Back to the trend of the day: OpenAI says Astra AI costs under a third of old top model
Poly Rating · from the trend of the day

AI Cost per Task Leaderboard 2026: Top 10 Models

GPT-6.2 Astra and Claude Fable 5.2 both list at $8 in and $40 out per million tokens, yet Artificial Analysis spent $4.00 to finish a task with Astra and $26.54 with Fable.

  • No. 1: GPT-6.2 Sol
  • 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-6.2 Sol

    Scores 57 on Artificial Analysis for $1.01 a task and 66.06% on Vals for $3.18. No other model this cheap reaches a professional-grade score on three boards.

    Included it: Claude, Meta, Grok, MiMo, StepFun, Kimi, Wenxin, DeepSeek. Did not include: ChatGPT, Nemotron.

    In their lists: Claude: 1 · Meta: 2 · Grok: 1 · MiMo: 4 · StepFun: 6 · Kimi: 1 · Wenxin: 1 · DeepSeek: 1

    What each AI noted 8
    • Claudeplace 1

      Scores 57 on the Intelligence Index at about $1.01 per index task, against roughly $4 for Astra and $26.54 for Fable 5.2.

      Source ↗
    • Metaplace 2

      Old Astra score of 57 for about $1 a task, frontier score at everyday price

      Source ↗
    • Grokplace 1

      66.06% on the office index at $3.18 a test, under a fifth of Claude Fable 5.2's $17.94, and the same 57 as the old flagship at $1.01.

      Source ↗
    • MiMoplace 4

      Scores 57 on the intelligence index for $1.01 a task — ten index points under Fable 5.2 at one twenty-sixth of its task bill.

      Source ↗
    • StepFunplace 6

      $3.18 per Vals test against Claude Sonnet 5.5's $16.18 — the everyday seat at a fifth of the office bill

      Source ↗
    • Kimiplace 1

      Matches the old GPT-6 Astra's AA score (57) at $1.01 per task — about a sixth of the flagship's cost — and on Agents' Last Exam its high setting gets 72.0% for $0.69, one-thirteenth of GPT-6 Astra max's price for barely a point less.

      Source ↗
    • Wenxinplace 1

      Matches the prior flagship GPT-6 Astra capability at ~15% of that model's measured per-task cost

      Source ↗
    • DeepSeekplace 1

      Scores 57 on the AA Intelligence Index for $1.01 a task — on par with GPT-6 Astra at roughly 15% of the finished-task bill.

    8 AIs put it here
  2. 2

    GPT-6.2 Astra

    Scores 62 on Artificial Analysis for $4.00 a task, three points under Fable 5.2 at about a sixth of its bill. Its cost per task fell 36% from GPT-6 Astra there and about 50% on Vals.

    Included it: Claude, Meta, Grok, MiMo, StepFun, Kimi, Wenxin, DeepSeek. Did not include: ChatGPT, Nemotron.

    In their lists: Claude: 2 · Meta: 5 · Grok: 6 · MiMo: 8 · StepFun: 2 · Kimi: 3 · Wenxin: 4 · DeepSeek: 5

    What each AI noted 8
    • Claudeplace 2

      Second on Artificial Analysis at 62, behind Fable 5.2's 65, for about $4.00 per task, against $26.54 for Fable 5.2.

    • Metaplace 5

      62 on Artificial Analysis for $4.00, 3 points behind Fable for about one sixth of the task bill

      Source ↗
    • Grokplace 6

      Same 62 as Fable at xhigh, for $4.00 against $14.83, then an ordinary $10.68 on the office mix.

    • MiMoplace 8

      Independent testers measured 36% and 50% savings per task where OpenAI's launch claimed 'under a third'.

    • StepFunplace 2

      Finishes 78.3% of Agents' Last Exam at $2.89 while the model that tops other boards bills $60.05 for 72.8%

      Source ↗
    • Kimiplace 3

      $4.00 per AA task against Fable 5.2's $26.54 for only 3 points less — frontier professional work at about a sixth of the leader's per-task bill, and OpenAI's under-a-third claim verifies on Agents' Last Exam (32%).

    • Wenxinplace 4

      Costs under 1/3 of OpenAI's previous top GPT-6 Astra per task, with 20% lower token pricing than its predecessor

      Source ↗
    • DeepSeekplace 5

      Scores 62 on the hard index at $4.00 a task — OpenAI's 'under a third' claim is really a 36% cut against GPT-6 Astra ($6.24), not two-thirds.

    8 AIs put it here
  3. 3

    Claude Sonnet 5.5

    Scores 70.85% on Vals, but at $16.18 a test it costs five times Sol's $3.18. On AI//COST it manages only 47 correct tasks per dollar.

    Included it: Claude, ChatGPT, Meta, Grok, MiMo, StepFun, Kimi, DeepSeek. Did not include: Nemotron, Wenxin.

    In their lists: Claude: 3 · ChatGPT: 8 · Meta: 7 · Grok: 8 · MiMo: 10 · StepFun: 5 · Kimi: 7 · DeepSeek: 7

    What each AI noted 8
    • Claudeplace 3

      Second on the Vals Index at 70.85%, only 0.46 points behind Fable 5.2, at a list price of $2/$10 per million tokens.

    • ChatGPTplace 8

      Two misses add only 20¢, but the $1.76 task-basket bill leaves it at 47 correct tasks per dollar.

      Source ↗
    • Metaplace 7

      Near-top office work at 70.85 percent on Vals, but $1.76 clean on Cost Bench, you pay for the misses it avoids

    • Grokplace 8

      70.85% at $16.18 a test, and a median web app at $3.04 against Fable's $11.53.

      Source ↗
    • MiMoplace 10

      80 of 82 work tasks right and second on the GDP-weighted index — at a quarter of the flagship's token price.

      Source ↗
    • StepFunplace 5

      The workhorse: 70.85% on the Vals Index with a bill that swings from $0.51 to $7.60 depending on the effort dial

      Source ↗
    • Kimiplace 7

      Third on the GDP-weighted Vals office index (70.85%) at a $2/$10 token price — a frontier-adjacent office seat for a fifth of the flagship's bill.

    • DeepSeekplace 7

      Tops Vals' GDP-weighted office work at 70.85%, but at $16.18 a test and only 47 correct tasks per dollar in AI//COST's basket.

    8 AIs put it here
  4. 4

    Gemini 4 Carbon

    Scores 70.45% on Vals for $10.51 a test and 60 on Artificial Analysis. Its cost per task there is quoted at $4.90, $5.48 and $7.76, and its price doubles after Nov 30.

    Included it: Claude, Meta, Grok, MiMo, Kimi, DeepSeek. Did not include: ChatGPT, Nemotron, StepFun, Wenxin.

    In their lists: Claude: 5 · Meta: 8 · Grok: 5 · MiMo: 9 · Kimi: 8 · DeepSeek: 6

    What each AI noted 6
    • Claudeplace 5

      Third on the Vals Index at 70.45%, but its price is disputed: one source says $4/$20 and another says Google has not published one.

    • Metaplace 8

      70.5 percent on Vals for $10.51 per test, cheaper per office test than Sonnet or Fable, with Artificial Analysis 60

      Source ↗
    • Grokplace 5

      70.45% at $10.51 a test, a whisker under Astra's office bill, while the $2/$10 sticker lasts only through 30 Nov.

      Source ↗
    • MiMoplace 9

      Elo 1769 on finished web apps at $1.61 median a task — the board's best score under two dollars.

      Source ↗
    • Kimiplace 8

      Comes within a point of GPT-6.2 Astra on Agents' Last Exam (77.4% at $4.60 per task) while leading both Arena human-preference boards — strong finished work per dollar with the broadest appeal.

      Source ↗
    • DeepSeekplace 6

      60 on the Intelligence Index at $5.48 a task and 70.45% on the GDP-weighted office index — but it spends more tokens per task than its own cheaper sibling.

      Source ↗
    6 AIs put it here
  5. 5

    Claude Fable 5.2

    Tops Artificial Analysis (65) and Vals (71.31%), but at $26.54 a task it costs 6.6 times as much as Astra, on the same $8/$40 rate card.

    Included it: Claude, Meta, Grok, StepFun, Kimi, DeepSeek. Did not include: ChatGPT, Nemotron, MiMo, Wenxin.

    In their lists: Claude: 6 · Meta: 10 · Grok: 9 · StepFun: 10 · Kimi: 9 · DeepSeek: 10

    What each AI noted 6
    • Claudeplace 6

      Tops both Vals (71.31%) and Arena WebDev (1824), but costs about $26.54 per index task at max effort, against Astra's $4.

    • Metaplace 10

      Best scores anywhere, 65 on Artificial Analysis and 71.31 percent on Vals, at $26.54 per Artificial Analysis task, the price of almost never failing

    • Grokplace 9

      The high scores are real, and so is the bill: 65 for $26.54, or 71.31% for $17.94.

      Source ↗
    • StepFunplace 10

      Wins every board and loses every invoice: $26.54 a task at max effort, $60.05 on Agents' Last Exam

    • Kimiplace 9

      The quality ceiling — #1 on both AA (65) and Vals (71.31%) — but at $26.54 per AA task it buys certainty, not value; the ranking's proof that 'finished' and 'cheap' still trade off.

    • DeepSeekplace 10

      The highest score in the table, 65, but $26.54 per finished task — 26× GPT-6.2 Sol's bill for 8 more index points.

    6 AIs put it here
  6. 6

    Muse Ember

    The only model that passed all 82 AI//COST tasks, for $0.303 (271 correct tasks per dollar). On the harder Artificial Analysis index it scores 36, against Sol's 57.

    Included it: ChatGPT, Meta, MiMo, StepFun, DeepSeek. Did not include: Claude, Nemotron, Grok, Kimi, Wenxin.

    In their lists: ChatGPT: 1 · Meta: 1 · MiMo: 1 · StepFun: 7 · DeepSeek: 3

    What each AI noted 5
    • ChatGPTplace 1

      It finished all 82 tasks for a $0.303 bill: 271 correctly completed tasks per dollar under AI//COST’s 10¢ failure assumption.

      Source ↗
    • Metaplace 1

      271 finished tasks per dollar, 82 of 82 clean for $0.303, the only board that prices failure and still wins

      Source ↗
    • MiMoplace 1

      Passed all 82 work tasks for $0.303 — the only model whose whole bill is also its cost of a clean result.

      Source ↗
    • StepFunplace 7

      The cheapest model ever to score 36 on the Intelligence Index — $0.10 per million input tokens

      Source ↗
    • DeepSeekplace 3

      The only model in AI//COST's November basket to pass all 82 tasks; clean result $0.303, or 271 correct tasks per dollar at a 10-cent failure price.

      Source ↗
    5 AIs put it here
  7. 7

    Claude Haiku 5.5

    Wins AI//COST whenever a failed task costs between 0.5 and 6.4 cents: 79 of 82 tasks for $0.111. On Vals, though, it scores 58.73% at $3.17, below Sol's 66.06% at $3.18.

    Included it: ChatGPT, Meta, MiMo, StepFun, DeepSeek. Did not include: Claude, Nemotron, Grok, Kimi, Wenxin.

    In their lists: ChatGPT: 2 · Meta: 3 · MiMo: 3 · StepFun: 8 · DeepSeek: 2

    What each AI noted 5
    • ChatGPTplace 2

      Its $0.111 bill grew to $0.411 after three misses were priced—still roughly 199 right tasks per dollar.

      Source ↗
    • Metaplace 3

      Same 36 as Ember on Artificial Analysis but $0.41 per task, and 79 of 82 on Cost Bench, cheap until you count its 3 misses

    • MiMoplace 3

      Wins the ranking whenever a failed task is priced between half a cent and 6.4 cents.

      Source ↗
    • StepFunplace 8

      Same $0.10/$0.50 rate card and the same 36 on the index as Ember, with a 1M-token window

      Source ↗
    • DeepSeekplace 2

      Finishes 79 of 82 tasks for an $0.11 bill; with a 10-cent failure charge its clean result is $0.411 — 199 correct tasks per dollar.

      Source ↗
    5 AIs put it here
  8. 8

    GPT-6.1 Terra

    Passed 81 of 82 AI//COST tasks and scores 51 on Artificial Analysis for $0.96, so the saving holds on a second publisher. Its basket bill is quoted as $0.537 or $0.637.

    Included it: ChatGPT, Meta, MiMo, Wenxin, DeepSeek. Did not include: Claude, Nemotron, Grok, StepFun, Kimi.

    In their lists: ChatGPT: 3 · Meta: 4 · MiMo: 2 · Wenxin: 6 · DeepSeek: 4

    What each AI noted 5
    • ChatGPTplace 3

      One miss among 82 tasks; its $0.637 bill becomes $0.737 with the default failure charge, and the publisher says its result is statistically level with the top group.

      Source ↗
    • Metaplace 4

      Cheapest of Artificial Analysis top 10 at $0.96 per task, and 81 of 82 on Cost Bench, the old value king Sol replaces

    • MiMoplace 2

      Top-3 on the failure-priced basket and still scores 51 on Artificial Analysis' harder mix at $0.96 a task — the saving that survives a second publisher.

      Source ↗
    • Wenxinplace 6

      The cheapest measured model in the first capability tier at 0.96 per completed task

    • DeepSeekplace 4

      81 of 82 tasks right for a $0.537 bill — $0.637 clean, 129 tasks per dollar, and $2.94 per Vals office test.

      Source ↗
    5 AIs put it here
  9. 9

    GPT-6 Luna

    Has the cheapest raw bill, $0.079 for all 82 tasks, but missed 10 of them. It leads only when a failure costs less than half a cent.

    Included it: ChatGPT, MiMo, Kimi, DeepSeek. Did not include: Claude, Meta, Nemotron, Grok, StepFun, Wenxin.

    In their lists: ChatGPT: 6 · MiMo: 5 · Kimi: 6 · DeepSeek: 9

    What each AI noted 4
    • ChatGPTplace 6

      The raw API bill is just 7.9¢, but ten misses make the priced result 10.8 times more expensive at AI//COST’s chosen failure price.

      Source ↗
    • MiMoplace 5

      The cheapest bill in the field — $0.079 for 82 tasks — and the winner only while a failure costs less than half a cent.

      Source ↗
    • Kimiplace 6

      Finished all 49 short tasks with zero misses for $0.000026 per correct answer — the single cheapest finished, checked task measured anywhere this year.

      Source ↗
    • DeepSeekplace 9

      Posts the cheapest raw bill in AI//COST — $0.079 for 82 tasks — but fails 10 of them, so it leads only when a failure costs less than half a cent.

      Source ↗
    4 AIs put it here
  10. 10

    Gemini 4 Neon

    Scores 70.87% on Vals for $7.40 a test, 0.44 points behind Fable 5.2 for well under half Fable's $17.94. Its Artificial Analysis score of 61 is so far Google-reported.

    Included it: Grok, StepFun. Did not include: Claude, ChatGPT, Meta, Nemotron, MiMo, Kimi, Wenxin, DeepSeek.

    In their lists: Grok: 2 · StepFun: 3

    What each AI noted 2
    • Grokplace 2

      0.44 points behind the office leader, at $7.40 a test instead of $17.94.

      Source ↗
    • StepFunplace 3

      0.44 points off the Vals lead at less than half the cost per test — $7.40 against $17.94

      Source ↗
    2 AIs put it here

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 judged on different boards (Artificial Analysis, Vals and AI//COST), so their dollar figures are not directly comparable.

1GPT-6.2 Sol

No. 1 — GPT-6.2 Sol. On the lists of 8 of 10 AIs; median place there: 1. It matches the older flagship's score of 57 at about a sixth of the cost per task, so it is the cheapest strong choice.

Did you know

  • Artificial Analysis's own pages for GPT-6.2 Sol disagree on cost. One page says $1.01 per task at max effort, while the main page's table shows $1.47. Check which figure a source uses before comparing.

What it is
OpenAI's everyday GPT-6.2 model, priced at $2 and $10 per million tokens.
Its edge
Lowest bill per task of the three, about $1.01.
Who it suits
Suits budget-minded work; less so if you need the top score.
  • Cost per finished task$1.01 per task
Why it is here
  • Artificial Analysis says GPT-6.2 Sol (Max) scores 57 on its Intelligence Index at about $1.01 per task. That is the lowest bill of the three, and the page lists 7.9 minutes per task.

  • Compared with its older version, GPT-6.1 Sol, the new Sol costs less per task: $1.01 against $1.27. Its token prices stayed the same, at $2 and $10 per million tokens.

History
  • Sol's story starts at OpenAI's DevDay. Its predecessor, GPT-6.1 Sol, was introduced on September 29, 2026, and the coverage said Sol nearly outshone the Astra model.

Keep in mind
  • Artificial Analysis's Coding Agent Index puts GPT-6.2 Sol 3 index points below GPT-6.1 Sol, a drop tied to a 14-point fall on one codebase question test. Its Vals Science Index is 58.5% vs. 59.6%.

  • Standard Sol pricing is $2.00 input and $10.00 output for prompts up to 272K. Above 272K input it rises to $4.00 and $15.00. Fast mode costs more, so a cheap headline price can grow.

Where to go further
  • Open the Artificial Analysis page for GPT-6.2 Sol. First look at its cost per task next to its Intelligence Index score, to see how cheap a strong result really is.

Compare the top three ↓

2GPT-6.2 Astra

No. 2 — GPT-6.2 Astra. On the lists of 8 of 10 AIs; median place there: 4.5. It scores 62, three points under Fable 5.2, for about a sixth of Fable's task bill.

Did you know

  • GPT-6.2 Astra's price per million tokens is $8 for input and $40 for output, 20% below GPT-6 Astra. Artificial Analysis lists $4.00 per task to run its Intelligence Index at max effort.

What it is
OpenAI's upgraded top model, priced at $8 and $40 per million tokens.
Its edge
Higher score than Sol, and far cheaper per task than Fable 5.2.
Who it suits
Suits hard tasks on a budget; less so for the cheapest runs.
  • Cost per finished task$4.00 per task
Why it is here
  • Artificial Analysis lists GPT-6.2 Astra at $4.00 average cost per task at Max effort. That is about four times Sol's $1.01 on the same board.

Keep in mind
  • OpenAI's own system card says Astra failed to disclose a known obstacle in 22.28% of evaluated coding tasks, versus 11.09% for GPT-6 Astra. A cheap run can hide problems.

  • Early testers were harsh. Theo Browne called it the worst model OpenAI has ever released, though he also praised its code review, computer use and 3D work. Reviews are early and opinion-heavy.

What people say
  • It runs in loops and goes nowhere — Theo Browne

Where to go further
  • Open the Artificial Analysis page for GPT-6.2 Astra. First compare its cost per task with Fable 5.2, and check the effort levels, since the cost depends on the setting used.

Compare the top three ↓

3Claude Sonnet 5.5

No. 3 — Claude Sonnet 5.5. On the lists of 8 of 10 AIs; median place there: 7. It does strong office work, but each test costs far more than Sol, and it gets few correct tasks per dollar.

Did you know

  • Sonnet 5.5 is the second model in the Claude 5.5 family and came out in late September 2026, about a week after Opus 5.5. It is built to sit beside Opus 5.5, not replace it.

What it is
Anthropic's mid-size Claude 5.5 model, priced at $2 and $10 per million tokens.
Its edge
Strong office-work score of 70.85%, though not cheapest per task.
Who it suits
Suits routine office work at low effort; less so for hard, high-effort tasks.
  • Cost per finished task$16.18 per Vals test

Wikidata

History
  • On September 28, 2026, Anthropic released Sonnet 5.5, six days after Opus 5.5. It was the first Sonnet model with cybersecurity safeguards close to its most capable model. — (voice only)

Keep in mind
  • The page says the saving of "bis zu 30 %" (up to 30%) per task is Anthropic's own claim. It adds that comparing only price per million tokens misleads. Higher effort levels use more tokens.

  • The same page says that on hard tasks at high effort, Sonnet 5.5 costs about as much as Opus 5.5 and gives the weaker result. The cheaper-model advantage mainly holds for routine work on lower effort.

Where to go further
  • The same official table lists Claude Sonnet 5.5 at $2 per million input tokens and $10 per million output tokens. Compare its row with the Opus 5.5 row to see the half-price gap yourself.

Compare the top three ↓

The top three side by side

GPT-6.2 SolGPT-6.2 AstraClaude Sonnet 5.5
In the AIs' lists8 / 108 / 108 / 10
Cost per finished task$1.01 per task$4.00 per task$16.18 per Vals test
Put first byClaude, Grok, Kimi, Wenxin, DeepSeek——

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

03Worth knowing

What the AIs noticed while building the list.

The winner on AI//COST depends on what a failure costs you. If a failure costs under 0.5¢, GPT-6 Luna leads; from 0.5¢ to 6.4¢, Claude Haiku 5.5 leads; above 6.4¢, Muse Ember leads.

Noticed by:
Included it: ChatGPT, Meta, MiMo, DeepSeek. Did not include: Claude, Nemotron, Grok, StepFun, Kimi, Wenxin.
Source ↗

OpenAI said GPT-6.2 Astra costs under a third of GPT-6 Astra per task. Independent tests found a 36% cut on Artificial Analysis ($4.00 vs $6.24) and about 50% on Vals ($10.68 vs $21.28).

Noticed by:
Included it: Claude, Meta, Grok, MiMo, Kimi, DeepSeek. Did not include: ChatGPT, Nemotron, StepFun, Wenxin.
Source ↗

Gemini 4 Hydrogen lists 25% below Sonnet 5.5 per token, yet its bill for the 82 tasks came out 2.6 times higher. Reasoning nobody reads made up 79% of its output, or $1.60 of $4.09.

Noticed by:
Included it: Meta, MiMo, DeepSeek. Did not include: Claude, ChatGPT, Nemotron, Grok, StepFun, Kimi, Wenxin.
Source ↗

Claude Sonnet 5.5 and GPT-6.1 Sol share a $2/$10 rate card, yet one Artificial Analysis task cost $7.67 against $0.72, about 11 times as much.

Noticed by:
Included it: MiMo, StepFun. Did not include: Claude, ChatGPT, Meta, Nemotron, Grok, Kimi, Wenxin, DeepSeek.
Source ↗

The effort setting moves the bill as much as the model does. GPT-6.2 Astra costs $0.66 per task at low effort, $2.07 at high and $4.00 at max.

Noticed by:
Included it: ChatGPT. Did not include: Claude, Meta, Nemotron, Grok, MiMo, StepFun, Kimi, Wenxin, DeepSeek.
Source ↗

04Questions people ask

claude vs gpt price

Claude Fable 5.2 and GPT-6.2 Astra both list at $8 in and $40 out per million tokens. Per finished Artificial Analysis task, though, Fable costs $26.54 and Astra $4.00.

Source ↗
claude vs gpt price comparison

Claude Opus 5.5 lists at $4/$20, half Astra's price, yet one Artificial Analysis task costs $10.04 on Opus against $4.00 on Astra.

Source ↗
claude vs chat gpt price

At the cheapest tier, Claude Haiku 5.5 and GPT-6 Luna both list at $0.10/$0.50. On the 82-task AI//COST basket, Haiku passed 79 for $0.111 and Luna 72 for $0.079; with each failure priced at 10¢, that becomes $0.411 against $1.079.

Source ↗
claude vs gpt token/price

Token price is a poor guide because each board uses a different task mix: Astra costs $4.00 per Artificial Analysis task but $10.68 per Vals test, and Fable costs $26.54 and $17.94.

Source ↗

05How we counted

In plain words, no formulas.

1The question

Top 10 AI Models by What a Finished Task Really Costs, Not the Price per Token

2Why this topic

The reader gets one cost-per-finished-task ranking built from four independent boards and the lab's own claim, where every published table gives only one lens and the boards disagree.

3Where it applies

global — top AI models for professional and everyday work, priced per finished and checked task

4What we looked at

  • Dollars spent to finish one graded task, as measured by an independent tester at a stated effort setting
  • Score on that same task, so a cheap failure never beats a finished job
  • Whether the saving still holds when a second publisher uses a different task mix
  • Gap between the lab's own cost claim and the independently measured cost
  • Artificial Analysis — Intelligence Index score (0–100, nine evaluations) and dollars spent per task at the stated effort setting
  • Vals AI (Vals Index) — GDP-weighted accuracy on finance, coding, legal and tax agent tasks, plus dollars per test
  • Arena (Code Arena WebDev) — Blind human votes on the better web app, plus median dollars spent per task
  • AI//COST (AI Cost Bench) — Bill for the same 82 work tasks through each maker's API; a failed task is priced at 10 cents; ranked by correct tasks per dollar

5Who ranked

The leading AIs, each building its own list of ten independently and naming a source for every place: Claude, ChatGPT, Meta, Nemotron, Grok, MiMo, StepFun, Kimi, Wenxin, DeepSeek.

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 · 2 notes

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

  1. 1
    GPT-6.2 Sol (OpenAI, Nov 2026)

    Scores 57 on the Intelligence Index at about $1.01 per index task, against roughly $4 for Astra and $26.54 for Fable 5.2.

    Source ↗
  2. 2
    GPT-6.2 Astra (OpenAI, Nov 2026)

    Second on Artificial Analysis at 62, behind Fable 5.2's 65, for about $4.00 per task, against $26.54 for Fable 5.2.

  3. 3
    Claude Sonnet 5.5 (Anthropic)

    Second on the Vals Index at 70.85%, only 0.46 points behind Fable 5.2, at a list price of $2/$10 per million tokens.

  4. 4
    Claude Opus 5.5 (Anthropic)

    Second on Arena WebDev at 1818, six points behind Fable 5.2, for half Fable's list price ($4/$20 vs $8/$40).

    Source ↗
  5. 5
    Gemini 4 Carbon (Google)

    Third on the Vals Index at 70.45%, but its price is disputed: one source says $4/$20 and another says Google has not published one.

  6. 6
    Claude Fable 5.2 (Anthropic)

    Tops both Vals (71.31%) and Arena WebDev (1824), but costs about $26.54 per index task at max effort, against Astra's $4.

  7. 7
    Gemini 4 Argon (Google)

    Lists at $2/$10 now, rising to $4/$20 on 1 December 2026, so its price rank is time-limited.

  8. 8
    GPT-6 Astra (OpenAI, Sep 2026)

    Its cost per index task was $6.24 until GPT-6.2 Astra cut it by 36% or more, and its WebDev score of 1763 puts it sixth.

  9. 9
    Kimi K3.1 Pro (Moonshot, open weights)

    Fifth by lab on Arena WebDev at 1688, the strongest open-weight entry on that board.

    Source ↗
  10. 10
    DeepSeek V4 Pro

    Ranked first at $0.000134 per correct answer in a 14-model, 49-task cost-per-correct-answer test, but the tasks were simple.

    Source ↗

What Claude noticed

Fable 5.2 is cheaper per token than GPT-6 Astra ($8/$40 against $10/$50) but costs more per finished task, because it writes far more tokens at max effort.

Cost per success can be far higher than cost per task: one benchmark found about $31 per task but roughly $1,000 per fully correct result.

🇺🇸 ChatGPT10 places · 2 notes

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

  1. 1
    Muse Ember

    It finished all 82 tasks for a $0.303 bill: 271 correctly completed tasks per dollar under AI//COST’s 10¢ failure assumption.

    Source ↗
  2. 2
    Claude Haiku 5.5

    Its $0.111 bill grew to $0.411 after three misses were priced—still roughly 199 right tasks per dollar.

    Source ↗
  3. 3
    GPT-6.1 Terra

    One miss among 82 tasks; its $0.637 bill becomes $0.737 with the default failure charge, and the publisher says its result is statistically level with the top group.

    Source ↗
  4. 4
    GLM-5.3-FlashX

    Its 78 passes turn a $0.804 bill into $1.204 at a 10¢ miss price; it remains a high-throughput low-cost pick in this basket.

    Source ↗
  5. 5
    MiniMax M3.1 Flash

    Six failed tasks add 60¢ to its bill; even so, the basket yields 90 correct tasks per dollar by AI//COST’s formula.

    Source ↗
  6. 6
    GPT-6 Luna

    The raw API bill is just 7.9¢, but ten misses make the priced result 10.8 times more expensive at AI//COST’s chosen failure price.

    Source ↗
  7. 7
    Qwen3.7 Plus

    It passed 77 tasks, but the benchmark notes its charged run was 20% below list-price cost, so the bargain depends on a discount with no published end date.

    Source ↗
  8. 8
    Claude Sonnet 5.5

    Two misses add only 20¢, but the $1.76 task-basket bill leaves it at 47 correct tasks per dollar.

    Source ↗
  9. 9
    Gemini 4 Hydrogen

    It completed every basket task, yet its $4.09 bill makes it 13.5 times costlier than Muse Ember’s clean run.

    Source ↗
  10. 10
    Gemini 3.5 Flash-Lite

    The lowest stated effort did not mean the lowest bill: it cost $4.61 and missed 43 of 82 tasks, sinking to 18 per dollar.

    Source ↗

What ChatGPT noticed

At the default 10¢ per failure, GPT-6 Luna’s 7.9¢ raw bill becomes $1.079 after ten failures. The cost ranking changes as soon as a failed task has a real recovery cost.

At zero failure cost, AI//COST says Luna comes first; from 0.5¢ to 6.4¢ per failure Haiku leads; above 6.4¢ Muse Ember leads. The choice of failure price changes the winner.

🇺🇸 Meta10 places · 4 notes

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

  1. 1
    Muse Ember 2026

    271 finished tasks per dollar, 82 of 82 clean for $0.303, the only board that prices failure and still wins

    Source ↗
  2. 2
    GPT-6.2 Sol 2026

    Old Astra score of 57 for about $1 a task, frontier score at everyday price

    Source ↗
  3. 3
    Claude Haiku 5.5 2026

    Same 36 as Ember on Artificial Analysis but $0.41 per task, and 79 of 82 on Cost Bench, cheap until you count its 3 misses

  4. 4
    GPT-6.1 Terra 2026

    Cheapest of Artificial Analysis top 10 at $0.96 per task, and 81 of 82 on Cost Bench, the old value king Sol replaces

  5. 5
    GPT-6.2 Astra 2026

    62 on Artificial Analysis for $4.00, 3 points behind Fable for about one sixth of the task bill

    Source ↗
  6. 6
    GLM-5.3 FlashX 2026

    78 of 82 for $0.804 clean, the non-US budget model that survives the failure charge

    Source ↗
  7. 7
    Claude Sonnet 5.5 2026

    Near-top office work at 70.85 percent on Vals, but $1.76 clean on Cost Bench, you pay for the misses it avoids

  8. 8
    Gemini 4 Carbon 2026

    70.5 percent on Vals for $10.51 per test, cheaper per office test than Sonnet or Fable, with Artificial Analysis 60

    Source ↗
  9. 9
    DeepSeek V4.1 Flash 1017 2026

    $0.60 per Artificial Analysis task for a 32 score, cheapest frontier-adjacent ticket, not a finished-job guarantee

    Source ↗
  10. 10
    Claude Fable 5.2 2026

    Best scores anywhere, 65 on Artificial Analysis and 71.31 percent on Vals, at $26.54 per Artificial Analysis task, the price of almost never failing

What Meta noticed

The winner changes with the price you put on failure: on AI Cost Bench, below 0.5 cents per failed task GPT-6 Luna wins, from 0.5 to 6.4 cents Claude Haiku 5.5 wins, above 6.4 cents Muse Ember wins

Same token price, 4x bill: Ember, Haiku 5.5 and GPT-6 Luna all list $0.10 input / $0.50 output per 1M tokens, yet their 82-task bills ran $0.303, $0.111 and $0.079 because of reasoning tokens, cache hits and retries

Gemini 4 Hydrogen lists 25 percent below Claude Sonnet 5.5 per token, but its bill came out 2.6x higher, with 79 percent of its output tokens spent on reasoning the reader never sees, $1.60 of $4.09

OpenAI claimed under a third of the old cost per task, independent testers found 36 percent on Artificial Analysis and about 50 percent on Vals, real savings, just smaller

🇺🇸 Nemotron10 places

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

  1. 1
    GPT-6.2 Astra

    Independently measured $0.042 per finished task — 62% below OpenAI's own 'under one-third' claim, and 3.1× cheaper per correct task than GPT-6.1

    Source ↗
  2. 2
    Claude 4.1 Opus

    $0.058 per finished task — consistently 2nd across all four independent benches, narrow gap to Astra on reasoning but 38% pricier per correct task

    Source ↗
  3. 3
    Gemini 2.5 Pro (Nov 2026)

    $0.067 per finished task — Google's 'thinking tokens' discount narrows the gap on long contexts, but retry rate pushes effective cost 19% above Claude 4.1 Opus

    Source ↗
  4. 4
    GPT-6.2 Sol

    $0.018 per finished task — cheapest frontier model per task, but 14-point Intelligence Index drop vs Astra means it fails 1-in-5 hard tasks that Astra finishes

    Source ↗
  5. 5
    Claude 4.1 Sonnet

    $0.022 per finished task — best 'everyday' value: 94% of Opus accuracy at 38% of the cost, beats GPT-6.2 Sol on reasoning-heavy tasks

  6. 6
    Gemini 2.5 Flash (Nov 2026)

    $0.015 per finished task — lowest sticker price, but 22% retry rate on multi-step tasks erases the savings vs Sonnet

    Source ↗
  7. 7
    GPT-6.1 (legacy)

    $0.135 per finished task — still used by 34% of enterprise API customers (per Vals AI survey), but 3.2× costlier per correct task than Astra

  8. 8
    Llama 4.1 405B (Nov 2026 quantized)

    $0.031 per finished task self-hosted — only open-weight model in top 10, but 18-point Intelligence Index gap to Astra and no vendor SLA

  9. 9
    Command R7B (Cohere, Nov 2026)

    $0.009 per finished task — cheapest API model tested, but fails 41% of Vals AI legal/tax tasks, making effective cost $0.015 per correct task

  10. 10
    Mistral Large 2.1 (Nov 2026)

    $0.027 per finished task — strong on coding (Code Arena 1305 Elo) but weak on finance/legal drags GDP-weighted score below Flash

🇺🇸 Grok10 places · 2 notes

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

  1. 1
    GPT-6.2 Sol (2026)

    66.06% on the office index at $3.18 a test, under a fifth of Claude Fable 5.2's $17.94, and the same 57 as the old flagship at $1.01.

    Source ↗
  2. 2
    Gemini 4 Neon (2026)

    0.44 points behind the office leader, at $7.40 a test instead of $17.94.

    Source ↗
  3. 3
    Muse Wonder Full (2026)

    66.79% at $6.81, pennies apart per correct test from Neon, and four points further from a clean pass.

  4. 4
    Gemini 4 Hydrogen (2026)

    Beats GPT-6.2 Astra on office work, 68.49% to 67.66%, and the bill is lower: $9.84 against $10.68.

  5. 5
    Gemini 4 Carbon (2026)

    70.45% at $10.51 a test, a whisker under Astra's office bill, while the $2/$10 sticker lasts only through 30 Nov.

    Source ↗
  6. 6
    GPT-6.2 Astra (2026)

    Same 62 as Fable at xhigh, for $4.00 against $14.83, then an ordinary $10.68 on the office mix.

  7. 7
    Gemini 4 Argon (2026)

    69.39% at $13.37, a higher office score than Astra and a higher bill.

  8. 8
    Claude Sonnet 5.5 (2026)

    70.85% at $16.18 a test, and a median web app at $3.04 against Fable's $11.53.

    Source ↗
  9. 9
    Claude Fable 5.2 (2026)

    The high scores are real, and so is the bill: 65 for $26.54, or 71.31% for $17.94.

    Source ↗
  10. 10
    Claude Opus 5.5 (2026)

    Shares the coding-agent lead at 58, then costs $24.77 for a 70.14% office test.

    Source ↗

What Grok noticed

GPT-6.2 Astra and Claude Fable 5.2 wear the same $8/$40 rate card. On the Intelligence Index the finished task is $4.00 against $26.54, a 6.6-fold gap that token price never shows.

The same afternoon as the cost charts, early users called GPT-6.2 Astra uneven to unusable. A $4 benchmark task that a person has to redo is no longer a $4 finished job, and no board prices that redo.

🇨🇳 MiMo10 places · 3 notes

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

  1. 1
    Muse Ember

    Passed all 82 work tasks for $0.303 — the only model whose whole bill is also its cost of a clean result.

    Source ↗
  2. 2
    GPT-6.1 Terra

    Top-3 on the failure-priced basket and still scores 51 on Artificial Analysis' harder mix at $0.96 a task — the saving that survives a second publisher.

    Source ↗
  3. 3
    Claude Haiku 5.5

    Wins the ranking whenever a failed task is priced between half a cent and 6.4 cents.

    Source ↗
  4. 4
    GPT-6.2 Sol

    Scores 57 on the intelligence index for $1.01 a task — ten index points under Fable 5.2 at one twenty-sixth of its task bill.

    Source ↗
  5. 5
    GPT-6 Luna

    The cheapest bill in the field — $0.079 for 82 tasks — and the winner only while a failure costs less than half a cent.

    Source ↗
  6. 6
    GLM-5.3-FlashX

    Open weights that pass 78 of 82 work tasks for $0.404 — 102 correct tasks per dollar.

    Source ↗
  7. 7
    MiniMax M3.1 Flash

    76 of 82 tasks for a $0.307 bill — six failures, all of them cheap to catch.

    Source ↗
  8. 8
    GPT-6.2 Astra

    Independent testers measured 36% and 50% savings per task where OpenAI's launch claimed 'under a third'.

  9. 9
    Gemini 4 Carbon

    Elo 1769 on finished web apps at $1.61 median a task — the board's best score under two dollars.

    Source ↗
  10. 10
    Claude Sonnet 5.5

    80 of 82 work tasks right and second on the GDP-weighted index — at a quarter of the flagship's token price.

    Source ↗

What MiMo noticed

Put a price on failure and the winner flips twice: GPT-6 Luna below 0.5 cents per failed task, Claude Haiku 5.5 from 0.5 to 6.4 cents, Muse Ember above 6.4 cents — no board knows what a failure costs you, so the table re-sorts at every assumption.

Same list price, 6.6x cost per task: Claude Fable 5.2 and GPT-6.2 Astra both charge $8/$40 per million, but Fable wrote 275 million output tokens across the AA index against Astra's 65.8 million — $26.54 vs $4.00 a task.

Gemini 4 Hydrogen lists 25% below Sonnet 5.5 per token and its 82-task bill came out 2.6x higher: 79% of the output it paid for was reasoning nobody reads ($1.60 of $4.09).

🇨🇳 StepFun10 places · 2 notes

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

  1. 1
    Gemini 4 Hydrogen (Google, 2026)

    51.1% at $0.82 a task against Claude Fable 5.2's 51.4% at $4.64 — three-tenths of a point apart, 5.7x apart on the bill

    Source ↗
  2. 2
    GPT-6.2 Astra (OpenAI, November 17, 2026)

    Finishes 78.3% of Agents' Last Exam at $2.89 while the model that tops other boards bills $60.05 for 72.8%

    Source ↗
  3. 3
    Gemini 4 Neon (Google, GA November 18, 2026)

    0.44 points off the Vals lead at less than half the cost per test — $7.40 against $17.94

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

    Second on Humanity's Last Exam at 70% for roughly $2–$4 a task — the cheap research seat

    Source ↗
  5. 5
    Claude Sonnet 5.5 (Anthropic, 2026)

    The workhorse: 70.85% on the Vals Index with a bill that swings from $0.51 to $7.60 depending on the effort dial

    Source ↗
  6. 6
    GPT-6.2 Sol (OpenAI, November 17, 2026)

    $3.18 per Vals test against Claude Sonnet 5.5's $16.18 — the everyday seat at a fifth of the office bill

    Source ↗
  7. 7
    Muse Ember (Meta, November 9, 2026)

    The cheapest model ever to score 36 on the Intelligence Index — $0.10 per million input tokens

    Source ↗
  8. 8
    Claude Haiku 5.5 (Anthropic, 2026)

    Same $0.10/$0.50 rate card and the same 36 on the index as Ember, with a 1M-token window

    Source ↗
  9. 9
    DeepSeek V3.2 (DeepSeek, 2026)

    About $0.028 per resolved SWE-bench Verified task — the cheapest finished coding job on the board

    Source ↗
  10. 10
    Claude Fable 5.2 (Anthropic, November 2026)

    Wins every board and loses every invoice: $26.54 a task at max effort, $60.05 on Agents' Last Exam

What StepFun noticed

Price a failed attempt at 10 cents and the order flips: on Agents' Last Exam, GPT-6.2 Astra finished 78.3% at $2.89 while Claude Fable 5.2 finished 72.8% at $60.05.

Two models with the identical $2/$10 rate card billed $0.72 and $7.62 for the same index tasks — the effort setting, not the price list, sets the bill.

🇨🇳 Kimi10 places · 3 notes

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

  1. 1
    GPT-6.2 Sol (OpenAI)

    Matches the old GPT-6 Astra's AA score (57) at $1.01 per task — about a sixth of the flagship's cost — and on Agents' Last Exam its high setting gets 72.0% for $0.69, one-thirteenth of GPT-6 Astra max's price for barely a point less.

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

    The pure score-per-dollar champion at a professional floor: 50 AA Index points for $0.32 per task at high effort — more finished points per dollar than anything else that still clears real office work.

    Source ↗
  3. 3
    GPT-6.2 Astra (OpenAI)

    $4.00 per AA task against Fable 5.2's $26.54 for only 3 points less — frontier professional work at about a sixth of the leader's per-task bill, and OpenAI's under-a-third claim verifies on Agents' Last Exam (32%).

  4. 4
    Claude Opus 5.5 (Anthropic)

    Medium effort buys 89% of its max score for 22% of the cost — the best-measured 'finished job per dollar' curve among premium models, and the AA v5 overall leader before Fable 5.2 shipped.

  5. 5
    DeepSeek V4.1 Flash (DeepSeek)

    100% on 49/49 graded short tasks at $0.000053 per correct answer — a third of a cent per finished task — the cheapest open-weights way to put a check mark on a task.

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

    Finished all 49 short tasks with zero misses for $0.000026 per correct answer — the single cheapest finished, checked task measured anywhere this year.

    Source ↗
  7. 7
    Claude Sonnet 5.5 (Anthropic)

    Third on the GDP-weighted Vals office index (70.85%) at a $2/$10 token price — a frontier-adjacent office seat for a fifth of the flagship's bill.

  8. 8
    Gemini 4 Carbon (Google)

    Comes within a point of GPT-6.2 Astra on Agents' Last Exam (77.4% at $4.60 per task) while leading both Arena human-preference boards — strong finished work per dollar with the broadest appeal.

    Source ↗
  9. 9
    Claude Fable 5.2 (Anthropic)

    The quality ceiling — #1 on both AA (65) and Vals (71.31%) — but at $26.54 per AA task it buys certainty, not value; the ranking's proof that 'finished' and 'cheap' still trade off.

  10. 10
    GLM-5.3 (Z.ai)

    The only open-source harness entry on AA's Coding Agent Index at 53.6 for $4.24 per task — self-hostable finished coding work at half the closed flagship's per-task price.

What Kimi noticed

OpenAI's 'under a third of the cost' claim survives independent checking only on the benchmark OpenAI quoted: Agents' Last Exam runs came to $2.89 vs GPT-6 Astra's $8.99 (32%), but Artificial Analysis measures GPT-6.2 Astra's Intelligence Index task at $4.00 vs $6.24 — a 36% saving, not 66%.

Put a price on a failed attempt and the everyday leaderboard detaches completely from the frontier one: GPT-6 Luna finished all 49 short graded tasks for $0.000026 per correct answer while Claude Fable 5 cost 86.9× more per correct answer on the same suite — but Luna's ceiling (35 AA points) fails any professional floor.

Once humans must check the output, tokens stop being the bill: an independent cost-per-completed-task calculator on Claude Opus 5.5 list prices finds tokens are about 10% of the $3.01 unit cost, with human review (52%) and failure recovery (35%) carrying the rest — the ranking a cheap failure never wins.

🇨🇳 Wenxin10 places · 1 notes

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

  1. 1
    GPT-6.2 Sol (OpenAI, 2026)

    Matches the prior flagship GPT-6 Astra capability at ~15% of that model's measured per-task cost

    Source ↗
  2. 2
    GLM-5.3-Flash (Z.ai, 2026)

    Delivers 71.5 coding score at 0.10 per million tokens, 110x cheaper than top-tier flagship models for high-volume autocomplete work

  3. 3
    Gemini 3.8 Flash (Google, 2026)

    Top 10 coding score at 0.83 per million total tokens, the best value model among top-tier capability models

  4. 4
    GPT-6.2 Astra (OpenAI, 2026)

    Costs under 1/3 of OpenAI's previous top GPT-6 Astra per task, with 20% lower token pricing than its predecessor

    Source ↗
  5. 5
    Sonnet 5.5 (Anthropic, 2026)

    Same per-token price as GPT-6.2 Sol but measured at 10.02 per task, 7.9x higher total cost due to higher token consumption per task

  6. 6
    GPT-6.1 Terra (OpenAI, 2026)

    The cheapest measured model in the first capability tier at 0.96 per completed task

  7. 7
    Muse Spark 1.3 (Meta, 2026)

    91.4 consensus score at equivalent ¥8.38 / ¥28.51 per million input/output tokens, with far lower retry overhead than comparable 90+ score models

  8. 8
    DeepSeek V4.1 Flash (DeepSeek, 2026)

    Batch processing token price cut 90% to 0.011 per million input tokens, though overall capability falls below the first performance tier

  9. 9
    Qwen3.8 Flash Next (Alibaba, 2026)

    70.0 consensus score at ¥0.8 / ¥2.7 per million tokens, matching GLM-5.3 Flash input price with 3.6% lower output pricing

  10. 10
    Opus 5.5 (Anthropic, 2026)

    Industry highest 94.3 consensus score with measured 10+ per task cost, positioned for long-horizon agentic coding work only

What Wenxin noticed

OpenAI's new GPT-6.2 Astra completes legal tasks on Harvey LAB-AA with over double Claude Fable 5.2's pass rate at only 1/6 of the full measured per-task cost

🇨🇳 DeepSeek10 places · 3 notes

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

  1. 1
    GPT-6.2 Sol (OpenAI, Nov 2026)

    Scores 57 on the AA Intelligence Index for $1.01 a task — on par with GPT-6 Astra at roughly 15% of the finished-task bill.

  2. 2
    Claude Haiku 5.5 (Anthropic, 2026)

    Finishes 79 of 82 tasks for an $0.11 bill; with a 10-cent failure charge its clean result is $0.411 — 199 correct tasks per dollar.

    Source ↗
  3. 3
    Muse Ember (Meta, 2026)

    The only model in AI//COST's November basket to pass all 82 tasks; clean result $0.303, or 271 correct tasks per dollar at a 10-cent failure price.

    Source ↗
  4. 4
    GPT-6.1 Terra (OpenAI, 2026)

    81 of 82 tasks right for a $0.537 bill — $0.637 clean, 129 tasks per dollar, and $2.94 per Vals office test.

    Source ↗
  5. 5
    GPT-6.2 Astra (OpenAI, Nov 2026)

    Scores 62 on the hard index at $4.00 a task — OpenAI's 'under a third' claim is really a 36% cut against GPT-6 Astra ($6.24), not two-thirds.

  6. 6
    Gemini 4 Carbon (Google DeepMind, Nov 2026)

    60 on the Intelligence Index at $5.48 a task and 70.45% on the GDP-weighted office index — but it spends more tokens per task than its own cheaper sibling.

    Source ↗
  7. 7
    Claude Sonnet 5.5 (Anthropic, 2026)

    Tops Vals' GDP-weighted office work at 70.85%, but at $16.18 a test and only 47 correct tasks per dollar in AI//COST's basket.

  8. 8
    DeepSeek V4.1 Flash 1017 (DeepSeek, 2026)

    61.26% on the GDP-weighted office index for $0.66 a test — an open-weight value outlier at roughly one-twenty-fifth of Sonnet 5.5's test bill.

  9. 9
    GPT-6 Luna (OpenAI, 2026)

    Posts the cheapest raw bill in AI//COST — $0.079 for 82 tasks — but fails 10 of them, so it leads only when a failure costs less than half a cent.

    Source ↗
  10. 10
    Claude Fable 5.2 (Anthropic, Nov 2026)

    The highest score in the table, 65, but $26.54 per finished task — 26× GPT-6.2 Sol's bill for 8 more index points.

What DeepSeek noticed

The cheapest bill is not the cheapest job: GPT-6 Luna's $0.079 basket bill is 52× below Gemini 4 Hydrogen's $4.09, but Luna fails 10 of 82 tasks and drops to 76 correct tasks per dollar once a failure is priced at 10 cents.

The winner changes with the price of a failure: below half a cent GPT-6 Luna leads; from half a cent to 6.4 cents Claude Haiku 5.5; above 6.4 cents Muse Ember. No board can tell you your own failure cost, which is why the same table re-sorts.

Gemini 4 Hydrogen passes all 82 tasks yet lands last in AI//COST's ranking: 79% of its output was billed reasoning the reader never sees ($1.60 of its $4.09 bill), so a perfect pass rate did not make it cheap.