On October 7, Anthropic released Claude Haiku 5.5, its fastest model, and immediately put it up against OpenAI's GPT-6 Luna, which came out on September 22. In Anthropic's tests Haiku beats Luna across the board, but in live Agent Arena sessions Luna ranks higher. The per-token price of the two models is nearly equal in our price list, so the choice comes down to the task.
All the models in one table
Haiku 5.5 is the first model of its class from Anthropic with an adjustable effort setting: less effort is cheaper and faster, more effort is smarter. Luna has had such a setting from the start. Next to the new models we put their predecessors:
| Haiku 5.5 (current) | GPT-6 Luna (current) | Haiku 4.5 (previous) | GPT 5.6 Luna (previous) | |
|---|---|---|---|---|
| Price on GPTunneL, $ per 1K, input / output | 0.0003 / 0.0015 | 0.0004 / 0.002, now 0.0002 / 0.001 | 0.002 / 0.01 | 0.0004 / 0.0024 |
| Context | 1M | 1.05M | 200K | 1.05M |
| OSWorld 2.1, computer use | 72.4% | 48.9% | 15.7% | no data |
| Terminal-Bench 4.0, terminal coding | 39.2% | 16.4% | 0.0% | no data |
| FrontierCode 1.1 Main, coding agents | 46.4% | 42.4% | not published | no data |
| Chartography, charts without tools | 46.4% | 29.1% | 6.4% | no data |
| GDPval-AA v2.1, work tasks, Elo | 1620 | 1437 | 735 | no data |
| AA-Briefcase v1.1, work tasks, Elo | 1578 | 1336 | 614 | no data |
| Agent Arena, rank (net score) | 27th (+1.88%) | 19th (+3.26%) | no data | 37th (-1.80%) |
Bold marks the two current models and the better result between them in every row with benchmarks. The test results come from Anthropic's announcement, so this is vendor data: the company ran Luna itself through OpenAI's API. GPT-6 Luna is currently half its regular price because of the 50% launch discount; see the pricing page for current rates. "No data" means the model did not take part in that test or the source has no result.
What the table shows:
- Haiku 5.5 vs GPT-6 Luna. Haiku wins all six of Anthropic's tests: by the most on Terminal-Bench (39.2% vs 16.4%) and OSWorld (72.4% vs 48.9%), by the least on FrontierCode (46.4% vs 42.4%). But in Agent Arena, where models are rated in real sessions, Luna is higher, and the per-token price of the two is nearly equal. Winning the tests does not always mean winning at work.
- Haiku 5.5 vs 4.5. The price is 6-7 times lower, OSWorld went from 15.7% to 72.4%, and Terminal-Bench from zero to 39.2%. If you are still on Haiku 4.5, the switch pays off right away.
- GPT-6 Luna vs 5.6 Luna. Input costs the same, output is a little cheaper ($0.002 vs $0.0024), and in Agent Arena Luna climbed from 37th place to 19th. The settings differ: 5.6 on xHigh, 6 on Max.
Two caveats. OSWorld here is an offline set of 82 tasks, and 72.4% and 48.9% are partial-credit scores: on a strict pass rate Haiku solves 37.1% of tasks (system card), and the card gives no strict score for Luna. The independent Artificial Analysis index at maximum effort also favors Haiku, 43 vs 38, but at high effort Haiku scores the same 38 as Luna at max.
Where Luna is ahead
On knowledge without search, Luna is more accurate: 44% correct answers in AA-Omniscience against 36% for Haiku. On the other hand, Haiku is less likely to make up an answer when it does not know: its share of invented answers is 40% against 77%. This is an Artificial Analysis measurement from October 7, as relayed by Latent Space.
What Arena votes say
On arena.ai models are rated by users. The data is from October 13-14.
Agents (Agent Arena, 2.58M sessions). The ranking is built from task success, user ratings and steerability. Luna is higher: a net score of +3.26% (±0.49) against +1.88% (±0.98) for Haiku, and the intervals barely overlap. On steerability the models are equal (+6.17% and +6.40%). The settings differ: Haiku on high, Luna on max.
Web development (Code Arena, WebDev). This one is a tie: Haiku has 1585 points (±12, 2,771 votes), Luna 1582 (±6, 17,869 votes). Three points is less than the margin of error, and Haiku has six times fewer votes, so the number will still move.
In other words, on synthetic agent tests Haiku leads, while in real agent work Luna comes out higher. Which result is closer to your case is up to you.
The price of a task, not a token
Token use matters more than the price list. According to Artificial Analysis, at maximum effort Haiku spends about 162 thousand output tokens per index task, and Luna about 50 thousand. Haiku at high effort spends about 55 thousand, close to Luna at max.
An example at GPTunneL prices: 30 thousand input tokens and 3 thousand output tokens.
| Scenario | Haiku 5.5 | GPT-6 Luna (without / with discount) |
|---|---|---|
| Same output, 3K | $0.0135 | $0.018 / $0.009 |
| Haiku at max: three times the output, 9K | $0.0225 | $0.018 / $0.009 |
Haiku's token price is lower, but at triple the usage a task costs more than on Luna, even without the discount. So set the effort deliberately and measure cost on your own requests.
So which to choose: Haiku 5.5 or Luna 6
| Task | Model | Why |
|---|---|---|
| An agent controls a computer or terminal | Haiku 5.5 | about a 23-point lead in Anthropic's tests, but check it on your own scenarios: in live Arena sessions Luna is higher |
| Factual questions without search | Luna | 44% correct in AA-Omniscience against 36% |
| Better to stay silent than to invent | Haiku 5.5 | 40% invented answers against 77% |
| Layout and web development | either | a tie on Arena, 1585 vs 1582 |
| Bulk requests on a tight budget | Luna | three times fewer tokens at max effort, and a 50% discount right now |
What this means for you
Do not choose from a single table. Run ten of your own real requests through both models at the same effort and compare not tokens but the cost of an accepted result: how much an answer you did not have to redo costs. If Haiku wins, the extra tokens pay off in quality; if not, Luna is cheaper.
Where to try them
Both models are on GPTunneL: Claude Haiku 5.5 and GPT-6 Luna, one balance for both calls. On choosing between Sonnet and Opus in the same lineup, see Sonnet 5.5 or Opus 5.5; on the Luna release itself, see GPT-6 Sol and Luna.
Sources
- Claude Haiku 5.5 announcement, Anthropic, 07.10.2026 - benchmarks, effort settings, test methodology.
- Claude Haiku 5.5 documentation, Anthropic - context window, release date.
- GPT-6 Luna in the API, OpenAI - context window, effort settings.
- Claude Haiku 5.5 system card, Anthropic - strict and partial OSWorld scores, methodology.
- AINews: Claude Haiku 5.5, Latent Space, 07.10.2026 - a recap of Artificial Analysis measurements: index, token use, AA-Omniscience.
- Arena, Code Arena WebDev on 13.10.2026 and Agent Arena on 14.10.2026.