.125 cache writes and $0.50 output. Haiku's cache writes are also $0.125.
So the comparison comes down to capability per dollar and to the pricing edges:
- Haiku 5.5 jumps to $0.50 input / $2.50 output once a prompt is over 100k tokens. Anthropic says prompts of that size are only about 10% of past Haiku requests.
- GPT-6 Luna doubles input and cache rates and charges 1.5x output once a prompt is over 272K tokens. Batch and Flex are 50% of standard rates.
Benchmarks: what Anthropic's own table shows
The only head-to-head numbers come from Anthropic's launch post, so treat them as vendor-reported:
| Benchmark (Anthropic-reported) | Haiku 5.5 | GPT-6 Luna |
|---|---|---|
| GDPval-AA v2.1 (knowledge work) | 1620 | 1437 |
| AA-Briefcase v1.1 | 1578 | 1336 |
| OSWorld 2.1 offline (computer use) | 72.4% | 48.9% |
| Terminal-Bench 4.0 (agentic coding) | 39.2% | 16.4% |
| FrontierCode 1.1 Main | 46.4% | 42.4% |
| Chartography, no tools (visual reasoning) | 46.4% | 29.1% |
Source: Introducing Claude Haiku 5.5. OpenAI has not published a matching table, and we did not find independent replication at the time of writing.
Vendor benchmarks show you what a model can do, not what it will do on your workload. Re-run your own eval set.
Where Anthropic says Haiku still falls short
Anthropic itself says Sonnet 5.5 and Opus 5.5 "remain better choices for complex agentic coding tasks." It positions Haiku 5.5 for "compaction, summarization, or subagent work," live customer support and browser use.
Watch on YouTube: "Claude Haiku 5.5 Is a GAME CHANGER!" by WorldofAI
Third-party hands-on video by WorldofAI, an independent creator: "Claude Haiku 5.5 Is a GAME CHANGER! 90% CHEAPER & INSANE Performance! (Fully Tested)". AI Takeaway did not make this video and is not affiliated with its creator.
No time to watch the whole review? AI Takeaway is a tool that works on any YouTube video, including third-party reviews like this one. Paste the link to get the key points in minutes, jump to the original timestamps, and ask follow-up questions; every answer cites the video. You can also browse the AI Takeaway video summary hub for more AI and tech interviews.
What launch coverage and reviews add
Vendor posts give you the numbers. Hands-on reviews and interviews show how a model behaves in real use, such as speed, failure modes and prompting quirks. They are also long. When you go through them, keep claims you can check at a specific timestamp and set aside hype you can't verify. Weigh what you find against the official sources linked in this article.
Features and platform fit
GPT-6 Luna (docs) offers:
- Text and image input
- A 1,050,000-token context window and 128,000 max output tokens
- A knowledge cutoff of May 18, 2026
- Reasoning effort settings from
nonetomax - Built-in tools in the Responses API: web search, file search, code interpreter, computer use, MCP and more
- EU data residency
Claude Haiku 5.5 (launch post) is the first Haiku with an adjustable effort setting. It is available on the Claude Platform, Amazon Web Services, Google Cloud and Microsoft Azure. It also comes with new beta computer-use and browser-use support in Anthropic's Python and TypeScript SDKs. Early customer quotes in the launch post include Asana (over 30% lower task-completion latency) and HubSpot (92.8% on its CRM suite).
Which one to choose
High-volume summaries, classification and subagents
Choose Haiku 5.5 if you already run Claude Sonnet or Opus as the lead agent. Anthropic designed it as a subagent for those models, and its reported numbers are higher across the shared benchmarks.
Very long documents or huge context
Choose GPT-6 Luna when single requests regularly go past 100k tokens. Its documented window is 1.05M tokens, and its surcharge only starts above 272K.
Computer use and browser agents
Anthropic reports 72.4% vs 48.9% on OSWorld 2.1 offline. Test Haiku 5.5 first, but confirm the result on your own tasks.
Already on the OpenAI stack
GPT-6 Luna plugs straight into the Responses API's built-in tools, Batch/Flex discounts and EU data residency.
Practical checklist
- Measure your prompt-length distribution (what share is over 100k? over 272K?).
- Price a typical request on both models, cache reads included.
- Run 50–100 real tasks on each at matched effort settings.
- Compare latency, error rate and cost per finished task, not per token.
- Re-check the official pricing pages before launch, because prices in this tier have changed often.
Practical rule: short prompts + Claude stack → Haiku 5.5. Huge contexts + OpenAI stack → GPT-6 Luna.
Keep up without the watch time
Haiku 5.5 and GPT-6 Luna launched within weeks of each other, and new models, launch streams and interviews keep coming. If you would rather read than watch, the AI Takeaway video summary hub has free curated summaries of AI and tech interviews. They go well beyond this one comparison, for example:
- Sam Altman on Astra, AGI, and the future of OpenAI
- OpenAI Codex lead on the new shape of product work (Andrew Ambrosino)
- Sam Altman on AGI, Compute, and Human Agency
To keep up with ongoing AI news, check the hub regularly. You can also install the AI Takeaway Chrome extension to get key points right on YouTube, or submit any YouTube video yourself on aitakeaway.net.
FAQ
Are Claude Haiku 5.5 and GPT-6 Luna the same price?
For short prompts, yes: both list $0.10 per million input tokens and $0.50 per million output tokens. Long prompts are priced differently: Haiku goes up above 100k tokens, Luna above 272K.
Which model is better at coding?
On Anthropic's reported Terminal-Bench 4.0, Haiku 5.5 scored 39.2% and GPT-6 Luna 16.4%. Anthropic still recommends Sonnet 5.5 or Opus 5.5 for complex agentic coding.
What is GPT-6 Luna's context window?
OpenAI's docs list 1,050,000 tokens of context, a 922,000 max input and 128,000 max output tokens.
Where can I use Claude Haiku 5.5?
On the Claude Platform as claude-haiku-5-5, plus Amazon Web Services, Google Cloud and Microsoft Azure.
Sources
- Anthropic, Introducing Claude Haiku 5.5 (Oct 7, 2026)
- OpenAI, GPT-6 Luna model documentation
- OpenAI, Introducing GPT-6.1 Sol (lists GPT-6 Luna pricing)
- Simon Willison, Claude Opus 5.5, GPT-6 Sol, GPT-6 Luna, and a new price war (Sep 22, 2026)
- Hacker News, Claude Haiku 5.5 discussion
- Claude on X, launch post