Claude Haiku 5.5 vs GPT-5.6 Luna: Half the Sticker Price, 19% More per Finished Task

The price match everybody wrote about is real, and it is not with this model. Claude Haiku 5.5, which the vendor launched on October 7, 2026, lists at $0.10 per million input tokens and $0.50 per million output tokens up to 100,000-token prompts. GPT-5.6 Luna, the cost-optimised tier from July 9, 2026, lists at $0.20 and $1.20. So the vendor is exactly half the older model's rate on both meters — and exactly equal to GPT-6 Luna, the successor that replaced it on September 22, 2026 and which is the model the vendor's own launch post benchmarks against.

That is a strange place to be writing a comparison from, so it is worth being precise about what this matchup is for. GPT-5.6 Luna is the only one of the two that is deprecated on the independent board, it is no longer the Luna in OpenAI's flagship rate card, and it is still the Luna you will find on a lot of integration code today. Comparing it to Claude Haiku 5.5 is a question about a migration you may already owe, not a question about which of two current models to call.

And the answer to the obvious question is the interesting part. On the same independent evaluation run, one finished Intelligence Index task costs $0.21 on Claude Haiku 5.5 and $0.18 on GPT-5.6 Luna. Half the token price, 19% more per answer.

The two rate cards

Per million tokens in US dollars. Anthropic's and OpenAI's rates come from their own documentation; the shared scores are Artificial Analysis Intelligence Index v4.3.2, read 2026-10-08.

• Input — Claude Haiku 5.5 $0.10 up to 100,000 tokens and $0.50 above; GPT-5.6 Luna $0.20, doubling to $0.40 for requests over 272,000 tokens.

• Output — Claude Haiku 5.5 $0.50 up to 100,000 tokens and $2.50 above; GPT-5.6 Luna $1.20, rising to $1.80 on the long-context path.

• Cached input — Claude Haiku 5.5 $0.01 and $0.05, a 90% discount; GPT-5.6 Luna $0.02, also 90%, plus a cache-write charge of $0.25.

• Context — 1M tokens for Claude Haiku 5.5 and 1,050,000 for GPT-5.6 Luna. Both cap a single response at 128,000 tokens, and OpenAI publishes a February 16, 2026 knowledge cutoff for its model.

• Reasoning control — Claude Haiku 5.5 exposes an effort dial from Low to Max with medium as the default; GPT-5.6 Luna exposes six settings, none / low / medium (default) / high / xhigh / max.

• Independent score — 43.40 for Claude Haiku 5.5 (Max) against 37.32 for GPT-5.6 Luna (Max), a 6.07-point gap on a 182-model board.

• Status — Claude Haiku 5.5 is current, proprietary, API-only. GPT-5.6 Luna carries a deprecation flag on the independent board and no longer holds a row in OpenAI's flagship pricing table.

Half the price, more expensive per answer

The rate-card arithmetic here is unusually favourable to Anthropic — a clean 2x on both meters inside the short-prompt regime — so the per-task result is the one that needs explaining.

Claude Haiku 5.5 emits 162,164 output tokens per Intelligence Index task, split 129,047 reasoning and 33,118 answer. GPT-5.6 Luna emits 41,235, split 28,098 reasoning and 13,136 answer. Anthropic's model spends 3.9 times as many tokens on the same job, and at half the output rate that lands it 19% more expensive per finished task. The rate advantage and the verbosity disadvantage are within a factor of two of each other, so they very nearly cancel — and the residual goes OpenAI's way.

The two per-task costs are close, but the way each is composed is not. GPT-5.6 Luna's $0.1783 per task is roughly half cache writes and a quarter output; Claude Haiku 5.5's $0.2128 is $0.1317 of input — mostly cached reads — plus $0.0811 of output. Neither model is expensive in absolute terms, and they are expensive in different places, which is what makes the per-task totals converge while the rate cards sit a factor of two apart.

This is the same effect that makes a per-token comparison of reasoning models unreliable in general, and it is worth naming plainly: a cheap meter multiplies with a heavy token budget, and the product is what you pay. Anthropic's own launch notes are candid on this, saying Claude Haiku 5.5 is "especially good value when used for tasks with prompts up to 100,000 tokens, which make up around 90% of requests to our previous Haiku model."

The blend math tells the same story from the other direction. On Artificial Analysis's 7:2:1 input-heavy blend — the shape most production traffic actually has — Claude Haiku 5.5 comes to $0.077 per million tokens against $0.174 for GPT-5.6 Luna. That is the 2.26x you would expect from the sticker. It is only when you weight the output meter the way a reasoning workload does that the two converge, and past 100,000 tokens they cross.

The 100,000-token cliff against the 272,000-token one

Both vendors reprice long requests, and the thresholds are more than two and a half times apart.

Anthropic's break is at 100,000 input tokens. Above it, Claude Haiku 5.5's meters become $0.50 in and $2.50 out — five times the short-prompt rate, applied to the whole request. OpenAI's break is at 272,000 input tokens. Above it, GPT-5.6 Luna's meters become $0.40 in and $1.80 out — 2x input and 1.5x output, also applied to the whole request.

So the crossover is not a slow convergence; it is a step. A 120,000-token retrieval prompt pays Anthropic's long tier and sits comfortably inside OpenAI's short tier, which flips a 2x Anthropic advantage into a 1.25x OpenAI advantage on input and a 1.39x advantage on output. And Anthropic's threshold is low enough that a document-heavy pipeline crosses it routinely, while OpenAI's is high enough that most such pipelines never see it.

Under 100,000 tokens nothing about this changes the ordering. Over 272,000 the ordering inverts and stays inverted.

What replaced the model you are comparing against

This is the part that should decide how much time you spend on the comparison at all.

OpenAI shipped GPT-6 Luna on September 22, 2026 — fifteen days before Anthropic shipped Claude Haiku 5.5 — at $0.10 in and $0.50 out, with a 1M-token context, a 128,000-token response ceiling, and the same six-step reasoning effort control. That is the identical rate card to Claude Haiku 5.5, on both meters, in the short-prompt regime.

On the independent board the two are close enough that the composite ordering does not settle it: Claude Haiku 5.5 scores 43.40 against GPT-6 Luna's 38.12 — still a 5.28-point Anthropic lead — but the cost per finished task is $0.21 against $0.068. That is a threefold OpenAI advantage on the number that actually bills, and it exists for the same verbosity reason as above: GPT-6 Luna emits 50,012 output tokens per Index task against Claude Haiku 5.5's 162,164.

Anthropic's launch post does the comparison itself, in a table that puts Haiku 5.5 against "GPT-6 Luna" on GDPval-AA v2.1 (1620 against 1437), AA-Briefcase v1.1 (1578 against 1336), OSWorld 2.1 offline subset (72.4% against 48.9%) and Terminal-Bench 4.0 (39.2% against 16.4%). Those are Anthropic's own numbers and they are vendor-reported, but the choice of opponent is the tell: the model Anthropic picked to beat is the one after the one in this article's title.

Where the old Luna still wins

Three things, and none of them is capability.

Video and audio. Neither model is video-native — Claude Haiku 5.5 reads text and images, and GPT-5.6 Luna reads text and images — so this row is a tie and both lose to anything with a media pipeline.

Cache writes. OpenAI publishes a $0.25 per million cache-write rate for GPT-5.6 Luna and Anthropic publishes none for Claude Haiku 5.5, which means the honest statement is that the write cost is inside Anthropic's $0.10 input meter rather than absent. A prefix-heavy workload with a short-lived cache pays that difference every turn.

And the long-context path itself, where OpenAI's higher threshold is a genuine structural advantage that has nothing to do with which model is smarter.

Both are callable, one of them is not ours

GPT-5.6 Luna is on the OrcaRouter catalogue today at openai/gpt-5.6-luna — $0.20 in, $1.20 out, a $0.25 cache write, 1,050,000-token context and a 128,000-token response ceiling, with the 272,000-token long-context tier carried as a separate step in the pricing record rather than averaged away. Passed through at OpenAI's list price with 0% markup, and automatic failover underneath, which is worth something specific for a deliberately superseded model: if OpenAI's own capacity for it tightens, the route moves before your error budget does.

Claude Haiku 5.5 is not on the catalogue. It is reachable through Anthropic's API and through AWS, Google Cloud and Microsoft Azure, and that is the accurate way to put it. What the catalogue does carry from Anthropic is Claude Sonnet 5.5 and Claude Opus 5.5, so the escalation path out of a cheap classification tier and into a mid-tier agentic one is a routing setting rather than an integration — and the same catalogue holds GPT-6 Luna at the exact rate card Claude Haiku 5.5 uses, which makes the honest A/B for this price point a two-string change rather than a procurement exercise.

What to do with this comparison

If your integration names gpt-5.6-luna, the useful conclusion is not that Claude Haiku 5.5 beats it by six index points. It is that you are on a model your vendor has already superseded twice in its own flash tier — first by Terra's generation, then by GPT-6 Luna, which Anthropic itself treats as the current competitor at this price. Migrating to GPT-6 Luna is a rate-card match and a threefold drop in cost per finished task on the same independent measurement.

If you are choosing between Claude Haiku 5.5 and the Luna line on capability, Anthropic wins on the composite, on terminal-and-code rows, and on computer use, and its own vendor table says so. If you are choosing on the invoice, the Luna line has won every version of this comparison by a wide margin, and the reason is a single number: 162,164 output tokens per task against 41,235.

The thing this matchup actually settles is that a two-times cheaper meter is worth less than a four-times leaner model. Anthropic priced Claude Haiku 5.5 to match GPT-6 Luna and it did — exactly, on both meters — and it still costs three times as much to get an answer out of it.