Back to the trend of the day: New Apple CEO says it won't ship AI for AI's sake
Poly Rating · from the trend of the day

Best AI Model for Agents 2026: Top 10 by Tests, Votes, Use

DeepSeek V4.1 Flash ran 45.5 trillion OpenRouter tokens in the week to 30 October, about six times as many as Claude Opus 5.5, the model blind voters rank first on agent tasks.

  • No. 1: Claude Opus 5.5
  • 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 Opus 5.5

    It is first on LMArena's blind agent board (20.9) and on WebDev. Its 89.9% on SWE-bench Pro is vendor-reported, not from a common harness. It is sixth in OpenRouter tokens (7.31T).

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

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

    What each AI noted 7
    • Claudeplace 2

      It is listed first on LMArena's WebDev board as of October 17, 2026, which makes it the practical coding pick.

      Source ↗
    • ChatGPTplace 1

      26.6 trillion OpenRouter tokens in 30 days: a premium model with substantial observed demand despite cheaper alternatives.

      Source ↗
    • Metaplace 1

      89.9% on SWE-bench Pro – 8.6 points above Sonnet 5.5 – yet only 6.91T tokens on OpenRouter vs 47.8T for DeepSeek, frontier quality at $4 input/$20 output

    • Grokplace 2

      Blind voters put Opus 5.5 first on agent tasks at 20.9, and a long coding session still costs about $1.39.

      Source ↗
    • Kimiplace 1

      The agentic triple crown at half the old flagship's price: 89.9% SWE-bench Pro, #1 Arena WebDev (1807), #1 Agent Arena (21.4% confirmed success) — and the first Anthropic flagship that costs $4/$20 instead of $10/$50.

    • DeepSeekplace 1

      Claude Opus 5.5 tops the vendor-aggregate SWE-bench Pro board at 89.9% and the Hermes agent bench at 76.7% — the most complete “do-the-task” agent model of October 2026, at $4.19–$5.98 per successful task.

      Source ↗
    • Qwenplace 1

      89.9% on SWE-bench Pro leads the market, proving it resolves the most real GitHub issues.

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

    DeepSeek V4.1 Flash

    This is the model agents actually run: 45.5T OpenRouter tokens a week, Hermes Agent's top model, and about $0.05 for a long coding session. It trails the leaders on hard tests.

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

    In their lists: Claude: 5 · ChatGPT: 5 · Meta: 4 · Grok: 1 · Kimi: 7 · DeepSeek: 5 · Qwen: 5

    What each AI noted 7
    • Claudeplace 5

      It was still first on OpenRouter at 32.8T tokens in the report dated October 20, 2026, though its usage fell 20%.

    • ChatGPTplace 5

      116 trillion tokens in 30 days, with published inference prices starting at $0.13 input and $0.52 output per million tokens.

      Source ↗
    • Metaplace 4

      47.8T tokens last 7 days – 29.4% of top 20 – 1017 snapshot alone, 43.6% of top 20 are DeepSeek models, $0.30 input $1.2 output

    • Grokplace 1

      Hermes Agent put 11.8 trillion tokens through DeepSeek V4.1 Flash this month, and a long coding session on it costs about $0.047.

      Source ↗
    • Kimiplace 7

      The token-economy king: DeepSeek is the #1 provider by live agent token share (39.1%) — more than Anthropic and Google combined — while this model lists at $0.15/$0.60 off-peak.

      Source ↗
    • DeepSeekplace 5

      DeepSeek V4 Flash is the real-world agent volume king: 11.4 trillion weekly tokens on OpenRouter and a $0.028 cost per successful task — roughly 180× cheaper than Claude Fable 5.1, though it trails the frontier on hard coding benchmarks.

      Source ↗
    • Qwenplace 5

      53.1T tokens processed on OpenRouter, making it the real workload king for developers routing agent traffic.

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

    Gemini 4 Argon

    It sits near the top of both the test and vote boards: 99.22% on SWE-Bench Pro V2 and about 1520 on LMArena text. Yet it runs only 1.09T OpenRouter tokens a week, rank 28.

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

    In their lists: Claude: 3 · ChatGPT: 8 · Meta: 8 · Grok: 8 · Kimi: 5 · DeepSeek: 2 · Qwen: 6

    What each AI noted 7
    • Claudeplace 3

      It leads LMArena Text on human votes, but the 1524±9 score is preliminary.

      Source ↗
    • ChatGPTplace 8

      A 50.4 intelligence-index score keeps Google's frontier model competitive, but general capability alone cannot settle agent reliability.

      Source ↗
    • Metaplace 8

      1524 Elo #2 on LMArena – within 4 points of Carbon – High reasoning tier, $10/M output, $2/$10 intro pricing until Nov 30

      Source ↗
    • Grokplace 8

      Gemini 4 Argon sits at 99.22% on the public Pro V2 board, 0.18 points behind Opus 5, on a test Scale says is near the ceiling.

      Source ↗
    • Kimiplace 5

      The human favorite that loses the lab tests: #1 LMArena text (1517.2) and the lowest hallucination rate measured (4%), yet 50.4 on Intelligence Index v5 — sixth among usable models.

      Source ↗
    • DeepSeekplace 2

      Gemini 4 Argon is the people’s choice — No. 1 on LMArena text at 1517 and multi-turn at 1544 — and it does it at $1.80 per successful task, roughly a quarter of Claude Opus 5.5’s cost.

      Source ↗
    • Qwenplace 6

      99.2% on SWE-bench Pro V2 shows Google's frontier model is catching up to Anthropic in long-horizon engineering.

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

    Kimi K3 (K3.1 Pro)

    Kimi K3 solved 214 of 272 private SWE-Bench Pro V2 tasks, two behind Opus 5, and K3.1 Pro leads an independent Terminal-Bench 4.0 run at 67.2%. It uses under 2T tokens a week.

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

    In their lists: Claude: 9 · ChatGPT: 9 · Meta: 9 · Grok: 7 · Kimi: 4 · DeepSeek: 10

    What each AI noted 6
    • Claudeplace 9

      Sources name it the strongest open-weight choice for agentic work in 2026. I could not confirm a score.

      Source ↗
    • ChatGPTplace 9

      A 1529 human-preference rating in an IT-focused arena puts Kimi close to Claude Fable 5's 1540, despite a very different market profile.

      Source ↗
    • Metaplace 9

      Best open weights at #18 LMArena 1488 – K3.1 Pro $3/$15 – 1.97T tokens, #1 Arena full-stack 1664 per Turiloop budget pick

      Source ↗
    • Grokplace 7

      On the private SWE-Bench Pro V2 count, Kimi K3 solved 214 of 272 tasks, two behind Opus 5.

      Source ↗
    • Kimiplace 4

      The open-weights surprise that beats the giants at their own agentic test: it leads the independent Terminal-Bench 4.0 run (67.2%, Oct 23) ahead of every closed model except cyber-locked Gemini 4 Carbon — at $2/$10.

      Source ↗
    • DeepSeekplace 10

      Kimi K3 is built for orchestration: it scales to 300 sub-agents and 4,000 coordination steps, making it a specialist pick when the agent’s job is to manage other agents rather than write every line itself.

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

    Claude Sonnet 5.5

    It beat Opus 5.5 on OpenRouter's flight-booking agent test, 78.7% to 75.5%, at half the price. It scores 81.3% on the vendor-reported SWE-bench Pro board.

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

    In their lists: Meta: 2 · Grok: 9 · Kimi: 2 · DeepSeek: 6 · Qwen: 2

    What each AI noted 5
    • Metaplace 2

      81.3% SWE-bench Pro – 8.6 points behind Opus 5.5 at half price $2/$10 – 2.61T tokens, best balance for everyday agents

    • Grokplace 9

      On OpenRouter's own flight-booking agent test, Sonnet 5.5 scored 78.7%, above Opus 5.5 at 75.5%, at half the token price.

      Source ↗
    • Kimiplace 2

      The model developers actually burn tokens on: it tops Terminal-Bench 4.0 (70.6%) and LiveBench Coding (91.4) at half of Opus's price — the highest agentic score per dollar in the frontier tier.

      Source ↗
    • DeepSeekplace 6

      Claude Sonnet 5.5 gives you 81.3% of Claude Opus 5.5’s SWE-bench Pro score at roughly 84% of its per-task cost — the value pick inside the Claude family for production agent loops.

      Source ↗
    • Qwenplace 2

      The best all-round model for Hermes Agent, jumping to 70.6% on Terminal-Bench 4.0 for long tasks.

    5 AIs put it here
    Source ↗
  6. 6

    GLM 5.3 (and GLM 5.3 Flash)

    GLM 5.3 Flash is second in weekly OpenRouter tokens (16.1T). The full GLM-5.3 cleared 211 of 272 private SWE-Bench Pro V2 tasks, at $0.36 per long session.

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

    In their lists: Claude: 8 · ChatGPT: 6 · Meta: 6 · Grok: 3 · Kimi: 9

    What each AI noted 5
    • Claudeplace 8

      It held flat at 12.2T tokens on OpenRouter while DeepSeek fell, and it is the cheap-tier model agent builders are choosing.

    • ChatGPTplace 6

      55.3 trillion monthly tokens put GLM's economical Flash tier ahead of many more celebrated frontier models in observed API volume.

      Source ↗
    • Metaplace 6

      15.7T tokens #2 on OpenRouter – $0.5/M output – open-weight surge from 27% to 72% of top 50 tokens year-over-year

    • Grokplace 3

      The Flash edition was second in the week's traffic at 16.1 trillion tokens, while the full GLM-5.3 cleared 211 of 272 private SWE-Bench Pro V2 tasks.

      Source ↗
    • Kimiplace 9

      The self-hostable workhorse: 95.4% on the archived SWE-bench Verified board — above Kimi K3 (93.4%) — under a plain MIT license, for $1.19/$6 hosted.

    5 AIs put it here
    Source ↗
  7. 7

    Claude Fable 5.1

    It leads LiveBench overall at 88.0 and scores 81.2% on the vendor-reported SWE-bench Pro board. It is also the priciest per task here: $5.10 on LiveBench, $7.63 at max effort.

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

    In their lists: ChatGPT: 3 · Meta: 7 · DeepSeek: 3 · Qwen: 4

    What each AI noted 4
    • ChatGPTplace 3

      64.5% resolution in SWE-rebench at $4.40 per problem: a strong completion result with a meaningful cost premium.

    • Metaplace 7

      81.2% SWE-bench Pro – only 0.1 point behind Sonnet 5.5 – 95.0% on SWE-bench Verified #1, but $7.63 per task vs Astra $3.26

    • DeepSeekplace 3

      Claude Fable 5.1 is LiveBench’s overall leader at 88.0 and scores a perfect 100 on browser use — the best “computer-use” agent model — but at $5.10 per successful task it is the most expensive in this top group.

    • Qwenplace 4

      Hits 81.2% on SWE-bench Pro, tying closely with Sonnet 5.5 but at a lower cost per completed task.

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

    GPT-6 Astra

    It is first on the Frontend Code Arena (1793) and scores 82.4 on LiveBench's agent tasks. But it costs $10/$50, and OpenAI holds only 7.8% of tracked live tokens.

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

    In their lists: ChatGPT: 4 · Kimi: 3 · DeepSeek: 4

    What each AI noted 3
    • ChatGPTplace 4

      Its 56.9 intelligence-index result exceeds GPT-6.1 Sol's 54.5, but the stronger general score does not prove cheaper agent execution.

      Source ↗
    • Kimiplace 3

      Human-vote darling, token-bill villain: #1 Frontend Code Arena (1793) and #2 WebDev (1774), but at $10/$50 it is 2.5x Opus 5.5's price — and OpenAI's share of live agent tokens is just 7.8%.

    • DeepSeekplace 4

      GPT-6 Astra is the balanced frontier agent: 86.1 LiveBench overall, 82.4 agentic, and $3.81 per successful task — stronger than Gemini 4 Argon on agentic tasks and cheaper than Claude Fable 5.1.

    3 AIs put it here
    Source ↗
  9. 9

    Muse Spark 1.1 (Meta)

    It tops Scale's standardized SWE-bench Pro board at 61.5% ± 3.1, the score from one common harness. No Muse model is in the weekly token top ten.

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

    In their lists: Meta: 10 · Grok: 10 · DeepSeek: 7

    What each AI noted 3
    • Metaplace 10

      61.5% ±3.1 on Scale official public leaderboard – highest standardized score – $0.2/M tokens cheapest in top 20, Contributor tier

      Source ↗
    • Grokplace 10

      Muse Spark 1.1 leads Scale's older private SWE-Bench Pro board at 61.5%, and it is not in the week's token top ten.

      Source ↗
    • DeepSeekplace 7

      Muse Spark 1.1 owns the only standardized top score on SWE-bench Pro — 61.5% on Scale’s public set — but its 51.5% commercial-set score shows how much scaffolding and data splits can move the number.

    3 AIs put it here
    Source ↗
  10. 10

    GPT-6 Luna

    It has the cheapest long coding session in OpenRouter's comparison, $0.024, and is fourth in weekly tokens at 10.9T, up 62%. No SWE-bench Pro row has been published.

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

    In their lists: Meta: 5 · Grok: 4

    What each AI noted 2
    • Metaplace 5

      10.3T tokens #4 on OpenRouter – cheapest frontier at $0.5/M output – $0.07 per agentic task on OSWorld vs $14.19 for Fable 5.1

      Source ↗
    • Grokplace 4

      A long coding session on GPT-6 Luna costs about $0.024, the cheapest in OpenRouter's comparison, on 10.9 trillion tokens in a week.

      Source ↗
    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.

The places are ranked on a mix of tests, votes and use. Scores come from different boards, and some are vendor-reported, so they are not directly comparable.

1Claude Opus 5.5

No. 1 — Claude Opus 5.5. On the lists of 7 of 9 AIs; median place there: 1. Blind voters put it first on agent tasks, and it leads the coding boards, even though far fewer tokens run through it than through DeepSeek.

Did you know

  • With Opus 5.5, extended thinking cannot be switched off. Most cybersecurity tasks are also sent to an older model, Opus 4.8, unless the user is in Anthropic's Cyber Verification Program.

    Source ↗
What it is
Anthropic's top model, priced at $4 input and $20 output per million tokens.
Its edge
It ranks first with blind voters on agent tasks, ahead of the other two.
Who it suits
Suits teams that want top agent quality; not those on a tight budget.
  • OpenRouter tokens per week7.31T tokens a week
Why it is here

  • The model card behind this rating lists Claude Opus 5.5 at 89.9% on SWE-bench Pro, the coding test. That is a vendor-reported score, not one from a shared test setup.

    Source ↗

  • On Terminal-Bench 4.0, a test of working in a command line, Claude Opus 5.5 scored 66.4%. Gemini 4 Argon, the third-place model, scored 57.4%. So the top pick leads the third pick on hard agent tasks.

    Source ↗
Did you know

  • Anthropic says Opus 5.5 costs about 40% less to run than Opus 5 at default settings. The top-voted model got cheaper, not pricier. It came out on September 22, 2026.

    Source ↗
Keep in mind

  • Max effort mode is warned to use "6x or more" usage. Redditors report 4x, 5.5x and 6x, and one says Max tends to overthink. Some users also expect the model to get worse soon after launch.

    Source ↗

  • One user said Opus 5.5 made questionable architectural decisions on a feature task. Another report says Fable 5.1 still wins some side-by-side coding tests.

    Source ↗
Where to go further

  • Read Anthropic's own announcement for Claude Opus 5.5. Start with the benchmark section, then the note that the gap to Fable 5.1 is narrower in real use than the scores suggest.

    Source ↗
Compare the top three ↓

2DeepSeek V4.1 Flash

No. 2 — DeepSeek V4.1 Flash. On the lists of 7 of 9 AIs; median place there: 5. It is the model agents actually run: the most tokens by far, at a tiny cost per session, though it trails on hard tests.

Did you know

  • This cheap 'Flash' model can read images as well as text. OpenRouter lists a context window of 1,048,576 tokens and up to 384,000 tokens of output in a single answer.

    Source ↗
What it is
DeepSeek's cheap, fast 'Flash' model that works with images and text.
Its edge
It handles far more real agent traffic than the other two, at far lower cost.
Who it suits
Suits high-volume, low-budget agents; not tasks needing the hardest-test accuracy.
  • OpenRouter tokens per week45.5T tokens a week
Why it is here

  • On cost, the same tracker puts an average coding session at $0.0477, for about 4.4M tokens. The list price is $0.04 per million input tokens, which is why agents choose it.

    Source ↗
History

  • DeepSeek had quadrupled its prices in August 2026, then reversed course. On September 10 it released V4.1 Flash and cut prices, with Bloomberg Intelligence estimating up to 32%. From September 14, every V4-Pro request went to Flash.

    Source ↗
Keep in mind

  • The reported agent gains come from DeepSeek's own numbers, run with its own harness. The reviewer says independent tests are still needed. Tools and company support also lag the big labs.

    Source ↗

  • Price is disputed. The page lists a 2x peak-hour multiplier. A Reddit user said changing 10 lines used 28% of their quota, though that post is unverified.

    Source ↗
Where to go further

  • Open the OpenRouter model page and check the price per million tokens. Then compare it with a long coding task of your own, since cheap runs are why agents use this model.

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

  • Grok: Hermes Agent put 11.8 trillion tokens through DeepSeek V4.1 Flash this month, and a long coding session on it costs about $0.047.

Compare the top three ↓

3Gemini 4 Argon

No. 3 — Gemini 4 Argon. On the lists of 7 of 9 AIs; median place there: 6. It scores near the top on tests and human votes, but few people run it yet, which holds its rank down.

Did you know

  • Gemini 4 Argon has a limited-time launch price that ends on November 30. After that date it may cost more, which could change how many developers try it.

    Source ↗
What it is
Google's agent-focused model with a context window of up to 1 million tokens.
Its edge
It has the strongest human-vote and test scores of the three, but the least real use.
Who it suits
Suits those who want a Google-based agent; not anyone who needs open access today.
  • OpenRouter tokens per week1.09T tokens a week
Why it is here

  • On DeepSWE v1.1, a software-fixing test, Gemini 4 Argon scored 77.9%. Claude Opus 5.5 scored 74.2%. So Argon beats the first-place model on this test, which helps explain its high rank.

    Source ↗

  • On the FrontierSWE v2 coding test, Gemini 4 Argon scored 54.96 and ranked 6th of 23 models. That is strong, but it is not first.

    Source ↗
Did you know

  • Google's own developer page says Argon is built for multi-step autonomous agents and can handle up to 1 million tokens of context. That is roughly a few long books at once.

    Source ↗
History

  • Google had promised Gemini 3.5 Pro for June 2026, then dropped it after months of delays. Several Gemini team leaders left in between. Argon was announced on September 30, 2026, and the article says the delays left Google behind Anthropic and OpenAI.

    Source ↗
Keep in mind

  • Not open to everyone. The reviewer's API calls on October 1 returned 404, and access ran through Google's Fairwind Program for vetted cyber defenders. Other sources say it later opened up, so reports conflict.

    Source ↗
Where to go further

  • Read the Gemini API page for this model. Then try a multi-step coding task with thinking set to high. The page links to Google AI Studio, where you can run it.

    Source ↗
Compare the top three ↓

The top three side by side

Claude Opus 5.5DeepSeek V4.1 FlashGemini 4 Argon
In the AIs' lists7 / 97 / 97 / 9
OpenRouter tokens per week7.31T tokens a week45.5T tokens a week1.09T tokens a week
Put first byChatGPT, Meta, Kimi, DeepSeek, QwenGrok—

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.

Agent apps don't run the model that wins the votes. Hermes Agent's top model this month is DeepSeek V4.1 Flash at 11.8T tokens, while Opus 5.5, first on the agent vote, is sixth in weekly tokens.

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

"SWE-bench Pro" is several boards. The common harness tops out at 61.5% (Muse Spark 1.1) and vendor reports at 89.9% (Opus 5.5). Pro V2 has Opus 5 at 99.4% on public tasks but 222 of 272 on private ones.

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

Cost depends on who measures it. A long coding session costs $0.024 on GPT-6 Luna, $0.047 on DeepSeek and $1.39 on Opus 5.5. Other indexes put Opus 5.5 at $4.19 to $13.43 per task.

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

A Terminal-Bench score is meaningless without its version: the same Gemini 3.8 Flash scores 89.4% on version 2.1 and 19.1% on 4.0.

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

04Questions people ask

AI models vs AI agents?

A model writes the next answer. An agent wraps a model in tools, memory and a loop until the job is done, and Hermes Agent alone has called 547 different models on OpenRouter.

Source ↗
What is the best AI model for agents?

Claude Opus 5.5 leads blind agent votes, at 20.9 on LMArena's agent board. DeepSeek V4.1 Flash leads real use, at 45.5T tokens a week and about $0.05 per long coding session.

Source ↗
What is the best AI model for Hermes Agent?

By actual use it is DeepSeek V4.1 Flash, with 11.8T of Hermes Agent's tokens this month, followed by GLM 5.3 Flash at 3.91T.

Source ↗
Why do SWE-bench Pro scores differ so much?

Different harnesses and task splits: 61.5% is the top score under Scale's common harness, while 89.9% is a vendor-reported figure. Most pages never say which one they quote.

05How we counted

In plain words, no formulas.

1The question

Top 10 AI Models for Running Agents: Test Scores Against Human Votes Against What Developers Actually Use

2Why this topic

Three independent sources that rarely agree, a lab test, a human vote and real usage, sit side by side, so the reader sees which model wins on paper and which one people actually run agents on.

3Where it applies

global, AI agents and orchestration of several AIs

4What we looked at

  • Agentic coding score on SWE-bench Pro
  • Human preference rating on LMArena
  • Share of real agent workloads, measured by tokens used in live applications
  • Cost per completed task
  • Scale AI, SWE-Bench Pro V2 (public and private) — Share of 642 public and 272 private long coding tasks an agent resolves; Scale warns that public scores near 100% can mix skill with memorised tasks
  • Scale AI, SWE-bench Pro standardized board — Resolution under one common harness; the top score is 61.5%
  • BenchLM, SWE-bench Pro aggregate — Vendor-reported Pass@1 on the 731-task public set, not run under one common harness
  • LMArena text leaderboard (BenchLeader copy, 28 October 2026) — Elo from blind human votes between two answers

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, 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 · 1 notes

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

  1. 1
    Claude Mythos 5 (Anthropic)

    It tops the Swfte October 2026 leaderboard, and the earlier Mythos Preview led SWE-bench Pro at 77.8%, about 20 points above the next two models listed.

    Source ↗
  2. 2
    Claude Opus 5.5 (Anthropic)

    It is listed first on LMArena's WebDev board as of October 17, 2026, which makes it the practical coding pick.

    Source ↗
  3. 3
    Gemini 4 Argon (Google)

    It leads LMArena Text on human votes, but the 1524±9 score is preliminary.

    Source ↗
  4. 4
    Claude Opus 4.8 (Anthropic)

    A June 2026 guide gives it 88.6% on SWE-bench Verified but 69.2% on the harder Pro. The guide does not say whether that is a leaderboard result or a vendor figure.

    Source ↗
  5. 5
    DeepSeek V4.1 Flash

    It was still first on OpenRouter at 32.8T tokens in the report dated October 20, 2026, though its usage fell 20%.

  6. 6
    GPT-5.4 (OpenAI)

    It sits at 57.7% on SWE-bench Pro (BenchLM, April 2026). I found no October 2026 OpenAI data, so this is its last verified place.

    Source ↗
  7. 7
    GLM-5.1 (Z.ai)

    It is the open-weight model scoring 58.4% on SWE-bench Pro, ahead of GPT-5.4 in the BenchLM table.

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

    It held flat at 12.2T tokens on OpenRouter while DeepSeek fell, and it is the cheap-tier model agent builders are choosing.

  9. 9
    Kimi K2.6 (Moonshot)

    Sources name it the strongest open-weight choice for agentic work in 2026. I could not confirm a score.

    Source ↗
  10. 10
    MiMo V2.5 Pro (Xiaomi)

    Reportedly leads SWE-bench Pro for agentic coding, but that is Xiaomi's own claim and I could not verify it independently.

    Source ↗

What Claude noticed

Vendor scaffolds inflate SWE-bench numbers. One analysis argues a standardised 46% can be worth more than a vendor-reported 81%.

Source ↗
🇺🇸 ChatGPT10 places · 3 notes

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

  1. 1
    Claude Opus 5.5 (2026)

    26.6 trillion OpenRouter tokens in 30 days: a premium model with substantial observed demand despite cheaper alternatives.

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

    A frontier orchestration candidate whose adoption is harder to judge than the token-heavy budget models dominating public traffic.

    Source ↗
  3. 3
    Claude Fable 5 (2026)

    64.5% resolution in SWE-rebench at $4.40 per problem: a strong completion result with a meaningful cost premium.

  4. 4
    GPT-6 Astra (2026)

    Its 56.9 intelligence-index result exceeds GPT-6.1 Sol's 54.5, but the stronger general score does not prove cheaper agent execution.

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

    116 trillion tokens in 30 days, with published inference prices starting at $0.13 input and $0.52 output per million tokens.

    Source ↗
  6. 6
    GLM 5.3 Flash (2026)

    55.3 trillion monthly tokens put GLM's economical Flash tier ahead of many more celebrated frontier models in observed API volume.

    Source ↗
  7. 7
    Grok 4.5 (2026)

    63.8% issue resolution at $1.47 per problem in SWE-rebench makes Grok a serious cost-adjusted coding contender.

  8. 8
    Gemini 4 Argon (2026)

    A 50.4 intelligence-index score keeps Google's frontier model competitive, but general capability alone cannot settle agent reliability.

    Source ↗
  9. 9
    Kimi K3 Max (2026)

    A 1529 human-preference rating in an IT-focused arena puts Kimi close to Claude Fable 5's 1540, despite a very different market profile.

    Source ↗
  10. 10
    MiniMax M3 (2026)

    47.2% issue resolution at $0.95 per problem offers a lower-cost agent option, though success trails the leading frontier systems.

What ChatGPT noticed

The leading OpenRouter model family, DeepSeek V4.1 Flash, records 116T tokens in 30 days, more than four times Claude Opus 5.5's 26.6T. This measures traffic, not solved tasks.

Source ↗

Scale's SWE-bench Pro contains 1,865 tasks, and its original analysis found a major drop from the easier SWE-bench Verified results. Benchmark selection can change the apparent winner.

Source ↗

Hermes Agent shows 1.52T tokens in OpenRouter's participating-app listing, ahead of Claude Code at 1.25T and Codex at 508B. These are application volumes, not rankings of their underlying models.

Source ↗
🇺🇸 Meta10 places · 3 notes

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

  1. 1
    Claude Opus 5.5 (Anthropic, Sep 22 2026)

    89.9% on SWE-bench Pro – 8.6 points above Sonnet 5.5 – yet only 6.91T tokens on OpenRouter vs 47.8T for DeepSeek, frontier quality at $4 input/$20 output

  2. 2
    Claude Sonnet 5.5 (Anthropic, 2026)

    81.3% SWE-bench Pro – 8.6 points behind Opus 5.5 at half price $2/$10 – 2.61T tokens, best balance for everyday agents

  3. 3
    Gemini 4 Carbon (Google DeepMind, Oct 28 2026)

    1528 Elo #1 on LMArena Text Arena – 20 points above Opus 5.5 = 53% win rate – preliminary on 4,403 votes, Intelligence Index 59

    Source ↗
  4. 4
    DeepSeek V4.1 Flash (DeepSeek, Oct 17 2026 snapshot)

    47.8T tokens last 7 days – 29.4% of top 20 – 1017 snapshot alone, 43.6% of top 20 are DeepSeek models, $0.30 input $1.2 output

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

    10.3T tokens #4 on OpenRouter – cheapest frontier at $0.5/M output – $0.07 per agentic task on OSWorld vs $14.19 for Fable 5.1

    Source ↗
  6. 6
    GLM 5.3 Flash (Z.ai, 2026)

    15.7T tokens #2 on OpenRouter – $0.5/M output – open-weight surge from 27% to 72% of top 50 tokens year-over-year

  7. 7
    Claude Fable 5.1 (Anthropic, Sep 1 2026)

    81.2% SWE-bench Pro – only 0.1 point behind Sonnet 5.5 – 95.0% on SWE-bench Verified #1, but $7.63 per task vs Astra $3.26

  8. 8
    Gemini 4 Argon (Google, Oct 2026)

    1524 Elo #2 on LMArena – within 4 points of Carbon – High reasoning tier, $10/M output, $2/$10 intro pricing until Nov 30

    Source ↗
  9. 9
    Kimi K3 (Moonshot, 2026)

    Best open weights at #18 LMArena 1488 – K3.1 Pro $3/$15 – 1.97T tokens, #1 Arena full-stack 1664 per Turiloop budget pick

    Source ↗
  10. 10
    Muse Spark 1.1 (Meta, 2026)

    61.5% ±3.1 on Scale official public leaderboard – highest standardized score – $0.2/M tokens cheapest in top 20, Contributor tier

    Source ↗

What Meta noticed

Top 10 frontier models now run on only 14% of developer tokens, down from 55% a year ago – cheap flashes eat volume per The Information Oct 12

Source ↗

SWE-bench Pro gap: vendor-reported 89.9% (Opus 5.5) vs Scale standardized 61.5% (Spark 1.1) – same public set, different harness, plus 30% tasks flagged broken

Source ↗

LMArena crowding: Gemini 4 Carbon 1528 vs Opus 5.5 1508 – 20 points = ∼53% win rate, within 95% CI, preliminary on 4,403 votes

Source ↗
🇺🇸 Grok10 places · 4 notes

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

  1. 1
    DeepSeek V4.1 Flash

    Hermes Agent put 11.8 trillion tokens through DeepSeek V4.1 Flash this month, and a long coding session on it costs about $0.047.

    Source ↗
  2. 2
    Claude Opus 5.5

    Blind voters put Opus 5.5 first on agent tasks at 20.9, and a long coding session still costs about $1.39.

    Source ↗
  3. 3
    GLM 5.3

    The Flash edition was second in the week's traffic at 16.1 trillion tokens, while the full GLM-5.3 cleared 211 of 272 private SWE-Bench Pro V2 tasks.

    Source ↗
  4. 4
    GPT-6 Luna

    A long coding session on GPT-6 Luna costs about $0.024, the cheapest in OpenRouter's comparison, on 10.9 trillion tokens in a week.

    Source ↗
  5. 5
    Claude Opus 5

    Scale counts 638 of 642 public SWE-Bench Pro V2 tasks for Opus 5, and 222 of 272 private ones.

    Source ↗
  6. 6
    Gemini 4 Carbon

    Voters ranked Gemini 4 Carbon first on everyday answers and second on agent tasks, and it is missing from the week's token top ten.

    Source ↗
  7. 7
    Kimi K3

    On the private SWE-Bench Pro V2 count, Kimi K3 solved 214 of 272 tasks, two behind Opus 5.

    Source ↗
  8. 8
    Gemini 4 Argon

    Gemini 4 Argon sits at 99.22% on the public Pro V2 board, 0.18 points behind Opus 5, on a test Scale says is near the ceiling.

    Source ↗
  9. 9
    Claude Sonnet 5.5

    On OpenRouter's own flight-booking agent test, Sonnet 5.5 scored 78.7%, above Opus 5.5 at 75.5%, at half the token price.

    Source ↗
  10. 10
    Muse Spark

    Muse Spark 1.1 leads Scale's older private SWE-Bench Pro board at 61.5%, and it is not in the week's token top ten.

    Source ↗

What Grok noticed

The model agent apps actually call is not the model that wins the voting booth. Hermes Agent's top model this month is DeepSeek V4.1 Flash at 11.8 trillion tokens. Opus 5.5, first on LMArena's agent board, is sixth in OpenRouter's week at 7.31 trillion.

Source ↗

A 10-to-49-turn coding session costs about $0.024 on GPT-6 Luna, $0.047 on DeepSeek V4.1 Flash, $0.36 on GLM-5.3 and $1.39 on Claude Opus 5.5. The bill, not the leaderboard, is why the cheap models dominate tokens.

Source ↗

Scale's own write-up of SWE-Bench Pro V2 says Opus 5 resolves 638 of 642 public tasks and 222 of 272 private ones. The 99.4% headline and the private count are the same model.

Source ↗

On one agent chore, booking flights, OpenRouter measured Sonnet 5.5 at 78.7% and Opus 5.5 at 75.5%. The pricier model does not win every agent test.

Source ↗
🇺🇸 Gemini10 places · 1 notes

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

  1. 1
    Claude 3.5 Sonnet

    Processes more than 28% of all agentic coding tokens on OpenRouter and achieves a benchmark-defining 49% to 54% SWE-bench Verified solve rate.

    Source ↗
  2. 2
    OpenAI o1

    Introduced native test-time reasoning compute that pushes autonomous software repair to over 48% on complex SWE issues while charging a premium per reasoning chain.

  3. 3
    DeepSeek-V3

    Slashed agent orchestration economics by delivering 42% SWE-bench capability at roughly one-tenth the per-token cost of proprietary frontier models.

    Source ↗
  4. 4
    GPT-4o

    Remains the most versatile multimodal backbone for multi-agent supervisors, handling 15% to 20% of general agent traffic with sub-second time-to-first-token.

  5. 5
    Qwen 2.5 Coder 32B

    The premier local agent engine that achieves 35% on SWE-bench and fits inside a single dual-GPU workstation without recurring cloud API fees.

    Source ↗
  6. 6
    Claude 3.5 Haiku

    Serves as the preferred worker bee for sub-agent triage and linting loops, executing 3x faster than Sonnet with 40% of its reasoning punch.

    Source ↗
  7. 7
    Gemini 1.5 Pro

    Swallows entire multi-repository codebases in a single 2-million-token context window, bypassing complex RAG fragmentation for comprehensive agent planning.

    Source ↗
  8. 8
    GPT-4o mini

    Powers high-iteration agent state checkers and summarizers at under $0.15 per million input tokens, anchoring high-throughput commercial swarms.

  9. 9
    Llama 3.1 405B

    The enterprise open-weights champion providing unmetered, privacy-compliant agent execution for sovereign on-premise deployments.

  10. 10
    Nous Hermes 3 (Llama-3.1-8B/70B)

    The open agent community's favorite fine-tune for unaligned function calling, JSON schema compliance, and autonomous multi-agent simulation.

    Source ↗

What Gemini noticed

Human preference leaderboards like LMArena often decouple from actual developer agent usage: while models may tie in single-turn conversational Elo, developers allocate over 65% of production agent tokens to models strictly optimized for rigid tool-calling and scratchpad reasoning.

Source ↗
🇨🇳 Kimi10 places · 2 notes

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

  1. 1
    Claude Opus 5.5 (Anthropic, Sept 22 2026)

    The agentic triple crown at half the old flagship's price: 89.9% SWE-bench Pro, #1 Arena WebDev (1807), #1 Agent Arena (21.4% confirmed success) — and the first Anthropic flagship that costs $4/$20 instead of $10/$50.

  2. 2
    Claude Sonnet 5.5 (Anthropic, Sept 28 2026)

    The model developers actually burn tokens on: it tops Terminal-Bench 4.0 (70.6%) and LiveBench Coding (91.4) at half of Opus's price — the highest agentic score per dollar in the frontier tier.

    Source ↗
  3. 3
    GPT-6 Astra (OpenAI, Sept 3 2026)

    Human-vote darling, token-bill villain: #1 Frontend Code Arena (1793) and #2 WebDev (1774), but at $10/$50 it is 2.5x Opus 5.5's price — and OpenAI's share of live agent tokens is just 7.8%.

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

    The open-weights surprise that beats the giants at their own agentic test: it leads the independent Terminal-Bench 4.0 run (67.2%, Oct 23) ahead of every closed model except cyber-locked Gemini 4 Carbon — at $2/$10.

    Source ↗
  5. 5
    Gemini 4 Argon (Google, public Oct 13 2026)

    The human favorite that loses the lab tests: #1 LMArena text (1517.2) and the lowest hallucination rate measured (4%), yet 50.4 on Intelligence Index v5 — sixth among usable models.

    Source ↗
  6. 6
    GPT-6.1 Sol (OpenAI, Sept 29 2026)

    The volume agent inside Codex: default in the Codex CLI since Oct 7, 56.1% Terminal-Bench 4.0 near-Astra at one-fifth of Astra's price, and the cheapest route past 50 on the Intelligence Index.

    Source ↗
  7. 7
    DeepSeek V4.1 Flash (DeepSeek, Sept 10 2026, open weights)

    The token-economy king: DeepSeek is the #1 provider by live agent token share (39.1%) — more than Anthropic and Google combined — while this model lists at $0.15/$0.60 off-peak.

    Source ↗
  8. 8
    Grok 4.7 (xAI, Sept 21 2026)

    Cheapest frontier-flavored agent: $2/$6, EQ-Bench top-ten (2012.2) and Grok Build harness at 38% Terminal-Bench 4.0 — but Arena's human voters rank its writing 78th, below Grok 4.6.

  9. 9
    GLM-5.3 (Z.ai, open weights)

    The self-hostable workhorse: 95.4% on the archived SWE-bench Verified board — above Kimi K3 (93.4%) — under a plain MIT license, for $1.19/$6 hosted.

  10. 10
    Gemini 3.8 Flash (Google, Sept 2 2026)

    The agentic sprinter: 308 output tokens/sec at $0.75/$3.75 with 1M context — the model you put behind high-volume agent loops where a 20-point Terminal-Bench gap is worth 4x the throughput.

    Source ↗

What Kimi noticed

Test scores, human votes and real usage are three different leaderboards: Claude Opus 5.5 sweeps agentic tests (89.9% SWE-bench Pro, 21.4% Agent Arena) yet Gemini 4 Argon tops the human text vote (1517.2 vs 1508.7), while DeepSeek — 30+ index points behind — leads all live agent token share at 39.1%. Price and habit beat benchmark glory in production.

Source ↗

Open weights closed the agent gap first in the terminal: Kimi K3.1 Pro (67.2%) leads the independent Terminal-Bench 4.0 run ahead of every usable closed model, and GLM-5.3 plus DeepSeek V4.1-Flash sit on the token-share podium — the frontier is now a routing problem, not a license problem.

Source ↗
🇨🇳 DeepSeek10 places · 1 notes

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

  1. 1
    Claude Opus 5.5

    Claude Opus 5.5 tops the vendor-aggregate SWE-bench Pro board at 89.9% and the Hermes agent bench at 76.7% — the most complete “do-the-task” agent model of October 2026, at $4.19–$5.98 per successful task.

    Source ↗
  2. 2
    Gemini 4 Argon

    Gemini 4 Argon is the people’s choice — No. 1 on LMArena text at 1517 and multi-turn at 1544 — and it does it at $1.80 per successful task, roughly a quarter of Claude Opus 5.5’s cost.

    Source ↗
  3. 3
    Claude Fable 5.1

    Claude Fable 5.1 is LiveBench’s overall leader at 88.0 and scores a perfect 100 on browser use — the best “computer-use” agent model — but at $5.10 per successful task it is the most expensive in this top group.

  4. 4
    GPT-6 Astra

    GPT-6 Astra is the balanced frontier agent: 86.1 LiveBench overall, 82.4 agentic, and $3.81 per successful task — stronger than Gemini 4 Argon on agentic tasks and cheaper than Claude Fable 5.1.

  5. 5
    DeepSeek V4 Flash 0731

    DeepSeek V4 Flash is the real-world agent volume king: 11.4 trillion weekly tokens on OpenRouter and a $0.028 cost per successful task — roughly 180× cheaper than Claude Fable 5.1, though it trails the frontier on hard coding benchmarks.

    Source ↗
  6. 6
    Claude Sonnet 5.5

    Claude Sonnet 5.5 gives you 81.3% of Claude Opus 5.5’s SWE-bench Pro score at roughly 84% of its per-task cost — the value pick inside the Claude family for production agent loops.

    Source ↗
  7. 7
    Muse Spark 1.1 (Meta)

    Muse Spark 1.1 owns the only standardized top score on SWE-bench Pro — 61.5% on Scale’s public set — but its 51.5% commercial-set score shows how much scaffolding and data splits can move the number.

  8. 8
    Qwen3.8-Flash-Next

    Qwen3.8-Flash-Next is the best open-weights agent coder on SWE-bench Pro at 62.5% — ahead of GLM-5.2 at 62.1% — and its Flash sibling runs at about $0.196 per successful LiveBench task.

    Source ↗
  9. 9
    GLM-5.2

    GLM-5.2 sits just behind Qwen3.8-Flash-Next at 62.1% on SWE-bench Pro with a $2.40 output-token price — the second-best open-weights option for cost-sensitive agent fleets.

  10. 10
    Kimi K3

    Kimi K3 is built for orchestration: it scales to 300 sub-agents and 4,000 coordination steps, making it a specialist pick when the agent’s job is to manage other agents rather than write every line itself.

    Source ↗

What DeepSeek noticed

The model that wins human votes is not the model that wins real agent traffic. Gemini 4 Argon tops LMArena at 1517, but DeepSeek V4 Flash leads OpenRouter’s live token usage at 11.4 trillion weekly tokens — because at $0.028 per successful task, cost dominates production agent loops.

Source ↗
🇨🇳 Wenxin10 places · 1 notes

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

  1. 1
    OpenAI GPT-5 / o3 (2026)

    SWE-bench Pro 71% pass and the heaviest token load — but cost per task still stings developers

    Source ↗
  2. 2
    Anthropic Claude 4 Opus (2026)

    Highest human-vote Elo on LMArena (1392) and devs pick it for long-context agent runs

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

    21% token share and strong multi-modal agent scores — Google's bet on native tool use

    Source ↗
  4. 4
    Meta Llama 4 405B (2026)

    Open-weights king: 18% token share, devs run it on-prem — but agent coding lags closed models

  5. 5
    Mistral Large 3 (2026)

    European darling with 9% token share — punches above weight on code agents per dollar

  6. 6
    xAI Grok 4 (2026)

    7% token share; fast but agentic coding still behind — fans cite real-time data edge

  7. 7
    DeepSeek V4 (2026)

    Chinese open model at 6% token share; SWE-bench 55% for the price of a coffee

  8. 8
    Cohere Command R+ 2026 (Agent-optimized)

    Built for RAG and tool-use agents; 5% token share but LangSmith scores top-quartile

  9. 9
    Microsoft Phi-5 (2026)

    Small on-device model (7B params) with 3% token share — fastest per-token, not best per-task

  10. 10
    Stability AI Stable Code 3 (2026)

    Code-specific model at 2% token share; niche but loved by open-source agent builders

What Wenxin noticed

Open-weight models now handle nearly half of all agent tokens (Llama 4 + DeepSeek + Mistral + Phi = ~34%), a shift that didn't exist two years ago and reshapes who controls agent infrastructure.

Source ↗
🇨🇳 Qwen10 places · 2 notes

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

  1. 1
    Claude Opus 5.5

    89.9% on SWE-bench Pro leads the market, proving it resolves the most real GitHub issues.

    Source ↗
  2. 2
    Claude Sonnet 5.5

    The best all-round model for Hermes Agent, jumping to 70.6% on Terminal-Bench 4.0 for long tasks.

  3. 3
    GPT-5.6 Sol

    Leads LMArena human preference votes with 1485 Elo, beating out models that score higher on rigid coding tests.

    Source ↗
  4. 4
    Claude Fable 5.1

    Hits 81.2% on SWE-bench Pro, tying closely with Sonnet 5.5 but at a lower cost per completed task.

    Source ↗
  5. 5
    DeepSeek V4.1 Flash

    53.1T tokens processed on OpenRouter, making it the real workload king for developers routing agent traffic.

    Source ↗
  6. 6
    Gemini 4 Argon

    99.2% on SWE-bench Pro V2 shows Google's frontier model is catching up to Anthropic in long-horizon engineering.

    Source ↗
  7. 7
    Qwen 3 Max

    Hit #1 on OpenRouter's trending chart with 77.3B tokens, proving Alibaba's model handles massive agent workloads in Asia.

  8. 8
    GLM-5.2 Max

    Ranks #2 in Code Arena: Frontend with 1595 Elo, dominating human preference for UI agent tasks.

    Source ↗
  9. 9
    Llama 4 Maverick

    1.3M context window at $0.20/M input tokens makes it the cheapest way to run massive agent orchestration.

    Source ↗
  10. 10
    Claude Fable 5

    80.3% on SWE-bench Pro keeps it in the top tier for agentic coding despite newer models passing it.

    Source ↗

What Qwen noticed

The gap between SWE-bench Verified (~80%) and SWE-bench Pro (~58%) quantifies the 'contamination effect' where models memorized older test answers.

Source ↗

Claude Sonnet 5.5 outperforms the much larger Opus 5.5 on Terminal-Bench 4.0 by 4.2 points, surprising developers running long agentic loops.

Source ↗