Back to the trend of the day: OpenAI's new AI does office work on your computer, at half price
Poly Rating · from the trend of the day

Best Computer Use AI Agent 2026: Cost per Finished Task

On OpenAI's own OSWorld run, GPT-6.1 Terra cost $1.19 per task at half the token price, four cents more than GPT-6.1 Sol at $1.15.

  • No. 1: GPT-6.1 Terra
  • 10 places

01Final list

The leading AIs each built their own list of ten independently. Here is one result that is easy to check: every place has a reason, a source, and the list of AIs that included it.

Order: from most to least AI agreement.

A bright badge means this AI put the entry on its list. A dim badge means it did not.

  1. 1

    GPT-6.1 Terra

    Half Sol's token price ($1/$5) and, by OpenAI's claim, 3x faster, yet it costs $1.19 per OSWorld task. The $0.40 figure from Artificial Analysis comes from a different benchmark. Strict completion is unpublished.

    Included it: Claude, Meta, Grok, Gemini, Nova, Kimi, DeepSeek, Qwen.

    In their lists: Claude: 1 · Meta: 1 · Grok: 2 · Gemini: 1 · Nova · Kimi: 1 · DeepSeek: 1 · Qwen: 1

    What each AI noted 8
    • Claudeplace 1

      At $1/$5 per million tokens it is the cheapest frontier computer-use model. OpenAI's own max-effort run on OSWorld 2.1 cost $1.19 per task, which was not cheaper than Sol's $1.15.

      Source ↗
    • Metaplace 1

      Half price at $1/$5 and 3x faster, matches flagship Astra at one-tenth token price

      Source ↗
    • Grokplace 2

      Half the token price of prior mid-tier models and claimed ~3× faster; OpenAI puts medium-effort Agents’ Last Exam at $0.29 per task.

      Source ↗
    • Geminiplace 1

      Cuts token prices by 50% while operating 3x faster, yielding a benchmark-leading real cost of just $0.18 per finished task.

    • Nova

      Named differently in its own list: open the full list.

    • Kimiplace 1

      Half the token price and 3x the speed: about $1.19 per OSWorld 2.1 task at max effort — OpenAI says roughly 1/9 the cost per task of Claude Opus 5.5.

      Source ↗
    • DeepSeekplace 1

      At $0.40 a task on Artificial Analysis and 3x the speed, Terra makes retries cheap enough that a misclick is no longer the main cost.

      Source ↗
    • Qwenplace 1

      Half the price of its predecessors and 3x faster, it finally makes frontier computer use affordable per finished task.

      Source ↗
    8 AIs put it here
    Source ↗
  2. 2

    GPT-6.1 Sol

    In OpenAI's own OSWorld 2.1 chart, Sol costs $1.15 a task at 78.0%, against Terra's $1.19 at 74.2%. Arena measures it at $0.81 a task.

    Included it: Claude, Meta, Grok, Kimi, DeepSeek, Qwen. Did not include: Gemini, Nova.

    In their lists: Claude: 2 · Meta: 6 · Grok: 7 · Kimi: 2 · DeepSeek: 2 · Qwen: 6

    What each AI noted 6
    • Claudeplace 2

      OpenAI says Sol gets within 2.1 points of Astra for about one-seventh of the cost per task: 71.4% partial score at $1.27, against Astra's 73.5% at $9.44.

      Source ↗
    • Metaplace 6

      $0.28 cached per task, 70.5% partial on OSWorld 2.0 — middle tier that beats Astra on cost per success

    • Grokplace 7

      Higher capability than Terra but roughly double the token price; self-reported stronger on some long-horizon offline sets.

      Source ↗
    • Kimiplace 2

      $0.81 per task measured on Arena's real-work leaderboard — and in OpenAI's own chart a finished OSWorld task at max costs $1.15, four cents less than Terra.

      Source ↗
    • DeepSeekplace 2

      $1.27 per OSWorld 2.0 task against Astra's $9.44 — a 7.4x gap for a 2.1-point drop in partial score.

    • Qwenplace 6

      Clears 75.2% on DeepSWE, but its premium token rates make it expensive for high-retry visual workflows.

      Source ↗
    6 AIs put it here
    Source ↗
  3. 3

    Claude Opus 5.5

    The only leading model with a published strict OSWorld 2.0 score: 48.7% full completion. It costs $5.98–6.31 a task, about five times Sol, but full completion is checked.

    Included it: Claude, Meta, Kimi, DeepSeek, Qwen. Did not include: Grok, Gemini, Nova.

    In their lists: Claude: 5 · Meta: 8 · Kimi: 6 · DeepSeek: 8 · Qwen: 7

    What each AI noted 5
    • Claudeplace 5

      The best strict score in the table, 48.7% full completion, but at $4/$20 and roughly 9x Terra's cost per task.

      Source ↗
    • Metaplace 8

      81.8% partial, 48.7% binary on OSWorld 2.0 v2.1 — highest verified full-completion rate

      Source ↗
    • Kimiplace 6

      Best strict result money can buy — 48.7% full completion on OSWorld 2.0, a third higher than anything else verified — at roughly $13 per finished task.

      Source ↗
    • DeepSeekplace 8

      48.7% strict completion at $23.21 a task — roughly $47.66 per finished workflow, the most expensive credible option in the top tier.

      Source ↗
    • Qwenplace 7

      Reaches 83.5% on OSWorld-Verified, but $4/$20 pricing means screenshot loops quickly inflate the bill.

    5 AIs put it here
    Source ↗
  4. 4

    Claude Fable 5.1

    Estimate: $7.63 a task at 41.7–42.8% strict completion (the sources differ) is about $18 per finished task, at $10/$50 per million tokens.

    Included it: Claude, Meta, Grok, Nova, DeepSeek. Did not include: Gemini, Kimi, Qwen.

    In their lists: Claude: 6 · Meta: 9 · Grok: 10 · Nova · DeepSeek: 7

    What each AI noted 5
    • Claudeplace 6

      It completes 41.7% of OSWorld 2.0 tasks outright, but at $10/$50 per million tokens it costs 2.5x Opus 5.5 per token.

      Source ↗
    • Metaplace 9

      85.96% OSWorld-Verified — above 72.36% human baseline but flagship $10/$50 pricing

      Source ↗
    • Grokplace 10

      Tops several OSWorld 2.0 partial boards at 77.9% but $10/$50 token prices make finished long tasks costly after retries.

      Source ↗
    • Nova

      Named differently in its own list: open the full list.

    • DeepSeekplace 7

      $7.63 a task and 42.8% strict completion — about $17.83 per finished workflow, not per attempt.

      Source ↗
    5 AIs put it here
    Source ↗
  5. 5

    Claude Sonnet 5.5

    Estimate: $4.51 a task on Arena and 83.9% partial on OSWorld 2.1, the top score in OpenAI's own chart, give about $5.40 per finished task. No strict score is published.

    Included it: Claude, Meta, Kimi, Qwen. Did not include: Grok, Gemini, Nova, DeepSeek.

    In their lists: Claude: 4 · Meta: 5 · Kimi: 3 · Qwen: 4

    What each AI noted 4
    • Claudeplace 4

      The mid-priced Claude at $2/$10 per million tokens. Its predecessor Sonnet 5 scored 81.2% on OSWorld-Verified, within five points of the leaders.

      Source ↗
    • Metaplace 5

      $0.31 cached vs $1.45 uncached per task — 4.7x saving from prompt caching alone

      Source ↗
    • Kimiplace 3

      Highest desktop score in OpenAI's own comparison chart — 83.9% on OSWorld 2.1 offline — at $2/$10 per million tokens, about $5.4 per finished task.

      Source ↗
    • Qwenplace 4

      At $2/$10 per million tokens, it offers Anthropic's best safety-to-cost ratio for mid-tier computer use.

    4 AIs put it here
    Source ↗
  6. 6

    GPT-6 Astra

    $9.44 a task for 73.5% partial: seven times Sol's cost for about 2 more points. One report says OpenAI cancelled it, but trackers still list it.

    Included it: Claude, Kimi, DeepSeek. Did not include: Meta, Grok, Gemini, Nova, Qwen.

    In their lists: Claude: 9 · Kimi: 5 · DeepSeek: 6

    What each AI noted 3
    • Claudeplace 9

      About 10x Terra's token price and $9.44 per task for 73.5% partial score; Sol gets close to that for $1.27.

      Source ↗
    • Kimiplace 5

      $4.42 per measured Arena task and 80.9% offline completion — but at $10/$50 per million tokens it pays frontier prices for mid-tier economics.

      Source ↗
    • DeepSeekplace 6

      90.8% on OSWorld-Verified, yet $23.80 a task at max effort: capability you pay for by the minute.

      Source ↗
    3 AIs put it here
    Source ↗
  7. 7

    Claude Opus 5

    Estimate: $23.70 a task at 44.33% strict completion is about $53 per finished task. On an earlier OSWorld 2.0 release the same model scored 31.43%.

    Included it: Grok, Kimi, DeepSeek. Did not include: Claude, Meta, Gemini, Nova, Qwen.

    In their lists: Grok: 8 · Kimi: 7 · DeepSeek: 9

    What each AI noted 3
    • Grokplace 8

      Strong partial scores on OSWorld 2.0 (~70%+) but binary completion still in the low-to-mid 30-40% range on official boards; higher token cost.

      Source ↗
    • Kimiplace 7

      The cheapest published OSWorld-Verified run — 50.7% at a $2/M blended rate — works out to about $1.10 per finished task, but only on the older, easier benchmark.

      Source ↗
    • DeepSeekplace 9

      44.33% strict completion and $23.70 a task — a 1-in-2 agent that costs about $53.46 per finished job.

    3 AIs put it here
    Source ↗
  8. 8

    Claude Opus 4.8 (batched tools)

    Estimate: about $72.4 a task at 20.6% strict completion in the OSWorld 2.0 paper, or roughly $351 per finished task, the costliest in this table.

    Included it: Claude, Grok, DeepSeek. Did not include: Meta, Gemini, Nova, Kimi, Qwen.

    In their lists: Claude: 10 · Grok: see list · DeepSeek: see list

    What each AI noted 3
    • Claudeplace 10

      The paper's best model at the time cost about $72.4 per task and finished 20.6%, which works out to roughly $350 per fully finished task.

      Source ↗
    • Grok

      Named differently in its own list: open the full list.

      see list
    • DeepSeek

      Named differently in its own list: open the full list.

      see list
    3 AIs put it here
    Source ↗
  9. 9

    Simular Sai

    Estimate: $14.34–15.70 a task at 28.25% strict completion is about $51–56 per finished task, even though its partial scores reach 73–79%.

    Included it: Grok, Qwen. Did not include: Claude, Meta, Gemini, Nova, Kimi, DeepSeek.

    In their lists: Grok: 9 · Qwen: 3

    What each AI noted 2
    • Grokplace 9

      Highest partial scores on some OSWorld 2.x boards (~73-79%) but ~$14–15 per task — expensive once retries and screenshots are counted.

      Source ↗
    • Qwenplace 3

      Beats major labs on OSWorld 2.0 binary completion at 28.25% by minimizing expensive reasoning loops.

      Source ↗
    2 AIs put it here
    Source ↗
  10. 10

    Qwen3.8-27B with Agent S4 (Simular)

    Estimate: $1.10 a task at 21.3% strict completion works out to about $5 per finished task. That is the cheapest strict figure found, but it comes from a single report.

    Included it: Grok. Did not include: Claude, Meta, Gemini, Nova, Kimi, DeepSeek, Qwen.

    In their lists: Grok: 1

    What each AI noted 1
    • Grokplace 1

      Open 27B model fully solves 21.3% of OSWorld V2 tasks at ~$1.10 each — edges Opus 4.8 binary while costing ~1.5% as much.

      Source ↗
    1 AIs put it here
    Source ↗

More about the top three ↓

02More about the top three

The ranking is above. Here is more about the three places at the top: what they are, why they are here, and what is worth knowing.

The three are not measured the same way: Terra and Sol come from OpenAI's own OSWorld 2.1 chart, while Opus's cost and strict score come from OSWorld 2.0 and other sources.

1GPT-6.1 Terra

No. 1 — GPT-6.1 Terra. On the lists of 8 of 8 AIs; median place there: 1. The AIs put Terra first for its half-price tokens and claimed triple speed, though its $1.19 task cost barely differs from Sol's.

Did you know

  • The name "GPT-6.1 Terra" first showed up in a code change, not a launch post. A merged Microsoft pull request on October 16, 2026 switched 18 evaluation settings from GPT-5.6 Terra to gpt-6.1-terra. OpenAI has not confirmed the model.

    Source ↗ ✓ confirmed by several AIs
What it is
An OpenAI computer-use model, priced at $1 and $5 per million tokens.
Its edge
Half Sol's token price and claimed 3x faster, so retries stay cheap.
Who it suits
Suits budget-minded users who accept unconfirmed numbers; not those needing a model that is released today.
  • Cost per OSWorld task$1.19 per OSWorld task
Keep in mind

  • A blog says a new Terra was reportedly still in development and could slip into November. It is an unconfirmed report, so Terra's numbers may change.

    Source ↗
Where to go further

  • Still unreleased. Its name appears in leaked test settings, not in OpenAI's catalog. Read the 'Release status' section first, then check OpenAI's own pricing page before counting its cost.

    Source ↗
Compare the top three ↓

2GPT-6.1 Sol

No. 2 — GPT-6.1 Sol. On the lists of 6 of 8 AIs; median place there: 4. The AIs rank Sol second because it finished tasks for $1.15, four cents under Terra, and came close to the pricier Astra.

Did you know

  • GPT-6.1 Sol came out only a week after GPT-6 Sol, on Sep 29, 2026. It ranks first of 69 models on BrowseComp, a web-browsing test, with 95.4%. The page notes that OpenAI ran many of its headline tests itself.

    Source ↗
What it is
OpenAI's mid-tier model, released September 29, 2026, priced at $2 and $10 per million tokens.
Its edge
Slightly cheaper per finished task than Terra, and about seven times cheaper than Astra.
Who it suits
Suits teams wanting near-top computer use at lower cost; not those who need independent test results.
  • Cost per OSWorld task$1.15 per task

Wikidata

Why it is here

  • On OSWorld 2.0 (offline, max effort), GPT-6.1 Sol scored 71.4% at $1.27 per task. The older GPT-6 Astra Max scored 73.5% but cost $9.44 per task, about seven times more. — Kingy.ai, citing OpenAI's launch report

    Source ↗

  • Sol's list price is $2.00 per million input tokens and $10.00 per million output tokens for prompts up to 272,000 tokens. Above that, it rises to $4.00 and $15.00. — Kingy.ai

    Source ↗ ✓ confirmed by several AIs
History

  • OpenAI dropped a planned GPT-6.1 Astra model because it often ignored the instructions it was given. On September 29, 2026, it shipped GPT-6.1 Sol instead, with near-Astra performance at lower prices.

    Source ↗
Keep in mind

  • OpenAI's own page says Sol nearly matches its top model, Astra, on computer use at one-fifth of the token price. Those are the vendor's claims, not independent tests.

    Source ↗ ✓ confirmed by several AIs
Compare the top three ↓

3Claude Opus 5.5

No. 3 — Claude Opus 5.5. On the lists of 5 of 8 AIs; median place there: 7. The AIs place Opus third: it costs about five times more per task, but it has a published strict full-completion score.

Did you know

  • Anthropic says Opus 5.5 "no longer supports turning thinking off." It also says the model "may suspect when it is being evaluated," so a benchmark score can't guarantee how it acts in real use.

    Source ↗
What it is
Anthropic's model from September 22, 2026, priced at $4 and $20 per million tokens.
Its edge
Only one of the three with a published strict score: 48.7% fully completed.
Who it suits
Suits users who must have tasks fully finished and can pay more; not tight budgets.
  • Cost per OSWorld task$5.98–6.31 a task
Why it is here

  • Claude Opus 5.5 scores 48.7% on the stricter OSWorld 2.0, according to its system card. Anthropic publishes this strict score, and the other two models' strict scores were not found. — BenchLM, citing Anthropic's system card

    Source ↗

  • Opus 5.5 lists at $4.00 per million input tokens and $20.00 per million output tokens. Cached input is $0.20 per million, which lowers the cost when a task reuses the same prompt. — BenchLM

    Source ↗ ✓ confirmed by several AIs
History

  • Anthropic released Claude Opus 5.5 on September 22, 2026. Anthropic says it performs at the level of Claude Fable 5.1 on most tasks and costs 40% less to run than Opus 5.

    Source ↗
Keep in mind

  • The same early user says max effort is still slow and token hungry. The advertised 40% cheaper figure is a claim, and max effort is where benchmark wins come from, so real cost per task can run higher.

    Source ↗

  • A migration guide warns that thinking can no longer be turned off and forced tool use now returns an error. Older code may need changes, and costs may rise.

    Source ↗
What people say

  • finally sounds like old claude again, less like ai — @notjazii

    Source ↗
Where to go further

  • Look at the two OSWorld 2.0 numbers side by side: 81.8% partial and 48.7% strict. The gap shows how many tasks were fully finished, not just partly done.

    Source ↗
Compare the top three ↓

The top three side by side

GPT-6.1 TerraGPT-6.1 SolClaude Opus 5.5
In the AIs' lists8 / 86 / 85 / 8
Cost per OSWorld task$1.19 per OSWorld task$1.15 per task$5.98–6.31 a task
Put first byClaude, Meta, Gemini, Kimi, DeepSeek, Qwen——

Places with the same number of AIs are ordered by their average place in the AIs' own lists.

More ratings

03Worth knowing

What the AIs noticed while building the list.

The same agents score 83–86% on OSWorld-Verified, above the 72.36% human baseline, but complete only 20–49% of tasks fully on OSWorld 2.0, where tasks average about 318 tool calls against about 30.

Noticed by:
Included it: Claude, Meta, Grok, Kimi, Qwen. Did not include: Gemini, Nova, DeepSeek.
Source ↗

A lower token price did not make Terra cheaper per task: on OpenAI's OSWorld run it cost $1.19 against $1.15 for Sol, because it used more steps and tokens.

Noticed by:
Included it: Claude, Kimi. Did not include: Meta, Grok, Gemini, Nova, DeepSeek, Qwen.
Source ↗

Claude Opus 5 rose from 31.43% to 44.33% strict completion between OSWorld 2.0 releases with no change to the model, after 14 tasks were rewritten to stop reward hacking.

Noticed by:
Included it: Claude, Kimi. Did not include: Meta, Grok, Gemini, Nova, DeepSeek, Qwen.
Source ↗

The screenshot tax: one independent run needed 400–750k input tokens and 43–68 screenshot-reason-click cycles per task, about 45x the cost of using the same app's API.

Noticed by:
Included it: Kimi. Did not include: Claude, Meta, Grok, Gemini, Nova, DeepSeek, Qwen.
Source ↗

04Questions people ask

Which AI can use your computer?

GPT-6.1 Terra and Sol, Claude Opus, Sonnet and Fable 5.x, Qwen3.8 and Gemini Flash can all see the screen and click. The best strict full-completion rate on OSWorld 2.0 is 48.7% (Claude Opus 5.5).

Source ↗
What is the best computer use AI agent?

On cost per finished task, GPT-6.1 Sol ($1.15 per OSWorld task at 78.0%) narrowly beats Terra ($1.19 at 74.2%) in OpenAI's own chart. On strict completion, Claude Opus 5.5 leads at 48.7%.

Source ↗
Is GPT-6.1 Terra really half the price?

Its token price, $1/$5 per million, is half of Sol's. On OpenAI's OSWorld run it still cost $1.19 per task, slightly more than Sol.

Source ↗
Why do computer-use scores differ so much between sources?

Vendors quote partial scores while strict completion is far lower: Opus 5 scores 77.67% partial but 44.33% strict. Boards also use different releases, step limits and harnesses.

Source ↗

05How we counted

In plain words, no formulas.

1The question

Top 10 AI Computer-Use Agents Ranked by Real Cost per Finished Task, Not Price per Token

2Why this topic

No published table gives the reader the cost of one finished task. This ranking combines the benchmark score, the token price and the number of steps an agent actually uses, and the order changes once failed and repeated attempts are counted.

3Where it applies

Global: AI agents that see the screen and operate desktop and web apps, including GPT-6.1 Terra, released October 26, 2026

4What we looked at

  • Success rate on OSWorld-Verified (official leaderboard)
  • Full-completion rate on the stricter OSWorld 2.0
  • List price per million input and output tokens, from each vendor's pricing page
  • Steps and screenshot tokens used per task, so each retry counts toward the cost
  • Speed each vendor claims, checked against independent runs where they exist
  • OSWorld 2.0 / v2.1 (XLang Lab, compiled by Presenc AI) — Partial score and strict binary full completion on 108 long-horizon workflows (about 318 tool calls each); release versions are not comparable
  • OSWorld 2.0 / Snorkel leaderboard — Strict full completion and partial score by model, effort level and release
  • OpenAI GPT-6.1 Terra launch chart (via Kingy AI) — OSWorld 2.1 offline-set score and cost per task at each effort level for Terra, Sol, Astra and Claude

5Who ranked

The leading AIs, each building its own list of ten independently and naming a source for every place: Claude, Meta, Grok, Gemini, Nova, Kimi, DeepSeek, Qwen.

6How the places were set

The same entry written differently by different AIs was counted as one. The more AIs put an entry on their list, and the higher they put it, the higher its place. "AI agreement" shows how many of the AIs had it on their own list. Sources are only those the AIs found themselves; nothing was added by hand.

06Each AI's own list

🇺🇸 Claude10 places · 4 notes

The same question, answered by Claude alone. The final list above combines all the AIs.

  1. 1
    GPT-6.1 Terra (OpenAI, Oct 2026)

    At $1/$5 per million tokens it is the cheapest frontier computer-use model. OpenAI's own max-effort run on OSWorld 2.1 cost $1.19 per task, which was not cheaper than Sol's $1.15.

    Source ↗
  2. 2
    GPT-6.1 Sol (OpenAI, Sept 2026)

    OpenAI says Sol gets within 2.1 points of Astra for about one-seventh of the cost per task: 71.4% partial score at $1.27, against Astra's 73.5% at $9.44.

    Source ↗
  3. 3
    Qwen3.8-Max (Alibaba, Aug 2026)

    It tops the OSWorld-Verified table at 86.1% while charging $2/$6, which is cheaper per output token than any Claude model with a comparable score. The score is self-reported.

    Source ↗
  4. 4
    Claude Sonnet 5.5 (Anthropic)

    The mid-priced Claude at $2/$10 per million tokens. Its predecessor Sonnet 5 scored 81.2% on OSWorld-Verified, within five points of the leaders.

    Source ↗
  5. 5
    Claude Opus 5.5 (Anthropic, Sept 2026)

    The best strict score in the table, 48.7% full completion, but at $4/$20 and roughly 9x Terra's cost per task.

    Source ↗
  6. 6
    Claude Fable 5.1 (Anthropic)

    It completes 41.7% of OSWorld 2.0 tasks outright, but at $10/$50 per million tokens it costs 2.5x Opus 5.5 per token.

    Source ↗
  7. 7
    Muse Spark 1.3 (Meta)

    It is second on the strict OSWorld 2.0 table at 39.25% full completion, behind only Claude Opus 5. I could not find a list price.

    Source ↗
  8. 8
    GPT-6 Luna (OpenAI)

    At $0.10/$0.50 per million tokens, its Agents' Last Exam run cost $0.14 per task. It has no published OSWorld result.

    Source ↗
  9. 9
    GPT-6 Astra (OpenAI)

    About 10x Terra's token price and $9.44 per task for 73.5% partial score; Sol gets close to that for $1.27.

    Source ↗
  10. 10
    Claude Opus 4.8, batched tools (Anthropic, mid-2026)

    The paper's best model at the time cost about $72.4 per task and finished 20.6%, which works out to roughly $350 per fully finished task.

    Source ↗

What Claude noticed

The same agent scores 83-86% on OSWorld-Verified but 20-49% on OSWorld 2.0 full completion. The easy benchmark flatters it; the long-horizon one is closer to a real work day. OSWorld 2.0 tasks average about 318 tool calls against about 30 for Verified.

Source ↗

Lower token prices did not make Terra cheaper per task on OpenAI's OSWorld run: $1.19 for Terra against $1.15 for Sol. Cost per task depends on how many steps and screenshots the model burns, not only on the price list.

Source ↗

On the Snorkel OSWorld 2.0 table, Claude Opus 5 goes from 18.81% to 44.33% full completion when effort is raised from low to max, so the same model can sit at very different points on cost versus success.

Source ↗

A 98% per-step success rate over 20 steps gives only about 67% end-to-end success. This is why long tasks cost far more per finished task than the per-token price suggests.

Source ↗
🇺🇸 Meta10 places · 3 notes

The same question, answered by Meta alone. The final list above combines all the AIs.

  1. 1
    GPT-6.1 Terra

    Half price at $1/$5 and 3x faster, matches flagship Astra at one-tenth token price

    Source ↗
  2. 2
    Gemini 3.8 Flash

    $0.58 per Intelligence Index task vs $3.26 for Astra — cheapest finished task in independent test

  3. 3
    Claude Haiku 5.5

    $0.07 per 20-step task uncached — cheapest Anthropic but 5x price after 100K tokens per prompt

  4. 4
    GPT-6 Luna

    $0.10/$0.50 — 10x cheaper than Terra, built for volume form-filling despite lower reasoning

  5. 5
    Claude Sonnet 5.5

    $0.31 cached vs $1.45 uncached per task — 4.7x saving from prompt caching alone

    Source ↗
  6. 6
    GPT-6.1 Sol

    $0.28 cached per task, 70.5% partial on OSWorld 2.0 — middle tier that beats Astra on cost per success

  7. 7
    Gemini 3.5 Flash

    78.4% OSWorld-Verified vs 78.7% for GPT-5.5 — near parity at $1.50/$9 with built-in computer use

  8. 8
    Claude Opus 5.5

    81.8% partial, 48.7% binary on OSWorld 2.0 v2.1 — highest verified full-completion rate

    Source ↗
  9. 9
    Claude Fable 5

    85.96% OSWorld-Verified — above 72.36% human baseline but flagship $10/$50 pricing

    Source ↗
  10. 10
    Qwen3.8 Max

    86.1% OSWorld-Verified leader Oct 26 2026 — open-weight alternative to $10/$50 flagships

    Source ↗

What Meta noticed

Token price lies: a $4/$20 model at $0.61 cached per task can be cheaper per finished task than a $1/$5 model that loops through misclicks and needs 3 retries — cost per success = cost per run / success rate

Source ↗

OSWorld-Verified is now saturated above human 72.36%, so OSWorld 2.0 binary 44-48% is the real ceiling — vendors quote 77-81% partial scores that hide full failures

Source ↗

Prompt caching cuts a 20-step task from $1.45 to $0.31, 4.7x saving — bigger lever than halving list price from $2 to $1

Source ↗
🇺🇸 Grok10 places · 4 notes

The same question, answered by Grok alone. The final list above combines all the AIs.

  1. 1
    Qwen3.8-27B with Agent S4 (Simular, 2026)

    Open 27B model fully solves 21.3% of OSWorld V2 tasks at ~$1.10 each — edges Opus 4.8 binary while costing ~1.5% as much.

    Source ↗
  2. 2
    GPT-6.1 Terra (OpenAI, 2026)

    Half the token price of prior mid-tier models and claimed ~3× faster; OpenAI puts medium-effort Agents’ Last Exam at $0.29 per task.

    Source ↗
  3. 3
    MiMo-V2.6-Pro (Xiaomi, 2026)

    82% on OSWorld-Verified at roughly $0.43/$0.87 per million tokens — one of the cheapest high-scoring open options.

    Source ↗
  4. 4
    Qwen3.8 Max (Alibaba, 2026)

    Leads OSWorld-Verified at 86.1% while staying well under Claude flagship token prices.

    Source ↗
  5. 5
    Gemini 3.6 Flash (Google, 2026)

    83% Verified at $0.75/$3.75 — strong score-to-price on short desktop tasks.

    Source ↗
  6. 6
    Claude Sonnet 5 (Anthropic, 2026)

    81.2% Verified at mid-tier pricing; still trails open models on pure cost per finished short task.

    Source ↗
  7. 7
    GPT-6.1 Sol (OpenAI, 2026)

    Higher capability than Terra but roughly double the token price; self-reported stronger on some long-horizon offline sets.

    Source ↗
  8. 8
    Claude Opus 5 (Anthropic, 2026)

    Strong partial scores on OSWorld 2.0 (~70%+) but binary completion still in the low-to-mid 30-40% range on official boards; higher token cost.

    Source ↗
  9. 9
    Simular Sai (hosted agent, 2026)

    Highest partial scores on some OSWorld 2.x boards (~73-79%) but ~$14–15 per task — expensive once retries and screenshots are counted.

    Source ↗
  10. 10
    Claude Fable 5.1 (Anthropic, 2026)

    Tops several OSWorld 2.0 partial boards at 77.9% but $10/$50 token prices make finished long tasks costly after retries.

    Source ↗

What Grok noticed

Open 27B models with evolved agent harnesses now match or beat closed frontier binary completion on OSWorld V2 at roughly 1–5% of the cost.

Source ↗

OSWorld-Verified scores cluster above the 72% human baseline while OSWorld 2.0 binary completion for the same models often stays under 45% — long-horizon state tracking remains the bottleneck.

Source ↗

GPT-6.1 Terra’s list price is half of GPT-6.1 Sol and roughly one-tenth of GPT-6 Astra; OpenAI claims matching Astra-level results on some professional agent tasks at a fraction of the cost.

Source ↗

Screenshot-heavy planning loops mean token price alone understates cost; models that reduce steps or cache better can be cheaper per finished task even at higher list rates.

Source ↗
🇺🇸 Gemini10 places

The same question, answered by Gemini alone. The final list above combines all the AIs.

  1. 1
    GPT-6.1 Terra (OpenAI)

    Cuts token prices by 50% while operating 3x faster, yielding a benchmark-leading real cost of just $0.18 per finished task.

  2. 2
    UI-TARS-7B (ByteDance)

    Lightweight open-weights architecture achieves $0.26 per finished task when routed via low-cost inference hosts.

    Source ↗
  3. 3
    Gemini 2.0 Flash Computer Use (Google)

    Ultra-low $0.10 input pricing offsets occasional screenshot misclicks to deliver $0.32 per completed task.

    Source ↗
  4. 4
    UI-TARS-72B (ByteDance)

    High visual grounding reduces retries by 24%, bringing heavy-model execution down to $0.62 per finished task.

    Source ↗
  5. 5
    Qwen-2.5-VL-72B-Instruct (Alibaba Cloud)

    Superior coordinate prediction saves frames on complex menus, stabilizing completed runs at $0.74 each.

    Source ↗
  6. 6
    Claude 3.5 Sonnet Computer Use (Anthropic)

    High raw accuracy of 64% is tempered by $3/$15 token rates, keeping real finished cost at $1.18.

    Source ↗
  7. 7
    Claude 3.7 Sonnet Computer Use (Anthropic)

    Extended thinking drives OSWorld 2.0 completion to 43.6% but inflates total billable tokens to $1.36 per completed task.

    Source ↗
  8. 8
    Gemini 2.0 Pro Computer Use (Google)

    Context caching trims prompt overhead by 60%, holding per-finished-task costs to $1.52 despite heavy reasoning steps.

    Source ↗
  9. 9
    OpenAI Operator / GPT-4o (OpenAI)

    Legacy vision architecture averages 18 retries per multi-app task, pushing final completion cost to $1.94.

    Source ↗
  10. 10
    Claude 3 Opus Agent Framework (Anthropic)

    Premium $15 input and $75 output rates produce an expensive $9.20 per completed task on long-horizon OS runs.

    Source ↗
🇨🇳 Kimi10 places · 4 notes

The same question, answered by Kimi alone. The final list above combines all the AIs.

  1. 1
    GPT-6.1 Terra (OpenAI, 2026)

    Half the token price and 3x the speed: about $1.19 per OSWorld 2.1 task at max effort — OpenAI says roughly 1/9 the cost per task of Claude Opus 5.5.

    Source ↗
  2. 2
    GPT-6.1 Sol (OpenAI, 2026)

    $0.81 per task measured on Arena's real-work leaderboard — and in OpenAI's own chart a finished OSWorld task at max costs $1.15, four cents less than Terra.

    Source ↗
  3. 3
    Claude Sonnet 5.5 (Anthropic, 2026)

    Highest desktop score in OpenAI's own comparison chart — 83.9% on OSWorld 2.1 offline — at $2/$10 per million tokens, about $5.4 per finished task.

    Source ↗
  4. 4
    Kimi K3.1 Pro (Moonshot AI, 2026)

    79.6% on OSWorld 2.1 with open weights — the only top-five entry you can self-host, which removes the per-token rent entirely.

  5. 5
    GPT-6 Astra (OpenAI, 2026)

    $4.42 per measured Arena task and 80.9% offline completion — but at $10/$50 per million tokens it pays frontier prices for mid-tier economics.

    Source ↗
  6. 6
    Claude Opus 5.5 (Anthropic, 2026)

    Best strict result money can buy — 48.7% full completion on OSWorld 2.0, a third higher than anything else verified — at roughly $13 per finished task.

    Source ↗
  7. 7
    Claude Haiku 4.5 (Anthropic, 2025)

    The cheapest published OSWorld-Verified run — 50.7% at a $2/M blended rate — works out to about $1.10 per finished task, but only on the older, easier benchmark.

    Source ↗
  8. 8
    Gemini 4 Argon / Gemini 3.1 Pro (Google, 2026)

    69.2% on OSWorld 2.0's offline subset — but Google needed 500 steps and up to 3 runs per task, so screenshot tokens quietly eat the savings.

  9. 9
    Qwen3.8 Max (Alibaba, 2026)

    86.1% OSWorld-Verified — the top self-reported score of any model — but with no public API price, cost per finished task is a number nobody can check.

    Source ↗
  10. 10
    GPT-5.6 Luna (OpenAI, 2026)

    The cheapest meter in the industry — $0.0144 for a 12-call task — but computer-use success does not survive the strict benchmarks, so finished tasks cost infinite retries.

    Source ↗

What Kimi noticed

The screenshot tax dominates: an independent Reflex run found a vision agent needs 400-750k input tokens and 43-68 screenshot-reason-click cycles per task, making computer use ~45x more expensive than hitting the same app's structured API — the step count is set by the interface, not the model.

Source ↗

Partial and full completion are two different products: Claude Opus 5 scores 77.67% partial but only 44.33% binary on the same OSWorld 2.0 run — vendors lead with the bigger number, but cost per FINISHED task must divide by the smaller one.

Source ↗

Cheaper per token can still be pricier per task: in OpenAI's own launch chart, Terra at max effort costs $1.19 per OSWorld task versus $1.15 for Sol — Terra burns more steps and tokens despite half the sticker price.

Source ↗

GPT-6.1 Terra's headline economics rest on vendor numbers: the 3x speed and 1/9-cost-vs-Opus claims are OpenAI's own, and as of today no independent lab has run Terra on OSWorld — the Arena leaderboard does not list it yet.

Source ↗
🇨🇳 DeepSeek10 places · 1 notes

The same question, answered by DeepSeek alone. The final list above combines all the AIs.

  1. 1
    GPT-6.1 Terra (2026)

    At $0.40 a task on Artificial Analysis and 3x the speed, Terra makes retries cheap enough that a misclick is no longer the main cost.

    Source ↗
  2. 2
    GPT-6.1 Sol (2026)

    $1.27 per OSWorld 2.0 task against Astra's $9.44 — a 7.4x gap for a 2.1-point drop in partial score.

  3. 3
    Holo3-35B-A3B (2026)

    $0.25 per million input tokens and 82.6% on OSWorld-Verified — the open-weight agent that undercuts the frontier by roughly 10x.

    Source ↗
  4. 4
    Gemini 3.6 Flash (2026)

    $1.50 per million input tokens and 83.0% on OSWorld-Verified; the cheapest way to buy a top-10 desktop score from a major lab.

  5. 5
    Qwen3.8 Max (2026)

    86.1% on OSWorld-Verified, but it burns 19.4x Astra's reasoning tokens — a cheap token that fills the boot.

    Source ↗
  6. 6
    GPT-6 Astra (2026)

    90.8% on OSWorld-Verified, yet $23.80 a task at max effort: capability you pay for by the minute.

    Source ↗
  7. 7
    Claude Fable 5.1 (2026)

    $7.63 a task and 42.8% strict completion — about $17.83 per finished workflow, not per attempt.

    Source ↗
  8. 8
    Claude Opus 5.5 (2026)

    48.7% strict completion at $23.21 a task — roughly $47.66 per finished workflow, the most expensive credible option in the top tier.

    Source ↗
  9. 9
    Claude Opus 5 (2026)

    44.33% strict completion and $23.70 a task — a 1-in-2 agent that costs about $53.46 per finished job.

  10. 10
    Intelligence-Indeed Agent (2026)

    90.19% on the official OSWorld-Verified board, 100 steps, no public per-task price — the leaderboard leader you cannot cost.

What DeepSeek noticed

Terra's 3x speed claim may matter more than its 50% price cut: at $0.40 a task, retries are cheap, but its strict OSWorld 2.0 completion rate is still unpublished.

Source ↗
🇨🇳 Qwen10 places · 3 notes

The same question, answered by Qwen alone. The final list above combines all the AIs.

  1. 1
    GPT-6.1 Terra

    Half the price of its predecessors and 3x faster, it finally makes frontier computer use affordable per finished task.

    Source ↗
  2. 2
    Qwen3.8 Max

    Leads OSWorld-Verified at 86.1%, proving that open-weight models can dominate real desktop tasks at $2/$6 per million tokens.

    Source ↗
  3. 3
    Sai (Simular)

    Beats major labs on OSWorld 2.0 binary completion at 28.25% by minimizing expensive reasoning loops.

    Source ↗
  4. 4
    Claude Sonnet 5.5

    At $2/$10 per million tokens, it offers Anthropic's best safety-to-cost ratio for mid-tier computer use.

  5. 5
    Google Gemini 2.5 Computer Use

    Powers the successor to Project Mariner with a 69.2% OSWorld 2.0 partial score and bundled enterprise pricing.

    Source ↗
  6. 6
    GPT-6.1 Sol

    Clears 75.2% on DeepSWE, but its premium token rates make it expensive for high-retry visual workflows.

    Source ↗
  7. 7
    Claude Opus 5.5

    Reaches 83.5% on OSWorld-Verified, but $4/$20 pricing means screenshot loops quickly inflate the bill.

  8. 8
    Coasty

    Claims 82% on OSWorld with a proprietary routing model that minimizes screenshot tokens per step.

    Source ↗
  9. 9
    OpenAI Operator (CUA)

    The original commercial web agent, now facing stiff competition from the newly released Terra model.

    Source ↗
  10. 10
    Agent S3 w/ bBoN

    Leads some open-source leaderboards at 63.5% but requires heavy compute overhead that adds to real-world costs.

    Source ↗

What Qwen noticed

Binary completion rates on OSWorld 2.0 are shockingly low—often below 30%—proving that partial success scores hide how often agents actually fail to finish long workflows.

Source ↗

GPT-6.1 Terra's release shifts the frontier from raw intelligence to efficiency; paying $1/$5 per million tokens means retrying a 10-step task 5 times is finally cheaper than hiring a human intern.

Source ↗

Open-weight models like Qwen3.8 Max now top verified leaderboards, forcing commercial labs to compete on cost-per-task rather than just benchmark bragging rights.

Source ↗