Does Haiku 5.5 really cost the same as GPT-6 Luna?

This is the video written out, with every figure in full, each linked to its source in the table below.

Three Claudes in fifteen days▶ 0:00

Anthropic released Opus 5.5 on 22 September 2026 and Sonnet 5.5 six days later. Haiku 5.5 followed on 7 October, which makes three new Claude models in fifteen days. Haiku 5.5 is the small, cheap one, and for prompts below a certain size it costs exactly what OpenAI's cheapest GPT-6 model, GPT-6 Luna, costs.

Same price, different bill▶ 0:22

For prompts up to 100K tokens, both charge $0.10 per million input tokens and $0.50 per million output tokens.

Artificial Analysis runs an independent score, the Intelligence Index, built from 10 tests. On it, Haiku 5.5 costs about three times as much per task as GPT-6 Luna at the same effort setting (effort is how hard you ask a model to think). At medium it costs 2.7× as much, at high 2.7×, at xhigh 2.9× and at max 3.1×. At low the gap is wider still, 5.4×: $0.024 a task against Luna's $0.0045. Luna's figure is from Artificial Analysis's full model list of 8 October; the release page we read on 15 October has no row for Luna at low.

The price per token is the same, but Haiku uses far more tokens, reading as well as writing. Over the whole index run at max, it wrote 435.18M output tokens against Luna's 144.38M, about 3.0× as many, and read 7733.95M input tokens against Luna's 3146.39M, 2.5× as many, with each company's own API counting the tokens. At medium Haiku wrote 1.9× as many and read 1.9× as many. Most of each model's cost per task is input: 62% of Haiku's at max. By Artificial Analysis's own cost split, 64% of the gap between the two at medium and 61% at max is input.

The basics▶ 1:02

Haiku 5.5 is Anthropic's small model. Per token it costs 10% of what Haiku 4.5 did, which was $1 per million input tokens and $5 per million output tokens.

Anthropic puts the average saving at around 75%, not 90%. Its footnote gives two reasons. Haiku 5.5 is 90% cheaper than Haiku 4.5 only for requests up to 100K tokens; above that it is 50% cheaper, and 90% of Haiku 4.5's requests were below the line. And Haiku 5.5 has a new tokenizer: according to Anthropic's docs, the same text counts as about 30% more tokens than on Haiku 4.5.

Both models have a context window of about a million tokens: 1,000,000 for Haiku 5.5 and 1,050,000 for GPT-6 Luna. Haiku 5.5 is the first Haiku with effort settings: low, medium, high, xhigh and max, with medium as the default. GPT-6 Luna has the same five plus a setting with no reasoning at all, and Haiku 5.5 can switch its thinking off at high effort or below.

On the Intelligence Index, Haiku 4.5 (with reasoning) scored 16.9. Haiku 5.5 scores 34.5 at medium and 43.4 at max. GPT-6 Luna scores 29.9 at medium and 38.1 at max.

Matched by cost, a tie▶ 2:13

Comparing the two at the same setting has a catch, because the same setting doesn't mean the same amount of work. Haiku costs more at each setting, but it also scores higher at each one, so here they are matched by what you pay instead.

Haiku at medium costs $0.047 a task and scores 34.5. GPT-6 Luna at xhigh costs $0.042 and scores 34.6. One step up, Haiku at high costs $0.079 and scores 37.8, and Luna at max costs $0.068 and scores 38.1. Over that stretch, paying about the same gets you about the same score, with Luna a little cheaper.

They come apart at the top. Luna's best score is at max, 38.1. Haiku at max reaches 43.4, for $0.213 a task, 3.1× what Luna costs at max.

The two tables below have the figures from Artificial Analysis that the video uses, at every setting, as we read them for the video. Its other test scores are on its release page.

The painting test▶ 3:10

This is the test we run on new models. Each one looks at Hokusai's Great Wave and paints it by writing a painting program, with no image model and no image library. We ran each model once at each setting shown.

Haiku 4.5 didn't paint a wave at all. Haiku 5.5 at medium painted a recognisable Great Wave, with the curling crest, the boats and Mount Fuji, in 329 s for $0.04 at API prices. At max it took 859 s and cost $0.77. It used more tokens, and its context grew past 100K tokens, where the price goes up (the section on the 100K-token step below has the arithmetic). To our eye that painting isn't clearly better. GPT-6 Luna at max painted the wave too, in 317 s, flatter and with more empty sky; Codex doesn't report a dollar cost for it. Sonnet 5.5 and Opus 5.5, painted on 28 September, are still the most detailed.

Five models, six runs, one prompt: the Great Wave

These are single runs, so they show what each model can do, not how often it does it.

Where Luna is the better buy▶ 4:11

For its Coding Agents board, Artificial Analysis runs models inside coding tools, Haiku 5.5 in Claude Code and GPT-6 Luna in Codex, and averages three tests. Haiku's best score there is 41.4, at xhigh, for $0.61 a task. Luna at max scores 41.1 for $0.18. That is about the same score for 0.29× the cost, under a third.

Luna draws level thanks to one of the three tests, a set of software engineering tasks called DeepSWE, where Luna at max gets 63.7% against Haiku at xhigh's 49.0%. Haiku is ahead on the other two. Turning Haiku up to max makes it worse overall on this board: 36.5, at $2.58 a task.

Then a test of our own, one job with one run per setting. We gave both models 610 real messages sent through a tutoring business's website contact form. Simple keyword rules had already labelled them: 128 genuine inquiries and 482 spam. Those labels are sometimes wrong themselves, so what we measured is how often each model agrees with them. We ran each model at low and at medium and kept its better result.

GPT-6 Luna at low agreed with the labels on 88.9% of the messages and Haiku 5.5 at medium on 82.0%. Message by message, that is a difference of +6.9 points, with a 95% interval from +4.4 to +9.3. By the labels, Haiku let 103 spam messages through and Luna 57, though Luna marked 11 genuine ones as spam, to Haiku's 7. Luna cost $0.081 per thousand messages and Haiku $0.176.

Each model's other setting did no better. Haiku at low minus Haiku at medium comes to −2.8 points (95% interval −5.1 to −0.7). Luna at low and at medium were within a point of each other: low minus medium is +0.5 points (interval −0.7 to +1.6). Haiku read more input tokens per message (775 against Luna's 409) and wrote more (198 against 81), which is where its higher cost comes from.

Where Haiku is ahead▶ 5:48

Artificial Analysis also runs a knowledge test, AA-Omniscience, which takes points off for a wrong answer and none for saying "I don't know". From medium to max, Luna knows more: it gets 42.8% to 44.2% of the questions right, Haiku 34.0% to 36.4%.

When a model doesn't have the right answer, though, Haiku makes one up 40.4% to 44.8% of the time, and Luna 76.7% to 84.7% of the time. So when it doesn't know, Haiku makes things up about half as often, because more often it gives no answer or only part of one. At max, Haiku answers 36.4% right, 25.7% wrong and leaves 37.9% unanswered or partly answered; Luna answers 43.8% right, 43.1% wrong and leaves 13.1%. On the test's own score, right answers minus wrong ones, Haiku comes out ahead: +10.67 at max against Luna's +0.65.

Haiku also leads on Terminal-Bench 4.0, a set of tasks done at a command line. At max it passes 32.8% of them in Artificial Analysis's run and 36.36% in Vals's, another independent tester, which also runs both models at max; Luna passes 12.6% and 13.64%. Artificial Analysis gives every model the same basic agent setup for this test. Terminal-Bench is also one of the three tests on the Coding Agents board, and Haiku leads it there too, inside the coding tools: 23.7% at xhigh and 29.8% at max, against Luna's 15.2%. So Luna's tie on coding comes from the software engineering tasks, not from command-line work.

Vals runs many more tests than the video had time for. Here is every one it lists for both models.

On computer use, Anthropic's own numbers give Haiku a partial-credit score of 72.4% on a subset of 82 OSWorld 2.1 tasks and Luna 48.9%, both at max. Anthropic ran Luna itself, through OpenAI's API. Counting only tasks finished completely, Haiku's pass rate is 37.1%.

Haiku also streams its words faster than Luna, but it writes more, and at max it waits much longer before it starts. The video showed Artificial Analysis's figures as we read them on 15 October: at max, Haiku wrote 240now 243 tokens a second and Luna 128now 132, and Haiku took 237 snow 267 s to its first token against Luna's 90 snow 116 s. Per index task, Haiku took 209.9 snow 188.4 s against Luna's 163.5 snow 176.3 s at high, 295.4 snow 288.7 s against 247.3 snow 246.5 s at xhigh, and 424.8 snow 428.7 s against 433.8 snow 421.0 s at max. So on the index tasks it takes about as long as Luna, or longer, at the same setting.

Update, 17 October: Artificial Analysis has re-measured speed since our read, and each new value is shown beside the old one above. Haiku is still the faster writer at every setting it measured for both, and now takes longer than Luna per index task at all three settings. Every score, cost and price in the video is unchanged.

The 100,000-token step▶ 7:41

Below 100K tokens, the two price lists match on input and output. Once a Haiku prompt goes over 100K tokens, including any part read from the cache of earlier requests, the whole request costs 5× as much: $0.50 per million input tokens and $2.50 out. GPT-6 Luna's price doesn't move until 272K input tokens. Then, also for the whole request, input costs 2× and output 1.5× as much, $0.20 in and $0.75 out.

So between the two lines Haiku costs 5× as much per token, and past the second line 2.5× as much for input and 3.3× for output.

The prompts that cross the line are long documents, big codebases and long agent runs, where everything so far is sent again with each step. Our max-effort painting crossed it. Priced entirely at the lower rates, its own token counts come to $0.195, already 4.8× the medium run, because it used more tokens. Crossing the line partway through multiplied that by 3.9×, so at list price it cost 19× as much as the medium run. The factor is below 5× because some of the run's requests stayed under the line.

If you pay with a Claude plan▶ 8:47

If you use Claude through a Pro or Max plan, the usage limit matters more than the price per token. Anthropic's Claude Academy says "Haiku is the lightest, Sonnet is moderate, Opus is heavy", without giving numbers. OpenAI at least publishes estimates for Codex: on Plus, 350 to 3,000 local messages with GPT-6 Luna every five hours, for Codex running on your own computer.

For Haiku 5.5, a few people have checked their usage bars themselves. Tony Stevenson, on his blog Rambling Rows, ran the same 7 programming jobs on Sonnet 5.5 and Haiku 5.5 on a Pro plan. A run on Sonnet used 6 or 7 points of the five-hour meter. He ran Haiku 8 times in a row and the meter went from 22% to 30%, about one point a run. He calls that a ceiling: the meter shows whole points, and his own session drew on the same limit.

まっちゃん投資ラボ, writing on note.com on a Max 5x plan, got 1 point for Haiku 5.5 on one task at max effort, against 6 for Sonnet 5.5 and 7 for Opus 5.5, one run each. The Sonnet and Opus readings include about one point used by a judge model.

AI Dive, in a post on DEV, ran 140 small Haiku 5.5 tasks (70 at low and 70 at medium) without moving the five-hour bar off 21%. 21 Opus 5.5 tasks moved it from 22% to 23%.

Haiku is light on your limit, then, but nobody has measured exactly how light. And the price list is no guide to it. The one precise measurement we found, by tksfjt1024 on Zenn, was of the older Haiku 4.5 in September, on one Max 20x account. Per input token, the limit counted Haiku 4.5 at about half the weight of Opus 5.5 (2.2 to one, Opus to Haiku), even though Haiku 4.5's input price was a quarter of Opus's (4 to one).

To have Claude Code hand its small jobs to Haiku, set the model in the agent's file (model: haiku) or with environment variables (CLAUDE_CODE_SUBAGENT_MODEL=haiku, plus CLAUDE_CODE_SUBAGENT_MODEL_FORCE=1 to apply it to every subagent, the built-in Explore and Plan ones included), because most subagents run on your main model by default. Some of the most-liked posts about Haiku 5.5 recommend a --subagents flag. That flag doesn't exist. We checked Claude Code 2.1.296 when we made the video, and on 17 October the newest release, 2.1.302, still had no such flag.

Which one for what▶ 10:38

For general work at a fixed budget, the Intelligence Index says they're about level. For coding agents paid per token, Luna in Codex: about the same score as Haiku in Claude Code for under a third of the price, and well ahead on software engineering tasks. For agents that mostly work at a command line, for computer use going by Anthropic's own numbers, for work where a confident wrong answer costs you, and for the highest index score if you'll pay for it: Haiku. For sorting jobs like our test, and for prompts past 100K tokens: Luna, on the evidence here. And if you pay through a Claude plan, give the small jobs to Haiku. It's clearly light on your limit, even if nobody can say yet by how much.

AI Actions (made by the company behind this channel)▶ 11:23

Sorting came up earlier, and it's what our own app does too. The company behind this channel makes AI Actions, an iPhone app. One of its actions for iOS Shortcuts is called Choose Category, and it runs on Jev, a model from TypeSafe. It sorts a piece of text, like a reminder about an invoice, into categories you set. You choose how sure it has to be, and when it isn't sure enough about any category, it says it's unsure instead of guessing. It's linked in the video's description.

There'll probably be another new model next week, and we'll measure that one too. If you've used Haiku 5.5 or GPT-6 Luna on real work, tell us in the comments under the video what you ran and which one you kept.