Coasty Systems · Released, not on our provider

What is Coarena AI? The agent arena explained (2026)

Coarena AI is a website, run by Coasty Systems and backed by Y Combinator (Summer 2026), where two anonymous AI agents do the same real computer task and people vote on the better result before the models are named; on November 18, 2026 its arena rated 26 models.

Released, not on our providerLast updated:

Use it now

Chat

This chat starts with GPT-6.1 Terra, not Coarena AI.

≈ 2 credits per reply · See all models

New accounts get 5 free credits once, valid for 7 days. They work on the lower-priced models here; the priciest ones need credits you bought.

Start chatting — 5 free credits
Maker
Coasty Systems
Type
Text
Released
No date given
OpenRouter id
None

Checked

Coarena AI: facts, with sources

  • Coarena’s Computer Use Arena page describes itself as "Real tasks. Two anonymous agents. Blind human and model votes." A round goes task, two runs (A and B), blind vote, then rating. The home page says agents from the major AI labs do the task on a computer or an Android phone.Source: coarena.ai · checked 2026-11-18
  • The Governance page says Coarena is "built, owned and operated by Coasty Systems, Inc.", which also builds a computer-use agent, and is that company’s product, not a separate entity. The site header reads "Coarena by Coasty".Source: coarena.ai · checked 2026-11-18
  • Y Combinator lists CoArena in its Summer 2026 batch: founded 2026, team of 2, San Francisco, founders Prateek Jannu and Nitish Kovuru. Their bios say they previously built Coasty, a computer-use agent that reached 82% to 83% on the OSWorld benchmark (the founders’ own figures, not tested by us), and that they are "Now building CoArena".Source: ycombinator.com · checked 2026-11-18
  • Coarena’s Terms say that older sentences that name "Coarena" as the licensee mean Coasty Systems, Inc.; the earliest archived consent version, dated August 4, 2026, already uses the name Coarena. We found no page that states when or why Coasty became Coarena.Source: coarena.ai · checked 2026-11-18
  • The Governance page says the team’s own agent "is not on the roster": the roster refuses the lane that drives it, so no battle can be matched to it. It also says every competing agent is driven by one shared scaffold, with random side assignment, and that model identities are withheld by the API until a judgment is recorded.Source: coarena.ai · checked 2026-11-18
  • The separate Leaderboard page currently shows a Model API usage board ranked by tokens processed (for example, GPT-6 Luna first for the 7 days to 9:00 AM UTC), and says "This board shows usage and feedback only". It is not the arena’s preference ranking.Source: coarena.ai · checked 2026-11-18
  • The Terms say "there are no fees for using the Services" and that running an agent on your task costs Coasty, not you. Watching needs no account; posting a task or judging needs a Google sign-in and an age of 18 or older. The separate Model API sells prepaid credits that never expire and are not refundable, and the first purchase ends the free daily allowance.Source: coarena.ai · checked 2026-11-18
  • By signing in you license Coasty to use your tasks and votes in datasets it licenses to AI labs "for fees that Coasty charges and retains", and the Terms say you are not entitled to compensation, royalty or revenue share. The licence is revocable. The Privacy Policy says "Yes, we sell it" and explains that licensing datasets for money counts as a sale under California law.Source: coarena.ai · checked 2026-11-18
  • Coarena is a website and benchmark, not an AI model: no entry named Coarena or Coasty is on OpenRouter’s model list (478 models, checked November 18, 2026), so this site cannot run it. This site is a text chat with AI models.Source: openrouter.ai · checked 2026-11-18 · second-hand

Our table

Coarena AI arena ratings next to what the same models cost

  • The rating is "a Bradley-Terry fit over every admissible blind judgment", with 95% intervals clustered on the judge. A model with under 30 comparisons is listed but not rated, and judgments where both agents were bad count as ties. The page lists n = 26 models and shows no date for the standings.Source: coarena.ai · checked 2026-11-18

Rating and price

Coarena AI arena rating, OpenRouter price and typical reply cost

Coarena publishes the ratings but not what the models cost. This table adds the price. The top three, GPT-6.1 Terra, Claude Fable 5.2 and GPT-6.1 Sol, have overlapping 95% intervals, so the arena does not separate them; yet their per-token prices run from $1 / $5 to $8 / $40 per million tokens. The arena rates which result people preferred on computer tasks, not price or chat quality, and an agent run uses many more tokens than one chat reply, so read the cost column as a price guide only. Each model’s page here has the full sourced price, and Gemini 4 Argon and the others are linked from the home page directory.

Our table: ratings as read on coarena.ai on November 18, 2026; prices are the OpenRouter prices on each model’s page here; credits are our arithmetic · checked 2026-11-18
Arena rank and modelCoarena rating (±95%)Price per million tokens, in / outTypical reply here
1. GPT-6.1 Terra1112 ± 21$1 / $5about 2 credits
2. Claude Fable 5.21109 ± 18$8 / $40about 11 credits
3. GPT-6.1 Sol1107 ± 17$2 / $10about 3 credits
4. Hy4 (Tencent Hy4)1087 ± 23$0.417 / $1.2505about 1 credit
7. Gemini 4 Argon1049 ± 22$2 / $10about 3 credits
8. Claude Opus 5.51043 ± 20$4 / $20about 6 credits
9. Gemini 4 Carbon1038 ± 22$2 / $10about 3 credits
11. GPT-6.2 Sol1032 ± 69$2 / $10about 3 credits
18. Kimi K3984 ± 19$0.192 / $16about 4 credits
19. GPT-6 Luna973 ± 21$0.1 / $0.5about 1 credit
20. MiniMax M3.1 Flash972 ± 21$0.15 / $0.6about 1 credit
22. Claude Sonnet 5.5953 ± 22$2 / $10about 3 credits
24. DeepSeek V4.1 Flash 1017939 ± 25$0.3 / $1.2about 1 credit
25. Claude Haiku 5.5938 ± 21$0.1 / $0.5about 1 credit

Source: coarena.aiSource: openrouter.ai

Nearest available

Alternatives to Coarena AI you can use now

$1 / $5 (per million tokens)

  • OpenAI’s API changelog has an October 26, 2026 entry: "Released GPT-6.1 Terra", model gpt-6.1-terra, on the Responses and Chat Completions endpoints.Source: developers.openai.com · checked 2026-11-17
  • OpenAI’s model page calls it "A fast, efficient model for visual reasoning, precise computer use, and clear, natural writing." It lists these uses: matching product photos to catalog entries, interpreting technical drawings, acting on application screens, and drafting or revising text.Source: developers.openai.com · checked 2026-11-17
  • OpenAI’s models page lists four flagship models per million input / output tokens: GPT-6 Astra $10 / $50, GPT-6.1 Sol $2 / $10, GPT-6.1 Terra $1 / $5 and GPT-6 Luna $0.10 / $0.50. Its "Choosing a model" section names Astra, Sol and Luna but gives no guidance on when to pick Terra.Source: developers.openai.com · checked 2026-11-17

$8 / $40 (per million tokens)

  • Released on November 9, 2026 (Anthropic’s model page marks it "Latest").Source: platform.claude.com · checked 2026-11-13
  • Anthropic describes it as "for demanding reasoning and long-horizon agentic work": hours-long agent tasks, multi-step deep research, and finished documents, spreadsheets and presentations.Source: platform.claude.com · checked 2026-11-13
  • Context window 1M tokens; maximum output 128K tokens; text and image input, text output.Source: platform.claude.com · checked 2026-11-13

FAQ

Coarena AI FAQ

What is Coarena AI?

It is a live arena at coarena.ai: you give a real computer task, two anonymous AI agents do it, and you vote on the better result before Coarena reveals which models they were. The votes feed a public ranking and datasets that Coarena licenses to AI labs. It is a benchmark site, not an AI model or chatbot.

Is Coarena the same as Coasty?

They are linked. The site header reads "Coarena by Coasty", the Terms name Coasty Systems, Inc. as the operator, and the founders’ Y Combinator bios say they previously built Coasty, a computer-use agent. We found no page stating when or why the name changed, and the team’s own agent is not in the arena.

Who is number one on the Coarena leaderboard?

In the Computer Use Arena standings we read on November 18, 2026, GPT-6.1 Terra was first at 1112 ± 21, then Claude Fable 5.2 at 1109 ± 18 and GPT-6.1 Sol at 1107 ± 17. Their intervals overlap, so the top three are not clearly separated. Coarena’s separate Leaderboard page ranks Model API usage, not preference.

How does Coarena rank the agents?

It fits a Bradley-Terry model to blind judgments, with 95% intervals clustered on the judge. Models with under 30 comparisons are listed but not rated, and "both bad" judgments count as ties. A new submitter’s first three battles pair the two top-rated agents and count at a quarter weight.

How much does Coarena cost?

The Terms say there are no fees for using the arena. Watching needs no account; posting a task or voting needs a Google sign-in and an age of 18 or older. The separate Model API has a free daily allowance for some models and sells prepaid, non-refundable credits.

What happens to the tasks and votes I submit?

By signing in you license Coasty to use them in datasets it licenses to AI labs for fees it keeps, and you get no compensation. The licence is revocable, and the Privacy Policy says it counts this as selling data. Read the Terms and Privacy Policy before you submit anything private.

Can I chat with Coarena here?

No. Coarena is a website, not a model, and it is not on OpenRouter’s model list, so this site cannot run it. You can chat with the models it rates: on this site a typical reply to GPT-6.1 Terra costs about 2 credits (a credit is about $0.01 in the $5 pack), billed by its real token cost, at least 1 credit.

Is this the official Coarena website?

No. This page is unofficial and not affiliated with or endorsed by Coarena or Coasty. Facts link to the pages we read.