Back to the trend of the day: Meta's AI helper that does tasks for you now lives on your watch
Poly Rating · from the trend of the day

Best AI Agent Benchmark 2026: Top 10 Agents per Dollar

On Princeton's independent GAIA re-runs, Claude Opus 4.1 High spent $562 to score 68.48%, while Claude Sonnet 4.5 scored 74.55% for $178.

  • No. 1: Claude Sonnet 4.5 (HAL Generalist Agent)
  • 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

    Claude Sonnet 4.5 (HAL Generalist Agent)

    Highest verified GAIA score (74.55%) and highest Level 3 score (65.39%) for $178.20 a run. HAL marks it Pareto-optimal. Opus 4.1 High costs 3x more and scores lower.

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

    In their lists: Claude: 1 · ChatGPT: 1 · Kimi: 1

    What each AI noted 3
    • Claudeplace 1

      The only run to clear 74% overall and 65% on Level 3, at about $1.45 per solved question. HAL marks it Pareto-optimal.

    • ChatGPTplace 1

      It solved nearly three quarters of the validation set for $178.20 across the run, while leading this list on level 3 at 65.39%.

      Source ↗
    • Kimiplace 1

      Best verified accuracy on the board (74.55%) AND the best hardest-level score (65.4%) — for about a dollar per task.

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

    Claude 3.7 Sonnet High (HAL Generalist Agent)

    A February 2025 model that scores 64.24% overall and 57.69% on Level 3 for $122.49. It beats the newer Opus 4.1 High on Level 3 at about a fifth of the cost.

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

    In their lists: Claude: 4 · ChatGPT: 2 · Kimi: 2

    What each AI noted 3
    • Claudeplace 4

      A February 2025 model still scores 57.69% on Level 3, ahead of GPT-5 Medium's 46%, at about $1.16 per solved question.

      Source ↗
    • ChatGPTplace 2

      It trails the overall leader by 10.31 points, but costs about 31% less per estimated correct answer and gets 57.69% on level 3.

      Source ↗
    • Kimiplace 2

      A February 2025 model that beats the newer Opus 4.1 High on the hardest level (57.7% vs 53.8%) for a fifth of the cost.

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

    OpenAI o4-mini Low (HAL Generalist Agent)

    Pareto-optimal at $73.26 a run. It ties the $562 Opus 4.1 High run on Level 3 (53.85%) but is 16 points behind the leader overall.

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

    In their lists: Claude: 2 · ChatGPT: 4 · Kimi: 3

    What each AI noted 3
    • Claudeplace 2

      It reaches 58% overall and 53.85% on the hardest level, at about $0.76 per solved question, which is the best mix of price and Level 3 strength in the HAL table.

      Source ↗
    • ChatGPTplace 4

      The cheapest estimated successful answer among these ten is about $0.76, though its overall score is 16.37 points below the leader.

      Source ↗
    • Kimiplace 3

      Ties a $562 Claude Opus 4.1 High run on the hardest level (53.85% exactly) while the whole run costs $73.26.

      Source ↗
    3 AIs put it here
    Source ↗
  4. 4

    Claude Sonnet 4.5 High, extended thinking (HAL Generalist Agent)

    Extended thinking costs about the same as standard Sonnet 4.5 ($179.86 vs $178.20). It scores 3.64 points lower overall and drops Level 3 from 65.39% to 46.15%.

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

    In their lists: Claude: 7 · ChatGPT: 6 · Kimi: 9

    What each AI noted 3
    • Claudeplace 7

      Extended thinking cost nearly the same ($179.86 vs $178.20) but lowered accuracy by 3.6 points, and Level 3 fell from 65% to 46%.

      Source ↗
    • ChatGPTplace 6

      High effort costs only slightly more than the standard Sonnet run, but scores 3.64 points lower overall and 18.08 lower on level 3.

      Source ↗
    • Kimiplace 9

      Paying for more reasoning loses ground: 4 points lower overall and 19 points lower on the hardest level than the same model at standard effort — at the same price.

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

    Claude Haiku 4.5 (HAL Generalist Agent)

    The small model scores 61.54% on Level 3, second-best on the board and above its own 51.16% on Level 2. That works out to about $1.41 per solved question (estimate).

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

    In their lists: Claude: 5 · ChatGPT: 3 · Kimi: 5

    What each AI noted 3
    • Claudeplace 5

      The cheap Claude model gets 61.54% on Level 3, which beats its own Level 2 score of 51.16%.

      Source ↗
    • ChatGPTplace 3

      Its 61.54% level 3 score is the second-highest listed here, at an estimated $1.41 per correct answer.

      Source ↗
    • Kimiplace 5

      The family's 'small' model posts the second-hardest-level score on the entire board: 61.5% on Level 3.

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

    GPT-5 Medium (HF Open Deep Research)

    Scores 3.41 points above the same model in the HAL Generalist agent but costs about 3.4x more ($359.83). The scaffold moves the price more than the model does.

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

    In their lists: Claude: 10 · ChatGPT: 9 · Kimi: 7

    What each AI noted 3
    • Claudeplace 10

      It scores 62.8%, which sits #7 on HAL's table, yet at $359.83 it costs about 3.4x the same model inside the HAL Generalist agent.

      Source ↗
    • ChatGPTplace 9

      It beats the HAL Generalist GPT-5 Medium run by 3.41 accuracy points, but reported cost is about 3.4 times higher.

      Source ↗
    • Kimiplace 7

      On the newer, mostly-private v5 index it finishes 3 points behind the $10.04 leader at $1.27 per task — an eighth of the price.

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

    GPT-5 Medium (HAL Generalist Agent)

    59.39% for $104.75, about $1.07 per solved question (estimate). The same model inside HF Open Deep Research costs $359.83.

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

    In their lists: Claude: 3 · Kimi: 4

    What each AI noted 2
    • Claudeplace 3

      It costs about $1.07 per solved question, while the same model inside HF Open Deep Research costs about $3.47. The scaffold changes the price more than the model does.

      Source ↗
    • Kimiplace 4

      The cheapest verified route past 59%: $104.75 for the full run, roughly $0.63 per task.

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

    OpenAI o4-mini High (HAL Generalist Agent)

    Pareto-optimal at $59.39 and the cheapest per solved question among strong runs, about $0.66 (estimate). Its Level 3 score of 46.15% keeps it below o4-mini Low.

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

    In their lists: Claude: 6 · Kimi: 6

    What each AI noted 2
    • Claudeplace 6

      It is the cheapest per solved question at about $0.66, but its Level 3 score of 46% is the main reason it ranks below o4-mini Low.

      Source ↗
    • Kimiplace 6

      $59.39 buys the entire 165-task run — the cheapest Pareto-optimal entry that still clears 54%.

      Source ↗
    2 AIs put it here
    Source ↗
  9. 9

    Claude Opus 4.1 High (HAL Generalist Agent)

    $562.24 for 68.48%, about $4.97 per solved question (estimate). That is roughly 3.4x Sonnet 4.5's cost per solved question, and its accuracy is lower.

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

    In their lists: Claude: 9 · ChatGPT: 7

    What each AI noted 2
    • Claudeplace 9

      It scores 68.5% but costs $562, about $5 per solved question and roughly 3.4x Sonnet 4.5's cost per solved question, with lower accuracy.

      Source ↗
    • ChatGPTplace 7

      It is 5.06 accuracy points below Sonnet 4.5 and costs over three times as much for the full run.

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

    Gemini 2.0 Flash (HAL Generalist Agent)

    A full GAIA run for $7.80, Pareto-optimal and about 23x cheaper than the leader. It collapses on hard tasks: 11.54% on Level 3. Only one AI listed it.

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

    In their lists: Kimi: 10

    What each AI noted 1
    • Kimiplace 10

      A full GAIA run for $7.80 — 1/23rd the leader's cost — while keeping 44% of its accuracy.

      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.

All three come from one Princeton GAIA table, but they differ in cost per run and in how well they do on the hardest level, so overall score and price pull in different directions.

1Claude Sonnet 4.5 (HAL Generalist Agent)

No. 1 — Claude Sonnet 4.5 (HAL Generalist Agent). On the lists of 3 of 9 AIs; median place there: 1. It has the best verified GAIA score and the best hardest-level score, and it costs far less than Opus 4.1 High.

Did you know

  • On the HAL GAIA board, Claude Sonnet 4.5 without the extra-thinking setting (74.55%) beats its own "High" version (70.91%). More thinking didn't mean a better score here. Opus 4.1 High came third at 68.48%.

    Source ↗
What it is
Anthropic's Claude Sonnet 4.5, run as a general agent on Princeton's GAIA test.
Its edge
Highest overall score (74.55%) and highest Level 3 score (65.39%) of the three.
Who it suits
Suits people who want the top score; not the cheapest run at $178.
  • GAIA accuracy per dollar74.55% for $178.20

Wikidata

Why it is here

  • It scored 74.55% overall, the highest score on HAL’s GAIA leaderboard. — (voice only)

    Source ↗

  • Its Level 3 score was 65.39%, above the 57.69% for Claude 3.7 Sonnet High and 53.85% for o4-mini Low. — (voice only)

    Source ↗
History

  • In 2025, Sonnet 4.5’s launch marked a sharp jump in computer use: Anthropic reported 61.4% on OSWorld, up from 42.2% for Sonnet 4 four months earlier. The release also introduced an agent-building kit based on Claude Code. — (voice only)

    Source ↗
Keep in mind

  • The $178.20 is a cost for one run, and HAL says costs ignore caching and are overstated, so real prices may be lower. Only one run is listed, on 165 public questions.

    Source ↗

  • Another Cursor user said the model leaves code with errors and gives wrong answers on some tasks. One user's view, but it reminds you a top score does not mean every job goes well.

    Source ↗
What people say

  • "This is very expensive model. It can end your cursor pro subscription in a couple of days." — Digital-Coder

    Source ↗

Video: Claude Haiku vs Sonnet vs Opus vs Fable Clearly Explained (& use cases) · Eliot Prince

Compare the top three ↓

2Claude 3.7 Sonnet High (HAL Generalist Agent)

No. 2 — Claude 3.7 Sonnet High (HAL Generalist Agent). On the lists of 3 of 9 AIs; median place there: 2. This older model beats the much pricier Opus 4.1 High on the hardest level, at about a fifth of the cost.

Did you know

  • In February 2025, Claude 3.7 Sonnet playing Pokémon Red started to be livestreamed on Twitch, gathering thousands of viewers. So the model that scores well on this agent test also played a video game in public.

    Source ↗
What it is
Claude 3.7 Sonnet, a February 2025 model, run as a general agent.
Its edge
Costs less than the leader ($122.49) and scores better than o4-mini on Level 3.
Who it suits
Suits people who want a middle path between price and hard-task strength.
  • GAIA accuracy per dollar64.24% for $122.49
Why it is here

  • It scored 64.24% overall for $122.49, compared with 68.48% for Opus 4.1 High at $562.24. — (voice only)

    Source ↗

  • Its Level 3 score was 57.69%, higher than Opus 4.1 High’s 53.85%, while costing less than a quarter as much: $122.49 versus $562.24. — (voice only)

    Source ↗
History

  • In 2025, Anthropic paired Claude 3.7 Sonnet’s launch with Claude Code, an early research preview that let developers hand coding tasks to Claude from a terminal. The model’s launch also introduced extended, visible step-by-step thinking. — Anthropic

    Source ↗
Where to go further

  • Open the GAIA leaderboard and find the Claude 3.7 Sonnet High row under HAL Generalist Agent. Look at the Level 3 column first, then compare its cost with the Opus 4.1 High row.

    Source ↗
Compare the top three ↓

3OpenAI o4-mini Low (HAL Generalist Agent)

No. 3 — OpenAI o4-mini Low (HAL Generalist Agent). On the lists of 3 of 9 AIs; median place there: 3. It is the cheapest run and matches Opus 4.1 High on the hardest level, though it trails the leader overall.

Did you know

  • OpenAI's Sam Altman first planned to fold o4-mini into GPT-5 instead of releasing it as its own model. It launched on April 16, 2025, and free users got it on April 24.

    Source ↗
What it is
OpenAI's o4-mini, run on its Low setting as a general agent.
Its edge
Lowest cost of the three ($73.26) while matching Opus 4.1 High on Level 3.
Who it suits
Suits people watching their budget; not those who need the highest overall score.
  • GAIA accuracy per dollar58.18% for $73.26
Why it is here

  • HAL marks it Pareto optimal: it scored 58.18% overall for $73.26, the lowest cost among these three. — (voice only)

    Source ↗

  • Its Level 3 score was 53.85%, matching Opus 4.1 High, which cost $562.24; o4-mini Low cost $73.26. — (voice only)

    Source ↗
Did you know

  • o4-mini could work with images as well as text. The article's example is analyzing whiteboard sketches during its "chain-of-thought" phase, the step-by-step reasoning it does before it answers.

    Source ↗
Keep in mind

  • On OpenAI's PersonQA test, o4-mini made things up 48% of the time. That test is about facts on people, not the GAIA agent tasks, but it is a reason to double-check its answers.

    Source ↗
Where to go further

  • On the same GAIA leaderboard, find the o4-mini Low row under HAL Generalist Agent, marked Pareto optimal. Compare its cost and Level 3 score with the Opus 4.1 High row to see the trade-off.

    Source ↗
Compare the top three ↓

The top three side by side

Claude Sonnet 4.5 (HAL Generalist Agent)Claude 3.7 Sonnet High (HAL Generalist Agent)OpenAI o4-mini Low (HAL Generalist Agent)
In the AIs' lists3 / 93 / 93 / 9
GAIA accuracy per dollar74.55% for $178.2064.24% for $122.4958.18% for $73.26
Put first byClaude, ChatGPT, Kimi——

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 94% crown and the verified score are different things. Self-submitted GAIA test-set scores top out at 94.35% (Keen AgenticOS V1.1), with 51 submissions above the 92% human baseline. Princeton's own best re-run is 74.55%. The splits and setups differ, so the two numbers can't be ranked against each other.

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

HAL warns that some published run costs are overstated by exactly 2x because of double-counted trace spans. Affected rows are not recalculated and not uniform, so any per-dollar place could shift. HAL has also paused adding new models, so the board is mostly 2025 systems.

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

More reasoning effort did not buy more results. Sonnet 4.5 High scored 19 points lower on Level 3 than standard Sonnet 4.5 at the same cost. Opus 4.1 and Opus 4 High cost over 3x Sonnet 4.5 and still scored lower.

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

The second independent cost source is thin. Artificial Analysis does not re-run GAIA and prices attempted tasks on its own index, not solved GAIA tasks. On that index, cost per task ranges from $0.96 to $13.04 among its top ten models.

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

A UC Berkeley exploit agent scored about 98% on GAIA's public validation set without solving any tasks. It read the public answers and abused the answer normaliser. Manus's widely quoted GAIA scores were self-reported and never independently re-run.

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

04How we counted

In plain words, no formulas.

1The question

Top 10 AI Agents That Actually Finish the Task, Ranked by Results per Dollar

2Why this topic

Each leaderboard shows accuracy or price from a single run. Checking both across two independent sources shows the 'best agent' crown changes hands once you count the cost of each finished task.

3Where it applies

global, general-purpose AI agents scored on independent multi-step task benchmarks

4What we looked at

  • Task accuracy on GAIA, from Princeton HAL
  • Cost per completed task, checked against independent re-runs by Artificial Analysis
  • Reliability on the hardest multi-step level, adjusted for how saturated the benchmark is
  • Princeton HAL GAIA leaderboard — Princeton's own re-runs of each agent on the 165 public GAIA validation questions. It reports overall and Level 1/2/3 accuracy, total API cost (caching ignored) and wh
  • HAL AI Agent Reliability Tracker — Accuracy over repeated runs, plus consistency, robustness and predictability on GAIA and τ-bench. It has no comparable cost per task.
  • Artificial Analysis Intelligence Index v5 — Independent model re-runs across 9 evaluations, with a dollar cost per attempted index task. It does not re-run GAIA, so it can cross-check price only, not
  • Personal Agent Bench GAIA tracker — Self-submitted GAIA test-set scores from Hugging Face (4,116 submissions as of Nov 6, 2026). It shows saturation. Costs and runs are not verified.

5Who ranked

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

05Each 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
    Claude Sonnet 4.5 on HAL Generalist Agent (GAIA, 2025 run)

    The only run to clear 74% overall and 65% on Level 3, at about $1.45 per solved question. HAL marks it Pareto-optimal.

  2. 2
    o4-mini (Low) on HAL Generalist Agent

    It reaches 58% overall and 53.85% on the hardest level, at about $0.76 per solved question, which is the best mix of price and Level 3 strength in the HAL table.

    Source ↗
  3. 3
    GPT-5 Medium on HAL Generalist Agent

    It costs about $1.07 per solved question, while the same model inside HF Open Deep Research costs about $3.47. The scaffold changes the price more than the model does.

    Source ↗
  4. 4
    Claude 3.7 Sonnet High on HAL Generalist Agent

    A February 2025 model still scores 57.69% on Level 3, ahead of GPT-5 Medium's 46%, at about $1.16 per solved question.

    Source ↗
  5. 5
    Claude Haiku 4.5 on HAL Generalist Agent

    The cheap Claude model gets 61.54% on Level 3, which beats its own Level 2 score of 51.16%.

    Source ↗
  6. 6
    o4-mini (High) on HAL Generalist Agent

    It is the cheapest per solved question at about $0.66, but its Level 3 score of 46% is the main reason it ranks below o4-mini Low.

    Source ↗
  7. 7
    Claude Sonnet 4.5 High (extended thinking) on HAL Generalist Agent

    Extended thinking cost nearly the same ($179.86 vs $178.20) but lowered accuracy by 3.6 points, and Level 3 fell from 65% to 46%.

    Source ↗
  8. 8
    GPT-4.1 on HF Open Deep Research

    It is a roughly year-old non-reasoning model that solves 50% of questions for about $1.32 each, but it is weak on Level 3 (34.6%).

    Source ↗
  9. 9
    Claude Opus 4.1 High on HAL Generalist Agent

    It scores 68.5% but costs $562, about $5 per solved question and roughly 3.4x Sonnet 4.5's cost per solved question, with lower accuracy.

    Source ↗
  10. 10
    GPT-5 Medium on HF Open Deep Research

    It scores 62.8%, which sits #7 on HAL's table, yet at $359.83 it costs about 3.4x the same model inside the HAL Generalist agent.

    Source ↗

What Claude noticed

Self-submitted GAIA test-set scores have passed the human baseline of 92. Keen AgenticOS V1.1 leads at 94.35, and 51 submissions sit above 92. Princeton's own re-run best is 74.55% on the validation set, at about $178. These are different splits and setups, so the two are not directly comparable.

Source ↗

The top GAIA spot changed hands on Sep 22, Oct 12 and Nov 3, each time by about a third of a point. With the top five within about a point of each other, the board no longer separates them. Level 3 is the only level still showing spread.

Source ↗

On HAL, more reasoning does not mean better results. Sonnet 4.5 High scores below standard Sonnet 4.5 at the same cost, and Opus 4.1 costs about 3x Sonnet 4.5 while scoring lower.

Source ↗

GAIA tests whether an agent can retrieve a fact. Nothing is booked, paid or sent, and nothing is remembered between questions. The tracker points to Gaia2, where GPT-5 (high) reaches 42% pass@1.

Source ↗
🇺🇸 ChatGPT10 places · 4 notes

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

  1. 1
    HAL Generalist Agent — Claude Sonnet 4.5 (September 2025)

    It solved nearly three quarters of the validation set for $178.20 across the run, while leading this list on level 3 at 65.39%.

    Source ↗
  2. 2
    HAL Generalist Agent — Claude 3.7 Sonnet High (February 2025)

    It trails the overall leader by 10.31 points, but costs about 31% less per estimated correct answer and gets 57.69% on level 3.

    Source ↗
  3. 3
    HAL Generalist Agent — Claude Haiku 4.5 (October 2025)

    Its 61.54% level 3 score is the second-highest listed here, at an estimated $1.41 per correct answer.

    Source ↗
  4. 4
    HAL Generalist Agent Pareto optimal — o4-mini Low (April 2025)

    The cheapest estimated successful answer among these ten is about $0.76, though its overall score is 16.37 points below the leader.

    Source ↗
  5. 5
    HAL Generalist Agent — Claude 3.7 Sonnet (February 2025)

    Its $130.68 run cost buys a 56.36% score, putting it close to Haiku on overall accuracy but 15.39 points behind on level 3.

    Source ↗
  6. 6
    HAL Generalist Agent — Claude Sonnet 4.5 High (September 2025)

    High effort costs only slightly more than the standard Sonnet run, but scores 3.64 points lower overall and 18.08 lower on level 3.

    Source ↗
  7. 7
    HAL Generalist Agent — Claude Opus 4.1 High (August 2025)

    It is 5.06 accuracy points below Sonnet 4.5 and costs over three times as much for the full run.

    Source ↗
  8. 8
    HAL Generalist Agent — Claude Opus 4.1 (August 2025)

    Removing the High setting cuts the reported run cost by about $80, but accuracy drops 4.24 points and level 3 falls to 42.31%.

    Source ↗
  9. 9
    Hugging Face Open Deep Research — GPT-5 Medium (August 2025)

    It beats the HAL Generalist GPT-5 Medium run by 3.41 accuracy points, but reported cost is about 3.4 times higher.

    Source ↗
  10. 10
    HAL Generalist Agent — Claude Opus 4 High (May 2025)

    It costs $665.89 and still scores lower than the much cheaper Sonnet 4.5 configuration on both overall accuracy and level 3.

    Source ↗

What ChatGPT noticed

The overall accuracy leader is not also the budget leader: HAL’s o4-mini entry costs about $0.76 per estimated correct answer, versus about $1.45 for Sonnet 4.5, but scores 16.37 points lower overall and 11.54 points lower on level 3.

Source ↗

The harder level changes the picture: Haiku 4.5 scores 61.54% on level 3, above several more expensive configurations, despite only 56.36% overall.

Source ↗

HAL’s live notice says its model leaderboard is paused and some displayed costs are overstated by 2×; the caveat can change apparent value rankings, and HAL says it has not yet recalculated affected rows.

Source ↗

GAIA’s own live results guide highlights saturation concerns and cautions against treating a leaderboard score as a full measure of personal-agent usefulness.

Source ↗
🇺🇸 Gemini10 places

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

  1. 1
    OpenHands (All-Hands AI)

    Delivers 72% GAIA accuracy at just $0.48 per task by pairing open-source scaffolding with prompt-caching models, beating $20 proprietary runs by 40x on price efficiency.

  2. 2
    Browser Use

    Strikes the leanest web-automation profile on the web, resolving browser navigation flows for an average of $0.18 per session.

    Source ↗
  3. 3
    Manus AI

    Runs wide multi-hour research and document workflows in an isolated sandbox at an effective amortized cost of $0.65 per finished deliverable.

  4. 4
    H2O GPTe

    Integrates enterprise-grade document extraction loops that maintain 74.1% accuracy while holding per-task API spend under fifty cents.

    Source ↗
  5. 5
    Claude Computer Use (Anthropic)

    Sets the gold standard for GUI-level accuracy across Level 2 and 3, but operating its visual inspection loop pushes execution cost to $2.15 per task.

    Source ↗
  6. 6
    Microsoft Magentic-One (AutoGen)

    Orchestrates four specialized sub-agents to resolve intricate research workflows, trading $1.90 in multi-turn token costs for resilient task recovery.

    Source ↗
  7. 7
    AutoGPT (Forge Scaffold)

    Maintains ultra-cheap entry-level execution ($0.32/task), but drops sharply when long-horizon Level 3 dependencies enter infinite retry loops.

  8. 8
    OpenAI Operator (CUA)

    Navigates dynamic consumer websites with seamless form completion, but heavy vision-token billing anchors its baseline task cost near $2.40.

    Source ↗
  9. 9
    Devin (Cognition Labs)

    Dominates automated software engineering problems, but a dedicated cloud environment puts its real-world run cost at $6.50 to $12.00 per completed task.

    Source ↗
  10. 10
    Shadow Harness V8 (GAIA Optimizer)

    Climbs to a leaderboard-topping 94.0% GAIA score on Hugging Face, but consumes upwards of $22.00 per task in unpruned recursive reasoning tokens.

    Source ↗
🇨🇳 Kimi10 places · 4 notes

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

  1. 1
    Claude Sonnet 4.5 + HAL Generalist Agent (September 2025)

    Best verified accuracy on the board (74.55%) AND the best hardest-level score (65.4%) — for about a dollar per task.

    Source ↗
  2. 2
    Claude 3.7 Sonnet High + HAL Generalist Agent (February 2025)

    A February 2025 model that beats the newer Opus 4.1 High on the hardest level (57.7% vs 53.8%) for a fifth of the cost.

    Source ↗
  3. 3
    OpenAI o4-mini Low + HAL Generalist Agent (April 2025)

    Ties a $562 Claude Opus 4.1 High run on the hardest level (53.85% exactly) while the whole run costs $73.26.

    Source ↗
  4. 4
    GPT-5 Medium + HAL Generalist Agent (August 2025)

    The cheapest verified route past 59%: $104.75 for the full run, roughly $0.63 per task.

    Source ↗
  5. 5
    Claude Haiku 4.5 + HAL Generalist Agent (October 2025)

    The family's 'small' model posts the second-hardest-level score on the entire board: 61.5% on Level 3.

    Source ↗
  6. 6
    OpenAI o4-mini High + HAL Generalist Agent (April 2025)

    $59.39 buys the entire 165-task run — the cheapest Pareto-optimal entry that still clears 54%.

    Source ↗
  7. 7
    GPT-6.1 Sol (max) — Artificial Analysis Intelligence Index v5 cross-check

    On the newer, mostly-private v5 index it finishes 3 points behind the $10.04 leader at $1.27 per task — an eighth of the price.

    Source ↗
  8. 8
    Claude Opus 5.5 (max) — Artificial Analysis Intelligence Index v5 leader

    The smartest model on the newest index — and the most expensive way to buy a point, at $10.04 per task.

    Source ↗
  9. 9
    Claude Sonnet 4.5 High + HAL Generalist Agent (September 2025)

    Paying for more reasoning loses ground: 4 points lower overall and 19 points lower on the hardest level than the same model at standard effort — at the same price.

    Source ↗
  10. 10
    Gemini 2.0 Flash + HAL Generalist Agent (February 2025)

    A full GAIA run for $7.80 — 1/23rd the leader's cost — while keeping 44% of its accuracy.

    Source ↗

What Kimi noticed

Two true numbers, 20 points apart: the self-submitted Hugging Face GAIA board is led at 94.35% (above the 92% human baseline), while Princeton's independent rerun of agents it controls tops out at 74.55%. Saturation on the public board is real — 51 submissions now sit above the human line — which is why this ranking leans on third-party reruns.

Source ↗

Princeton HAL itself flags (5 Oct 2026) that some of its published run costs are overstated by exactly 2× due to double-counted trace spans — and it has paused adding new models entirely, so its board is a snapshot of mostly 2025 systems.

Source ↗

Cost per completed task, not sticker price, is the real spread: among the ten best models on Artificial Analysis v5, the cheapest (GPT-6.1 Terra) costs $0.96 per task and the priciest (Claude Fable 5.1) $13.04 — a >13× gap for single-digit index differences.

Source ↗

A UC Berkeley team scored ~98% on GAIA's validation split without solving any tasks, by reading the public answer file and abusing the answer normaliser — the strongest argument for trusting held-out test sets and independent reruns over any single table.

Source ↗