Back to the trend of the day: Meta says its cheap AI matches Anthropic's for a third the cost
Poly Rating · from the trend of the day

Cheapest AI for Agents 2026: Real Bill per Finished Task

Muse Ember and Claude Haiku 5.5 both score 36 and both charge $0.10 per million input tokens and $0.50 per million output tokens, yet one finished test task costs about $0.14 on Ember and $0.41 on Haiku.

  • No. 1: Meta Muse Ember
  • 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

    Meta Muse Ember

    Scores 36 for $0.14 a task, the cheapest model at that score. However, its 69.5% OSWorld result counts partial credit; only 33.8% of tasks fully pass, and both figures are Meta's own.

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

    In their lists: Claude: 2 · ChatGPT: 1 · Meta: 1 · Grok: 1 · Z.ai: 1 · Kimi: 1 · DeepSeek: 1 · Qwen: 1

    What each AI noted 8
    • Claudeplace 2

      $0.14 per index task for a score of 36, but the OSWorld claim reads 69.5% with partial credit and only 33.8% when a task must fully pass.

      Source ↗
    • ChatGPTplace 1

      Nearly Haiku's intelligence for $0.136 instead of $0.411 per index task; the real surprise is a $0.004 cached-input rate.

      Source ↗
    • Metaplace 1

      Same 36 as Haiku on AA v5 at max, for $0.14 a task instead of $0.41 - the cheapest way to buy that score.

      Source ↗
    • Grokplace 1

      Independent 36 for about $0.14 a finished task; Haiku's 36 costs about $0.41.

      Source ↗
    • Z.aiplace 1

      Score-for-score the cheapest model Artificial Analysis has measured: 36 on the index for $0.14 a task - about a third of Claude Haiku 5.5's $0.41 bill for the same score.

      Source ↗
    • Kimiplace 1

      $0.14 to finish one index task where Haiku 5.5 needs $0.41 at the same $0.10/$0.50 sticker price — 2.6x fewer output tokens.

      Source ↗
    • DeepSeekplace 1

      Matches Haiku's intelligence score at $0.14 per task — one-third the bill — and still leads the cheap field on business workflows.

      Source ↗
    • Qwenplace 1

      Matches Claude Haiku's brain for a third of the bill.

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

    Claude Haiku 5.5

    Same score of 36, but $0.41 a task, about three times Ember's bill. Above 100K prompt tokens its input price rises fivefold, to $0.50 per million.

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

    In their lists: Claude: 1 · ChatGPT: 2 · Meta: 2 · Grok: 2 · Z.ai: 3 · Kimi: 2 · DeepSeek: 4 · Qwen: 2

    What each AI noted 8
    • Claudeplace 1

      Same index score as Ember (36), but 3x the bill per task ($0.41 vs $0.14), and it has the cleanest record in the one independent test of unattended work.

      Source ↗
    • ChatGPTplace 2

      The same rounded intelligence score as Ember costs roughly three times as much, but Haiku wins the tougher code-merge comparison.

      Source ↗
    • Metaplace 2

      The coding-agent leader among the $0.10 models, but it writes about 3x the output tokens of Luna per AA task.

      Source ↗
    • Grokplace 2

      Same rounded 36, about three times the bill, and the one that still wins merge-ready code.

      Source ↗
    • Z.aiplace 3

      The proven one: ties Ember at 36 and beats every cheap peer on merge-ready code, but the bill is $0.41 a task - and past 100k prompt tokens the input rate jumps 5x to $0.50.

      Source ↗
    • Kimiplace 2

      Writes 298M output tokens to Ember's 115M per index run at the same output rate — the $0.27/task gap is verbosity, not the price card.

      Source ↗
    • DeepSeekplace 4

      The best OSWorld score in the cheap class at 72.4%, but the same $0.10/$0.50 list price bills $0.41 per task — nearly three times Muse Ember — and it finishes only 13.9% of strict business workflows.

      Source ↗
    • Qwenplace 2

      The baseline for cheap agents, now matched for pennies.

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

    DeepSeek V4.1 Flash

    Has the cheapest cached input, $0.003 per million, yet AA billed $0.60 a task for a score of 33. DeepSeek reports 42.3% on Terminal-Bench; Vals measured 31.82%.

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

    In their lists: Claude: 3 · ChatGPT: 4 · Meta: 5 · Grok: 4 · Z.ai: 5 · Kimi: 4 · DeepSeek: 3

    What each AI noted 7
    • Claudeplace 3

      Cache hits at $0.003 per million make it the cheapest cached input here, and it topped Bito's long-session test at 71.9%.

      Source ↗
    • ChatGPTplace 4

      A 98% cache discount sounds unbeatable, yet reasoning and output consumption bring the measured index bill to $0.60.

      Source ↗
    • Metaplace 5

      AA 33 for $0.60 a task and a Vals Terminal-Bench run at $0.54 - the low-bill terminal worker in this set.

    • Grokplace 4

      Index 33 costs $0.60 a task, and the official rate doubles when the clock hits peak.

      Source ↗
    • Z.aiplace 5

      MIT-licensed weights plus a $0.003 cache-hit rate let one developer run 445 terminal trials for $27 total - but its self-reported 42.3% on Terminal-Bench 4.0 fell to 19.7% in Vals' independent run.

      Source ↗
    • Kimiplace 4

      Top released open weights on the index at 32 — and 42.3% on Terminal-Bench 4.0 for roughly $0.60–$1.20 per 1M output tokens.

      Source ↗
    • DeepSeekplace 3

      At $0.60 per task it is not the cheapest token, but it finishes 42.3% of terminal tasks and 60.5% of business workflows — the best terminal-plus-office combination in the cheap tier.

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

    GPT-6 Luna

    Same $0.10 / $0.50 prices as Ember and Haiku, and about the same bill as Ember at $0.15 a task, but it scores only 27.8. On OSWorld only 17% of tasks fully pass.

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

    In their lists: Claude: 5 · ChatGPT: 3 · Meta: 4 · Grok: 9 · Z.ai: 6 · Kimi: 3 · DeepSeek: 5

    What each AI noted 7
    • Claudeplace 5

      Same $0.10 / $0.50 price tag as Ember and Haiku, but a score of 27.8 at max effort and only 16.4% on Terminal-Bench 4.0.

      Source ↗
    • ChatGPTplace 3

      Just $0.073 buys a low-effort index task, but the score falls to 21.4; max effort reaches 27.8 for $0.152.

      Source ↗
    • Metaplace 4

      Same $0.10 / $0.50 list price as Haiku and Ember, but about 50K output tokens per AA task at max vs Haiku's about 162K.

      Source ↗
    • Grokplace 9

      Same $0.10 sticker as Ember, 27.8 on the index, and 17% binary success on the computer test.

      Source ↗
    • Z.aiplace 6

      The bill champion: $0.01 per index task at low effort - 15x less than its own max setting - and one real-world task cost $0.071; the catch is it resolved only 13.6% of Vals' terminal tasks.

      Source ↗
    • Kimiplace 3

      Cheapest per task in the class ($0.07–$0.15) but only 16.4% on Terminal-Bench 4.0 — a routing model, not a finisher.

      Source ↗
    • DeepSeekplace 5

      Same $0.10/$0.50 price card as Haiku and Ember, but only 28 Intelligence and 16.4% on Terminal-Bench — the cheapest bill can still be a false economy for long jobs.

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

    MiniMax M3.1 Flash

    Scores 29 at $0.49 a task. MiniMax reports 38.1% on Terminal-Bench; Vals measured 25.25% with a shared harness. A 20-turn Bito run cost $3.62.

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

    In their lists: Claude: 4 · Meta: 6 · Grok: 7 · Z.ai: 2 · Kimi: 6 · DeepSeek: 6

    What each AI noted 6
    • Claudeplace 4

      Within one point of Haiku 5.5 on Bito's 20-turn test (75.0%), but a run took 3 to 5 hours and cost $3.62.

      Source ↗
    • Metaplace 6

      Office-workflow star at 72.5% on AutomationBench-AA, terminal worker at only about 26% - buy it for apps, not shells.

      Source ↗
    • Grokplace 7

      A self-reported 38.1% terminal score falls to 25.25% when Vals uses one harness.

      Source ↗
    • Z.aiplace 2

      Second-best office-workflow score of all 249 models AA tested (72.5%) while billing $0.21 per terminal test in Vals' independent run - roughly 15x cheaper per test than Haiku 5.5.

      Source ↗
    • Kimiplace 6

      38.1% on Terminal-Bench 4.0 at $0.60 output — the quiet third-party that out-terminals Luna by 22 points.

      Source ↗
    • DeepSeekplace 6

      The office-workflow specialist: 72.5% on AutomationBench-AA, the highest in the cheap group, but only 26% on Terminal-Bench and 21% on FrontierCode.

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

    GLM-5.3-Flash

    Its $0.15 / $0.50 prices look cheap, but it scores 28 at $0.58 a task. It writes more tokens than Luna, so the bill comes out higher. An older AA comparison shows 42 at $0.25.

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

    In their lists: ChatGPT: 5 · Meta: 7 · Grok: 3 · Kimi: 7 · DeepSeek: 8

    What each AI noted 5
    • ChatGPTplace 5

      A 35.8% official terminal-task resolution rate makes this open-weight Flash model a serious agent candidate.

      Source ↗
    • Metaplace 7

      $0.15 in / $0.50 out list price, AA 28 for $0.58 a task - cheap on paper, dearer per task than Luna because it writes more.

      Source ↗
    • Grokplace 3

      Open weights finish 60.4% of office objectives, then only 17.6% of merge-ready code tasks.

      Source ↗
    • Kimiplace 7

      35.8% on Terminal-Bench 4.0 for $0.36 a task — the cheapest measured pass rate above one-third on the board.

      Source ↗
    • DeepSeekplace 8

      A $0.58-per-task model with a strong 60% AutomationBench score but only 15% on FrontierCode and 33% on Terminal-Bench — office-capable, merge-shy.

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

    Gemini 3.8 Flash

    The widest gap between list price and real bill: a $0.75 input price turns into $7.77 per index task at high effort, for a score of 31.

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

    In their lists: ChatGPT: 6 · Meta: 9 · Grok: 10 · Z.ai: 9 · DeepSeek: 9

    What each AI noted 5
    • ChatGPTplace 6

      Its 1-million-token context and multimodal inputs are attractive, but $1.24 per index task weakens the cheap-agent argument.

      Source ↗
    • Metaplace 9

      AA 31 at high effort, but AA bills $7.77 per task - the clearest proof that list price is not the bill.

      Source ↗
    • Grokplace 10

      Highest knowledge accuracy in the Ember comparison, 50.5%, at $0.75 per million input tokens.

      Source ↗
    • Z.aiplace 9

      Fans call it Opus-level intelligence at a fifth of the price; Vals' meter says $8.65 per terminal test - the biggest bill in the budget class despite a mid-pack 19.7% accuracy.

      Source ↗
    • DeepSeekplace 9

      A low per-token price ($0.75 input) that turns into $7.77 per finished task — the widest gap between sticker price and real bill in the cheap class.

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

    Xiaomi MiMo-V2.6-Flash

    Its $0.13 bill per task matches Ember's, but it scores only 27, and each index task takes 64 minutes against Ember's 3.1.

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

    In their lists: Meta: 8 · Grok: 8 · Kimi: 9

    What each AI noted 3
    • Metaplace 8

      Joint-cheapest AA task in the set at $0.13, for AA 27 - a thrifty router, not a finisher, on the evidence found.

      Source ↗
    • Grokplace 8

      The $0.13 task bill looks like Ember's, then the task takes 64 minutes and scores 27.

      Source ↗
    • Kimiplace 9

      24.2% on Terminal-Bench 4.0 for $0.21 a task — the lowest measured cost per terminal task on the efficiency frontier.

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

    Thinking Machines Inkling 2 Small

    The word 'Small' covers a 594-billion-parameter model. It scores 31 at $1.18 a task and resolves only 12.63% of Vals' terminal tasks.

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

    In their lists: Claude: 6 · ChatGPT: 7

    What each AI noted 2
    • Claudeplace 6

      Scores 31 on the index, but only 13% on Terminal-Bench 4.0 against 30% for Sonnet 5.5 (medium).

      Source ↗
    • ChatGPTplace 7

      The word Small hides a 594-billion-parameter model and a $1.18 evaluation bill.

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

    Tencent Hy4

    Close to Ember's score at 35 and level with it on Vals' terminal test at 39.39%, but it costs $1.28 per terminal test against Ember's $0.22.

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

    In their lists: Meta: 10 · Grok: 5

    What each AI noted 2
    • Metaplace 10

      AA 35 for $1.03 a task - higher than every $0.10 model here, and the price of that extra score is visible.

      Source ↗
    • Grokplace 5

      Ties Ember on one terminal board at 39.4%, at $1.28 a test against Ember's $0.22.

    2 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.

Scores differ (36, 36, 33). The OSWorld and Terminal-Bench figures come from different sources: some are the makers' own, and Vals measured its own run.

1Meta Muse Ember

No. 1 — Meta Muse Ember. On the lists of 8 of 10 AIs; median place there: 1. It gets the top place because it matches Haiku's score of 36 for about a third of the bill, $0.14 a task.

Did you know

  • Meta's own table shows Muse Ember at 69.5% partial credit on OSWorld 2.1 but 33.8% fully passed. Claude Haiku 5.5 scores 68.9% partial yet only 27.8% fully passed, so the two look almost tied until you count full passes.

    Source ↗
What it is
Meta's smallest, most efficient model, released on 9 November 2026.
Its edge
Cheapest bill: $0.14 a task against $0.41 and $0.60 for the others.
Who it suits
Suits cheap agent runs; Meta's own OSWorld numbers have no independent check yet.
  • Bill per finished task$0.14 per task
Why it is here

  • Artificial Analysis calls Muse Ember the cheapest model scoring 36 on its Intelligence Index. It costs $0.14 a task, while Claude Haiku 5.5 scores 36 at $0.41.

    Source ↗
Did you know

  • Muse Ember is priced at $0.10 per million input tokens and $0.50 per million output. It takes a 1M-token context and can write up to 512K tokens in one reply, which is unusually long for a model this cheap.

    Source ↗
History

  • On 9 November 2026, Meta released Muse Ember, its smallest and most efficient model. It was the first model trained using Muse Wonder Full, Meta's larger model for the hardest problems.

    Source ↗
Keep in mind

  • Meta's methodology page says refusals get zero credit and stay in the count. It also says model-based checks are judged by Claude Sonnet 4.6. I found no independent review of Ember's OSWorld results.

    Source ↗
Where to go further

  • Start with the OSWorld 2.1 table on Meta's blog. Compare the partial-credit numbers with the binary completion numbers, since the gap shows how many tasks fully finish. The model ID is muse-ember-1.0.

    Source ↗
Compare the top three ↓

2Claude Haiku 5.5

No. 2 — Claude Haiku 5.5. On the lists of 8 of 10 AIs; median place there: 2. Same score of 36 as Ember, but the bill is about three times higher, and the input price jumps fivefold on long prompts.

What it is
Claude's small, cheap model, priced at $0.10 input and $0.50 output per million tokens.
Its edge
Ties Ember on score 36 and wins merge-ready code, but costs three times more.
Who it suits
Suits coding agents with short prompts; not long runs past 100K tokens.
  • Bill per finished task$0.41 per task

Wikidata

Why it is here

  • Haiku 5.5 matches Muse Ember's Intelligence Index score of 36 at $0.41 a task, against $0.14 for Ember. That is about three times the bill for the same score.

    Source ↗

  • Haiku 5.5 input costs $0.10 per million tokens up to 100K tokens. Past 100,000 tokens every rate is multiplied by five, so input becomes $0.50. Long agent runs can hit this jump.

    Source ↗
Claude put it first. What they saw in it

  • Claude: Same index score as Ember (36), but 3x the bill per task ($0.41 vs $0.14), and it has the cleanest record in the one independent test of unattended work.

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

Compare the top three ↓

3DeepSeek V4.1 Flash

No. 3 — DeepSeek V4.1 Flash. On the lists of 7 of 10 AIs; median place there: 4. It has the cheapest cached input, but it scores only 33 and the measured bill is $0.60 a task.

Did you know

  • DeepSeek's cached-input price depends on the hour. One write-up lists $0.003 per million tokens off-peak and $0.006 at peak, so the headline $0.003 is the off-peak rate.

    Source ↗
What it is
DeepSeek's small model, which took over from its larger V4-Pro.
Its edge
Cheapest cached input: $0.003 per million tokens, far below the other two.
Who it suits
Suits jobs that reuse the same text; not those chasing the lowest bill per task.
  • Bill per finished task$0.60 per task
Why it is here

  • DeepSeek V4.1 Flash is listed at $0.15 input, $0.60 output and $0.003 cached input per million tokens. That cache price is very low, but it only helps when the same text is reused.

    Source ↗
History

  • On 10 September 2026, DeepSeek launched V4.1-Flash and retired V4-Pro, its top model. From 14 September, every V4-Pro request was sent to Flash instead, and that arrangement runs until a V4.1-Pro arrives, with no date given.

    Source ↗ ✓ confirmed by several AIs
Keep in mind

  • The same user calls the model "a bit outdated" and says it needs reminders about what tools to use. The page notes the tasks and settings were not shared, so it is one person's experience, not a controlled test.

    Source ↗

  • On Humanity's Last Exam, a hard reasoning test, V4.1 Flash scores 36.8 against 44.5 for GPT-5.6 Sol and 56.3 for Claude Opus 5. The page says these figures come from DeepSeek's own table.

    Source ↗
Where to go further

  • Read DeepSeek's September 10 launch note first. Check the model name deepseek-flash and the peak and off-peak note, where off-peak rates are 50% of peak. The exact prices are only in a picture on this page.

    Source ↗
Compare the top three ↓

The top three side by side

Meta Muse EmberClaude Haiku 5.5DeepSeek V4.1 Flash
In the AIs' lists8 / 108 / 107 / 10
Bill per finished task$0.14 per task$0.41 per task$0.60 per task
Put first byChatGPT, Meta, Grok, Z.ai, Kimi, DeepSeek, QwenClaude—

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.

Ember, Haiku 5.5 and Luna all list at $0.10 input and $0.50 output, yet one index task costs about $0.14, $0.41 and $0.15. Two things explain Ember's lower bill: cached input at $0.004 per million against $0.01, and about a third of Haiku's output tokens.

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

Ember's 69.5% on OSWorld counts partial credit. Only 33.8% of its computer tasks fully pass, and both numbers come from Meta's own run, with no independent OSWorld score found.

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

Long prompts change the bill. Haiku's input price rises fivefold above 100K tokens, Ember's prices triple above 200K, and Luna's rise above 272K.

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

The cheapest run is not the most trustworthy one. In Bito's 20-turn coding sessions, Ember cost $0.95 but scored 56.1%, with about 47 misleading claims per session. Bito sells a model router, so it has a commercial interest.

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

Muse Spark 1.3 scores just one point above Ember (37 vs 36) but costs $2.39 a task, about 17 times as much.

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

04Questions people ask

What is the cheapest AI agent?

Measured by the bill per finished task, Muse Ember is the cheapest model scoring 36 on the Artificial Analysis index, at about $0.14 a task. Claude Haiku 5.5 reaches the same score for about $0.41.

Source ↗
What is the cheapest AI coding agent?

Ember is cheapest per run (a 20-turn Bito session cost $0.95), but Haiku 5.5 writes more merge-ready code: 34.9% vs 27.2% on Artificial Analysis's FrontierCode 2.0 run.

Source ↗
Muse Ember vs Haiku: which is cheaper?

Both list at $0.10/$0.50 per million tokens and round to 36 (35.8 vs 36.2), but Ember bills about $0.14 per task against Haiku's $0.41.

Source ↗
Why do sources give different costs for Ember?

Meta's chart shows $0.136 per task, calculated as an estimate at proposed pricing, while Artificial Analysis's live figure is $0.14. At xhigh effort the cost drops to $0.116 for a score of 33.9.

Source ↗

05How we counted

In plain words, no formulas.

1The question

Top 10 Cheap AI Models for Agents: Same Price Tag, Different Bill per Finished Task

2Why this topic

The reader sees, side by side, the sticker price, the real bill per finished task, and whether the maker's launch claim held, which no price list or leaderboard shows together.

3Where it applies

Global, small and low-price AI models used to run AI agents

4What we looked at

  • Intelligence score from an independent evaluator at a stated effort setting, not the maker's own number
  • Dollars actually billed to finish one test task, counting reasoning tokens and cache hits, not the price per million tokens
  • Official input, cached-input and output prices from each maker's price page
  • Success on long, step-by-step computer and browser tasks
  • Gap between the maker's launch claim and the independent result
  • Artificial Analysis Intelligence Index v5 — Independent composite score at a stated effort setting, plus dollars billed per index task including reasoning tokens and cache hits
  • Maker price pages (Meta, Anthropic, OpenAI, DeepSeek) — List price per million input, cached-input and output tokens, including higher rates for long prompts
  • OSWorld 2.1 (Meta launch table) — Multi-step computer-use tasks, scored with partial credit and as a strict binary pass. These are Meta's own runs.

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, Z.ai, Kimi, DeepSeek, Wenxin, 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
    Claude Haiku 5.5

    Same index score as Ember (36), but 3x the bill per task ($0.41 vs $0.14), and it has the cleanest record in the one independent test of unattended work.

    Source ↗
  2. 2
    Meta Muse Ember (2026)

    $0.14 per index task for a score of 36, but the OSWorld claim reads 69.5% with partial credit and only 33.8% when a task must fully pass.

    Source ↗
  3. 3
    DeepSeek V4.1 Flash

    Cache hits at $0.003 per million make it the cheapest cached input here, and it topped Bito's long-session test at 71.9%.

    Source ↗
  4. 4
    MiniMax M3.1 Flash

    Within one point of Haiku 5.5 on Bito's 20-turn test (75.0%), but a run took 3 to 5 hours and cost $3.62.

    Source ↗
  5. 5
    GPT-6 Luna

    Same $0.10 / $0.50 price tag as Ember and Haiku, but a score of 27.8 at max effort and only 16.4% on Terminal-Bench 4.0.

    Source ↗
  6. 6
    Inkling 2 Small (Thinking Machines)

    Scores 31 on the index, but only 13% on Terminal-Bench 4.0 against 30% for Sonnet 5.5 (medium).

    Source ↗
  7. 7
    Kimi K3.1 Pro

    Scores 64.9% on Bito's long sessions, but a 20-turn run cost $29.32, about 30x Ember's $0.95.

    Source ↗
  8. 8
    GLM-5.3 (Z.ai)

    $1.40 input list price, yet a 20-turn Bito run cost $20.28 for a 60.1% overall score.

    Source ↗
  9. 9
    Muse Spark 1.3 (Meta)

    One point above Ember on the index (37 vs 36) at 17x the bill per task ($2.39 vs $0.14).

    Source ↗
  10. 10
    Qwen3.8 Flash (Alibaba)

    Among the lowest list prices ($0.15 / $0.47), but I found no independent score or per-task bill, so it ranks last on missing evidence.

    Source ↗

What Claude noticed

Ember, Haiku 5.5 and Luna list at the same $0.10 / $0.50, yet cost per index task runs from $0.14 (Ember) to $0.15 (Luna) to $0.41 (Haiku). Cheap cache reads and about a third of Haiku's output tokens explain the Ember gap.

Source ↗

Haiku 5.5's score depends on effort: 24.6 at $0.092 per task, 36.2 at $0.411. The stated effort setting changes both numbers.

Source ↗

Above 200K tokens Ember's prices triple (to $0.30 / $1.50), so a long agent run costs more than the headline price suggests.

Source ↗

In Bito's tests the cheapest model per run was not the most trustworthy one: Ember made about 47 misleading claims per 20-turn session versus about 18 to 22 for Kimi and DeepSeek.

Source ↗
🇺🇸 ChatGPT10 places · 4 notes

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

  1. 1
    Meta Muse Ember (2026)

    Nearly Haiku's intelligence for $0.136 instead of $0.411 per index task; the real surprise is a $0.004 cached-input rate.

    Source ↗
  2. 2
    Claude Haiku 5.5 (2026)

    The same rounded intelligence score as Ember costs roughly three times as much, but Haiku wins the tougher code-merge comparison.

    Source ↗
  3. 3
    OpenAI GPT-6 Luna (2026)

    Just $0.073 buys a low-effort index task, but the score falls to 21.4; max effort reaches 27.8 for $0.152.

    Source ↗
  4. 4
    DeepSeek V4.1 Flash 1017 (2026)

    A 98% cache discount sounds unbeatable, yet reasoning and output consumption bring the measured index bill to $0.60.

    Source ↗
  5. 5
    Z.ai GLM-5.3-Flash (2026)

    A 35.8% official terminal-task resolution rate makes this open-weight Flash model a serious agent candidate.

    Source ↗
  6. 6
    Google Gemini 3.8 Flash (2026)

    Its 1-million-token context and multimodal inputs are attractive, but $1.24 per index task weakens the cheap-agent argument.

    Source ↗
  7. 7
    Thinking Machines Inkling 2 Small (2026)

    The word Small hides a 594-billion-parameter model and a $1.18 evaluation bill.

    Source ↗
  8. 8
    OpenAI GPT-6.1 Sol (2026)

    Too expensive to be a default micro-agent, but 58.2% terminal resolution makes it a credible escalation model.

    Source ↗
  9. 9
    Meta Muse Spark 1.3 (2026)

    Only one additional rounded intelligence point over Ember costs about 17.6 times as much per index task.

    Source ↗
  10. 10
    Meta Muse Spark Open (2026)

    Open weights improve deployment control, but 6.2% on FrontierCode leaves little margin for unattended code changes.

    Source ↗

What ChatGPT noticed

Equal token prices do not produce equal bills: Ember and Haiku both charge $0.10 input and $0.50 output per million, but their max-effort index tasks cost $0.136 and $0.411 respectively.

Source ↗

Ember's cached input costs $0.004 per million versus $0.01 for Haiku and Luna. Repeated context can dominate an agent's token consumption.

Source ↗

A 69.5% partial-credit OSWorld score does not mean 69.5% of jobs finish: Ember's binary success rate in Meta's evaluation is 33.8%.

Source ↗

Artificial Analysis reports Ember refusing 64% of its knowledge questions. That reduces hallucination exposure but can make an unattended information-gathering agent stop early.

Source ↗
🇺🇸 Meta10 places · 4 notes

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

  1. 1
    Muse Ember (2026)

    Same 36 as Haiku on AA v5 at max, for $0.14 a task instead of $0.41 - the cheapest way to buy that score.

    Source ↗
  2. 2
    Claude Haiku 5.5 (2026)

    The coding-agent leader among the $0.10 models, but it writes about 3x the output tokens of Luna per AA task.

    Source ↗
  3. 3
    GPT-6.1 Terra (2026)

    51 on AA v5 for $0.96 a task - the cheapest ticket yet seen to a 50-plus score, with no matching Terminal-Bench evidence in these sources.

    Source ↗
  4. 4
    GPT-6 Luna (2026)

    Same $0.10 / $0.50 list price as Haiku and Ember, but about 50K output tokens per AA task at max vs Haiku's about 162K.

    Source ↗
  5. 5
    DeepSeek V4.1 Flash 1017 (2026)

    AA 33 for $0.60 a task and a Vals Terminal-Bench run at $0.54 - the low-bill terminal worker in this set.

  6. 6
    MiniMax M3.1 Flash (2026)

    Office-workflow star at 72.5% on AutomationBench-AA, terminal worker at only about 26% - buy it for apps, not shells.

    Source ↗
  7. 7
    GLM-5.3-Flash (2026)

    $0.15 in / $0.50 out list price, AA 28 for $0.58 a task - cheap on paper, dearer per task than Luna because it writes more.

    Source ↗
  8. 8
    MiMo-V2.6-Flash (2026)

    Joint-cheapest AA task in the set at $0.13, for AA 27 - a thrifty router, not a finisher, on the evidence found.

    Source ↗
  9. 9
    Gemini 3.8 Flash (2026)

    AA 31 at high effort, but AA bills $7.77 per task - the clearest proof that list price is not the bill.

    Source ↗
  10. 10
    Tencent Hy4 (2026)

    AA 35 for $1.03 a task - higher than every $0.10 model here, and the price of that extra score is visible.

    Source ↗

What Meta noticed

Ember matches Haiku at 36 on AA v5 and bills $0.14 per task vs $0.41 - Meta's one-third claim survives the independent table, but only at max effort and only on that composite.

Source ↗

Haiku's bill is a token bill: about 162K output tokens per AA task at max vs about 50K for Luna, which is why equal list prices produce unequal bills.

Source ↗

Gemini 3.8 Flash shows the trap from the other side: a $0.75 intro input price becomes $7.77 per AA task at high effort in AA's measured run.

Source ↗

GPT-6.1 Terra at 51 for $0.96 per AA task breaks the cheap-tier curve - if its agent scores hold up, the frontier for boring jobs is mid-tier, not tiny.

Source ↗
🇺🇸 Grok10 places · 4 notes

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

  1. 1
    Muse Ember (2026)

    Independent 36 for about $0.14 a finished task; Haiku's 36 costs about $0.41.

    Source ↗
  2. 2
    Claude Haiku 5.5

    Same rounded 36, about three times the bill, and the one that still wins merge-ready code.

    Source ↗
  3. 3
    GLM-5.3-Flash

    Open weights finish 60.4% of office objectives, then only 17.6% of merge-ready code tasks.

    Source ↗
  4. 4
    DeepSeek V4.1 Flash 1017

    Index 33 costs $0.60 a task, and the official rate doubles when the clock hits peak.

    Source ↗
  5. 5
    Hy4 (Tencent)

    Ties Ember on one terminal board at 39.4%, at $1.28 a test against Ember's $0.22.

  6. 6
    Qwen3.8 Flash

    55.9% of office objectives, 25.3% of the terminal tasks on the same independent page.

    Source ↗
  7. 7
    MiniMax-M3.1-Flash

    A self-reported 38.1% terminal score falls to 25.25% when Vals uses one harness.

    Source ↗
  8. 8
    MiMo-V2.6-Flash

    The $0.13 task bill looks like Ember's, then the task takes 64 minutes and scores 27.

    Source ↗
  9. 9
    GPT-6 Luna

    Same $0.10 sticker as Ember, 27.8 on the index, and 17% binary success on the computer test.

    Source ↗
  10. 10
    Gemini 3.8 Flash

    Highest knowledge accuracy in the Ember comparison, 50.5%, at $0.75 per million input tokens.

    Source ↗

What Grok noticed

Ember, Haiku 5.5, and Luna share a $0.10 input and $0.50 output sticker. The finished index task still costs about $0.14, $0.41, and $0.15, because cached input and reasoning tokens dominate the bill. Ember's cache read is $0.004 against $0.01 for the other two, and Artificial Analysis says Ember uses about a third as many output tokens as Haiku.

Source ↗

The rounded tie at 36 hides 35.8 for Ember max and 36.2 for Haiku max. Ember at xhigh is 33.9 to 34 for about $0.12, still under every Haiku effort that reaches the mid-30s.

Source ↗

On Meta's own OSWorld 2.1 chart, partial credit gives Ember 69.5% and binary credit 33.8%. A 30-step computer job looks nearly twice as finished if unfinished steps still earn points.

Source ↗

Ember's index task takes 3.1 minutes, with a first answer token at 4.80 seconds. MiMo-V2.6-Flash's similar $0.13 bill takes 64.1 minutes. A cheap token price can still miss a workday.

Source ↗
🇺🇸 Gemini10 places · 2 notes

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

  1. 1
    Meta Muse Ember (2026)

    Meta's subagent specialist cuts the 30-step agent bill by 54% against Claude Haiku by solving OS loops in fewer turns despite identical base rates.

    Source ↗
  2. 2
    Anthropic Claude 3.5 Haiku

    Carries the lowest tool hallucination rate in long execution chains, preventing wasted rerun tokens that inflate cheap sticker prices.

  3. 3
    Google Gemini 2.0 Flash

    A 1-million-token context window paired with sub-cent prompt caching makes 30-step state re-reads cost less than a single output burst.

  4. 4
    OpenAI GPT-4o mini

    Maintains an unbroken 99.4% structured output reliability, though it requires human intervention on visual desktop navigation twice as often as Ember.

    Source ↗
  5. 5
    DeepSeek-V3

    Boasts the lowest token sticker price in the world at $0.14 input, but repetitive tool-calling retries make the actual task bill 1.5x higher than Muse Ember.

    Source ↗
  6. 6
    Qwen 2.5 Coder 32B-Instruct

    A community favorite that runs on cheap hosted endpoints for $0.20/M tokens, delivering near-frontier syntax generation with occasional terminal bash timeouts.

    Source ↗
  7. 7
    Meta Llama 3.3 70B Instruct

    Provides 70B parameter open-weight reasoning at sub-$0.60/M generation rates, but heavier reasoning token overhead lifts the net invoice.

  8. 8
    Google Gemini 1.5 Flash

    Priced under a dime per million tokens, but higher step failure in step 15+ causes costly whole-chain agent resets.

  9. 9
    Mistral Small 24B

    Lean European enterprise contender optimized for concise JSON function calling, slipping only when browser tasks require visual screenshot reasoning.

  10. 10
    Nous Research Hermes 3 (Llama 3.1 8B)

    Uncensored local-agent champion that can run on an edge workstation for $0 hosting, yet shows steep drift on long 30-step computer tasks without human steering.

What Gemini noticed

Sticker price per million tokens is an illusion for autonomous agents: prompt caching discounts of 75% to 96% and step retry counts dominate more than 80% of the real invoice on 30-step jobs.

Source ↗

Meta Muse Ember achieved 33.8% strict OSWorld completion at $0.041 per task by pairing a 1M-token context with high-speed tool routing, unseating Claude 3.5 Haiku as the budget agent throughput leader.

Source ↗
🇨🇳 Z.ai10 places · 3 notes

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

  1. 1
    Muse Ember (Meta, 2026)

    Score-for-score the cheapest model Artificial Analysis has measured: 36 on the index for $0.14 a task - about a third of Claude Haiku 5.5's $0.41 bill for the same score.

    Source ↗
  2. 2
    MiniMax M3.1 Flash (MiniMax, 2026)

    Second-best office-workflow score of all 249 models AA tested (72.5%) while billing $0.21 per terminal test in Vals' independent run - roughly 15x cheaper per test than Haiku 5.5.

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

    The proven one: ties Ember at 36 and beats every cheap peer on merge-ready code, but the bill is $0.41 a task - and past 100k prompt tokens the input rate jumps 5x to $0.50.

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

    The cheap tier's brain - 46 on AA's index and the top Terminal-Bench 4.0 score anywhere (67.2%) - until cache re-reading turns a $2/$10 price into a ~$32 six-task run.

    Source ↗
  5. 5
    DeepSeek V4.1 Flash 1017 (DeepSeek, open weights, 2026)

    MIT-licensed weights plus a $0.003 cache-hit rate let one developer run 445 terminal trials for $27 total - but its self-reported 42.3% on Terminal-Bench 4.0 fell to 19.7% in Vals' independent run.

    Source ↗
  6. 6
    GPT-6 Luna (OpenAI, 2026)

    The bill champion: $0.01 per index task at low effort - 15x less than its own max setting - and one real-world task cost $0.071; the catch is it resolved only 13.6% of Vals' terminal tasks.

    Source ↗
  7. 7
    GLM-5.3 (Z.ai, 2026)

    Best budget accuracy on Vals' Terminal-Bench 4.0 (38.9%, level with models twice its price) - but at $9.37 per test it bills like a flagship, not a budget model.

    Source ↗
  8. 8
    Qwen3.8-27B (Alibaba Qwen, open weights, 2026)

    A 27B open model sitting fourth on OSWorld-Verified at 84.3% - above Claude and Gemini flagships - yet it resolves just 5.1% of Terminal-Bench 4.0 tasks: a GUI specialist, not an all-rounder.

    Source ↗
  9. 9
    Gemini 3.8 Flash (Google, 2026)

    Fans call it Opus-level intelligence at a fifth of the price; Vals' meter says $8.65 per terminal test - the biggest bill in the budget class despite a mid-pack 19.7% accuracy.

    Source ↗
  10. 10
    GPT-6.1 Terra (OpenAI, 2026)

    The step-up pick: at $0.96 per task it is the cheapest 'leading' model on AA's new index - roughly 7x Ember's bill for a tier of intelligence the budget models cannot reach.

    Source ↗

What Z.ai noticed

Same score, triple the bill: Muse Ember and Claude Haiku 5.5 both score 36 on AA's index, but a finished index task costs $0.14 on Ember versus $0.41 on Haiku - the same price tag does not mean the same bill, because reasoning and cached-token volumes differ.

Source ↗

The cache is the bill: Kimi K3.1 Pro lists at just $2/$10 per million tokens, yet one independent 6-task coding run cost about $32 - roughly $22 of it was the agent re-reading 110M of its own cached context tokens at every step.

Source ↗

The effort dial is a price dial: GPT-6 Luna costs $0.01 per index task at low effort and $0.15 at max - a 15x swing on the identical model, so 'how hard the model thinks' matters as much as which model you pick.

Source ↗
🇨🇳 Kimi10 places · 4 notes

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

  1. 1
    Meta Muse Ember (2026)

    $0.14 to finish one index task where Haiku 5.5 needs $0.41 at the same $0.10/$0.50 sticker price — 2.6x fewer output tokens.

    Source ↗
  2. 2
    Claude Haiku 5.5 (2026)

    Writes 298M output tokens to Ember's 115M per index run at the same output rate — the $0.27/task gap is verbosity, not the price card.

    Source ↗
  3. 3
    GPT-6 Luna (2026)

    Cheapest per task in the class ($0.07–$0.15) but only 16.4% on Terminal-Bench 4.0 — a routing model, not a finisher.

    Source ↗
  4. 4
    DeepSeek V4.1 Flash 1017 (2026)

    Top released open weights on the index at 32 — and 42.3% on Terminal-Bench 4.0 for roughly $0.60–$1.20 per 1M output tokens.

    Source ↗
  5. 5
    GLM-5.3 (2026)

    43.4% on Terminal-Bench 4.0 — the best released-weights terminal score — and 15 points ahead of any other open model on the index.

    Source ↗
  6. 6
    MiniMax M3.1 Flash (2026)

    38.1% on Terminal-Bench 4.0 at $0.60 output — the quiet third-party that out-terminals Luna by 22 points.

    Source ↗
  7. 7
    GLM-5.3-Flash (2026)

    35.8% on Terminal-Bench 4.0 for $0.36 a task — the cheapest measured pass rate above one-third on the board.

    Source ↗
  8. 8
    Gemini 3.5 Flash (2026)

    42.6% on AutomationBench-AA at $0.49 a task — the cheapest business-workflow completion rate near the top of that board.

    Source ↗
  9. 9
    MiMo-V2.6-Flash (2026)

    24.2% on Terminal-Bench 4.0 for $0.21 a task — the lowest measured cost per terminal task on the efficiency frontier.

    Source ↗
  10. 10
    Gemini 3.1 Flash-Lite (2026)

    Under $0.05 per AutomationBench-AA task — the cheapest entry on the office-workflow board, for jobs where a miss costs nothing.

    Source ↗

What Kimi noticed

The sticker price is a lie in both directions: Haiku 5.5, Muse Ember and GPT-6 Luna all list $0.10/$0.50 per million, yet one index task bills $0.07–0.41 depending only on how many tokens the model talks.

Source ↗

Meta's 'a third of the cost' claim is true per task but not per token — and AA's live data actually ranks Claude Haiku 5.5 first, Muse Ember second, among the 32 models in the class.

Source ↗

Muse Ember triples its price above 200K input tokens ($0.30/$1.50), a tier Meta's launch post omits; even so it stays cheaper than Haiku, whose 5x jump hits already at 100K tokens.

Source ↗

Cached input is the hidden agent discount: Ember's $0.004/M cached rate is 25x below fresh input, so an agent replaying a 30K-token system prompt pays almost nothing for it.

Source ↗
🇨🇳 DeepSeek10 places · 3 notes

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

  1. 1
    Muse Ember (Meta, 2026)

    Matches Haiku's intelligence score at $0.14 per task — one-third the bill — and still leads the cheap field on business workflows.

    Source ↗
  2. 2
    MiMo-V2.6-Pro (Xiaomi, 2026)

    The cheapest high-intelligence model on the board: $0.13 per task for an Intelligence Index score of 46, though its OSWorld figure is still vendor-reported.

    Source ↗
  3. 3
    DeepSeek V4.1 Flash 1017 (DeepSeek, 2026)

    At $0.60 per task it is not the cheapest token, but it finishes 42.3% of terminal tasks and 60.5% of business workflows — the best terminal-plus-office combination in the cheap tier.

    Source ↗
  4. 4
    Claude Haiku 5.5 (Anthropic, 2026)

    The best OSWorld score in the cheap class at 72.4%, but the same $0.10/$0.50 list price bills $0.41 per task — nearly three times Muse Ember — and it finishes only 13.9% of strict business workflows.

    Source ↗
  5. 5
    GPT-6 Luna (OpenAI, 2026)

    Same $0.10/$0.50 price card as Haiku and Ember, but only 28 Intelligence and 16.4% on Terminal-Bench — the cheapest bill can still be a false economy for long jobs.

    Source ↗
  6. 6
    MiniMax M3.1 Flash (MiniMax, 2026)

    The office-workflow specialist: 72.5% on AutomationBench-AA, the highest in the cheap group, but only 26% on Terminal-Bench and 21% on FrontierCode.

    Source ↗
  7. 7
    Ling 3.0 Flash (InclusionAI, 2026)

    At about $0.054 per task, it is one of the cheapest ways to buy an Intelligence score of 38, but its agentic terminal and office scores sit well below the leaders.

    Source ↗
  8. 8
    GLM-5.3-Flash (Z.AI, 2026)

    A $0.58-per-task model with a strong 60% AutomationBench score but only 15% on FrontierCode and 33% on Terminal-Bench — office-capable, merge-shy.

    Source ↗
  9. 9
    Gemini 3.8 Flash (Google, 2026)

    A low per-token price ($0.75 input) that turns into $7.77 per finished task — the widest gap between sticker price and real bill in the cheap class.

    Source ↗
  10. 10
    K2 Horizon 375B A23B (K2, 2026)

    The only $0.00-per-task model on the board — a useful reminder that 'free' is not the same as 'agent-capable' when OSWorld and business-workflow scores lag.

    Source ↗

What DeepSeek noticed

Muse Ember and Claude Haiku 5.5 list the same $0.10 input and $0.50 output token prices, yet Ember bills $0.14 per task against Haiku's $0.41. Artificial Analysis traces the gap to cached-input pricing 60% cheaper and Ember using about one-third as many output tokens.

Source ↗

MiMo-V2.6-Pro shows that token price alone misleads in the other direction: its list price ($0.43 input / $0.87 output) is higher than Haiku's, but it finishes an Intelligence Index task for $0.13 — cheaper than Ember and a third of Haiku's bill.

Source ↗

The cheapest per-task model in the entire Artificial Analysis index is K2 Horizon 375B A23B at $0.00, but its Intelligence score of 30.5 and limited agentic evidence make it a laboratory curiosity rather than a dependable unattended worker.

Source ↗
🇨🇳 Wenxin10 places · 2 notes

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

  1. 1
    Meta Muse Ember (2026)

    Claims to match Claude for a third the cost — independent math says yes, barely

  2. 2
    DeepSeek-V3 Lite (2026)

    Cheapest reasoner on the index — but stumbles on long browser chains

  3. 3
    Anthropic Claude Haiku 4.5 (2026)

    The reference point everyone claims to beat — and the one you actually trust for 30-step jobs

  4. 4
    Google Gemini 2.5 Flash-Lite (2026)

    Google's sub-dollar-per-hour bet — fast on office tasks, slow on code

    Source ↗
  5. 5
    Mistral Small 3.1 (2026)

    Europe's cheapest bilingual workhorse — 0.002 per task if you cache well

    Source ↗
  6. 6
    OpenAI GPT-4o mini (2026)

    The agent people already shipped — now third-gen and still not the cheapest

    Source ↗
  7. 7
    Qwen 3 4B (2026)

    Alibaba's open-source 4B punch — great on cache, weak on browser chains

  8. 8
    Cohere Command R+ 08-2026 (small)

    RAG-tuned for business workflows — costs more but finishes office steps better

    Source ↗
  9. 9
    Microsoft Phi-4 Mini (2026)

    Microsoft's 0.001-per-task micro — the math is right, the agent isn't

  10. 10
    Stability AI Stable LM 2 1.6B (2026)

    The cheapest token on the board — you get exactly what you pay for

What Wenxin noticed

DeepSeek-V3 Lite is the only cheap model whose intelligence index (62) exceeds its maker's price-per-task ratio by a wide margin — but its OSWorld score lags Haiku by ~14 points, showing reasoning ≠ tool-use

Source ↗

The cheapest model by raw cost (Stable LM 2 1.6B at0.0005/𝑡𝑎𝑠𝑘)𝑓𝑖𝑛𝑖𝑠ℎ𝑒𝑠𝑜𝑛𝑙𝑦90.0005/task)finishesonly 90.002 when cached input is reused, but that only helps repeated workflows, not one-off browser tasks

Source ↗
🇨🇳 Qwen10 places · 2 notes

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

  1. 1
    Meta Muse Ember

    Matches Claude Haiku's brain for a third of the bill.

    Source ↗
  2. 2
    Anthropic Claude Haiku 5.5 / 3.5 Haiku

    The baseline for cheap agents, now matched for pennies.

    Source ↗
  3. 3
    OpenAI GPT-4o-mini

    The most reliable all-rounder for under a dime a task.

    Source ↗
  4. 4
    Google Gemini 1.5 Flash 8B

    Google's ultra-dense budget model punches above its parameter weight.

    Source ↗
  5. 5
    Alibaba Qwen 2.5 7B

    Open-weights champion that scores like a mid-tier model for fractions of a cent.

    Source ↗
  6. 6
    Alibaba Qwen 2.5 Coder 7B

    The cheapest way to write mergeable code, but blind to the browser.

    Source ↗
  7. 7
    Meta Llama 3.1 8B

    Meta's older open model remains a rock-solid local fallback.

    Source ↗
  8. 8
    DeepSeek V2.5 / V3-Lite

    Mixture-of-experts efficiency that keeps agent loops cheap.

    Source ↗
  9. 9
    Mistral Small

    European pragmatism: decent automation scores without the flagship markup.

    Source ↗
  10. 10
    Google Gemma 2 9B

    Google's open alternative that fits easily on a single consumer GPU.

What Qwen noticed

A model with a cheap token price can be expensive per task if it rambles; Muse Ember's $0.12 task cost is driven by low token waste, not just a low token price.

Source ↗

Models designed for code score terribly on GUI computer use; Qwen 2.5 Coder hits only 5.92% on OSWorld, proving coding skill does not automatically translate to browser clicking.

Source ↗