Claude Haiku 5.5 vs GPT-6 Luna: Cost and Benchmark Tested for Production Agents

Detail	Finding
Release date	Claude Haiku 5.5, October 7, 2026
Price (under 100k tokens)	$0.10 per million input tokens, $0.50 per million output tokens
Price (over 100k tokens)	$0.50 per million input, $2.50 per million output
Versus Haiku 4.5	About 75 percent cheaper on average, roughly 10x cheaper on short context input specifically
Versus Sonnet 5.5	Sonnet 5.5 input runs $2.00 per million, 20x Haiku 5.5's short context rate
New capability	First Haiku class model with adjustable effort: Low, Medium, High, Xhigh, Max
Benchmark comparison published by Anthropic	GDPval AA, AA Briefcase, OSWorld 2.1, Humanity's Last Exam, Terminal Bench 4.0, FrontierCode 1.1, Chartography, all measured against GPT-6 Luna, Haiku 4.5, and Sonnet 5.5

Anthropic's own release notes do something most small model launches skip: they put Haiku 5.5 next to a named competitor's cheap tier model, OpenAI's GPT-6 Luna, across eight benchmarks, rather than only comparing it to its own predecessor. That is useful, because Luna is not a hypothetical competitor. It is the model OpenAI just assigned to power ChatGPT's free and Go tiers as part of this week's GPT-6 rollout, which makes this the first real apples to apples look at what the two labs consider "good enough and cheap" for the volume end of their respective lineups.

What the benchmark numbers actually show

Across every published benchmark, Haiku 5.5 outscores GPT-6 Luna, and the gap is not small on agentic and computer use tasks specifically. On OSWorld 2.1's offline subset, a benchmark that measures whether a model can actually operate a computer environment to complete a task rather than just answer a question about one, Haiku 5.5 scored 72.4 percent against Luna's 48.9 percent. On Terminal Bench 4.0, which scores a model's ability to complete real command line tasks inside a sandboxed terminal, Haiku 5.5 hit 39.2 percent versus Luna's 16.4 percent, more than double. On Chartography, Anthropic's visual reasoning benchmark with no tool access, the gap was 46.4 percent to 29.1 percent.

The pattern across all eight benchmarks is consistent: Haiku 5.5 beats Luna by a meaningful margin on tasks that require taking actions in an environment (operating a computer, running terminal commands, using tools), and the gap narrows but does not close on tasks that are closer to pure question answering. On GDPval AA v2.1, a real world task completion benchmark Anthropic uses heavily across its model line, Haiku 5.5 scored 1620 against Luna's 1437, still ahead but the smallest relative gap of the set. None of this means Luna is a bad model for its price point; it means that if your workload actually involves an agent doing things, clicking, typing, running shell commands, rather than only generating text, the benchmark gap between these two specific cheap tier models is wider than the headline pricing comparison alone would suggest.

The honest caveat: these are Anthropic's own benchmarks

Every number above comes from Anthropic's release page, which means the benchmark selection, the prompting approach, and the tool access configuration for each model were chosen by the company selling Haiku 5.5. That is standard practice across the industry and does not make the numbers fabricated, OSWorld, Terminal Bench, and Humanity's Last Exam are independently maintained benchmarks that Anthropic did not invent, but it does mean you should treat the margins as directionally informative rather than as a guarantee that will replicate on your specific agent harness, your specific prompts, and your specific tool definitions. Before committing a production workload to either model, run your own held out evaluation set through both, ideally one built from real traffic your current agent has handled, not a public benchmark.

Why the effort setting changes the economics

Haiku 5.5 is the first model in Anthropic's Haiku line to expose an adjustable effort parameter, the same Low, Medium, High, Xhigh, Max scale Anthropic has used on its larger models. This matters more for a cost comparison than it might first appear, because it means Haiku 5.5 is not a single fixed point on the cost versus capability curve the way most small models are. A support ticket triage agent can run at Low effort for routine classification and jump to High only for ambiguous cases, with cost scaling accordingly inside a single model rather than requiring a fallback call to a larger, more expensive model when a request gets harder. GPT-6 Luna, as covered in OpenAI's own documentation, does not expose an equivalent per request effort dial; its cost profile is closer to fixed per token regardless of task difficulty. For an agent pipeline with genuinely variable task complexity, that architectural difference can matter more to your actual monthly bill than the per token price alone.

Where this fits against the rest of each lineup

Anthropic is explicit that Haiku 5.5 is not meant to replace Sonnet 5.5 or Opus 5.5 on complex agentic coding, and the benchmark gap backs that framing up: Sonnet 5.5 scored 70.6 percent on Terminal Bench 4.0 against Haiku 5.5's 39.2 percent, a real capability difference, not a rounding error. The release notes instead position Haiku 5.5 as a subagent that pairs with Opus 5.5 on coding work, handling the high volume, lower complexity steps in a pipeline while the larger model handles planning and the hardest edge cases. OpenAI's lineup follows a similar shape with GPT-6 Sol serving as the mid tier model and GPT-6 Luna as the free tier floor, which means the real comparison most teams should be running is not Haiku 5.5 against the entirety of OpenAI's lineup, it is Haiku 5.5 specifically against whichever OpenAI tier matches the volume and cost ceiling of the workload in question.

What two customer references suggest about real workloads

Anthropic's release notes cite two customer results worth noting with the usual caveat that vendor supplied case studies describe a specific configuration, not a universal outcome. HubSpot reports a 92.8 percent average score over three runs on an internal CRM evaluation suite, and AlphaSense reports a score of 0.84 on an internal metric versus 0.76 for Haiku 4.5. Neither company published the full methodology behind these numbers in Anthropic's materials, so treat them as directional evidence that real production users found Haiku 5.5 meaningfully better than its predecessor on their own workloads, not as benchmarks you can compare against your own evaluation.

When this applies to your stack

If you are running a high volume agent workload today on Haiku 4.5, GPT 5.5 mini, or an equivalent cheap tier model, and any part of that workload involves tool use, computer operation, or terminal style tasks rather than pure text generation, Haiku 5.5's benchmark margin over GPT-6 Luna on exactly those task types makes it worth a real side by side evaluation against whatever you are running now. If your workload is closer to pure classification or short form generation where the benchmark gap narrows, the pricing difference between Haiku 5.5 and Luna matters less and your existing vendor relationship, rate limits, and integration cost may be the deciding factor instead. Building that evaluation harness, and the effort tiered routing logic that lets you actually capture Haiku 5.5's adjustable cost curve in production, is the kind of agent infrastructure work we help engineering teams get right before they commit budget to one model family.

FAQ

How much does Claude Haiku 5.5 cost compared to Haiku 4.5? Haiku 5.5 is priced at $0.10 per million input tokens and $0.50 per million output tokens for prompts under 100,000 tokens, which Anthropic says is about 75 percent cheaper than Haiku 4.5 on average and closer to a 10x reduction on short context input specifically.

Is Claude Haiku 5.5 actually better than GPT-6 Luna, or just cheaper? Both. Across the eight benchmarks Anthropic published, Haiku 5.5 outscored Luna on every one, with the largest margins on tasks involving tool use, computer operation, and terminal commands. The gap is narrower on pure text generation and reasoning tasks.

Should I use Haiku 5.5 for complex agentic coding? No. Anthropic positions Sonnet 5.5 and Opus 5.5 as the right choice for complex coding work, with Haiku 5.5 better suited as a lower cost subagent handling high volume, lower complexity steps inside a larger pipeline.

What is the adjustable effort setting and why does it matter for cost? It is a Low through Max dial that lets you scale a single model's reasoning depth, and therefore its cost, per request based on task difficulty, rather than falling back to a larger model whenever a task gets harder. GPT-6 Luna does not currently expose an equivalent setting.