Berkeley researchers demonstrated a fake curl wrapper scoring perfectly on 89 Terminal-Bench tasks without solving them; benchmark gaming can invalidate a headline score.
Best AI Agent 2026: Top 10 AI Agents
This list ranks coding agents people run. It does not price LangChain's $3 billion round. Claude Code is first on the merged ranks, and Google Antigravity CLI is third.
- No. 1: Claude Code
- 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.
Blind human pairwise votes on agent tasks
A bright badge means this AI put the entry on its list. A dim badge means it did not.
- 1
Claude Code
Strong across the lists: Opus 5.5 was reported at 61.7 on Coding Index v2.0; Fable 5.2 leads Arena, but those are different models and board versions.
Included it: Meta, Gemini, Claude, MiMo, StepFun, Kimi, DeepSeek, Qwen. Did not include: ChatGPT, Nemotron, Nova, Wenxin.In their lists: Meta: 1 · Gemini: 1 · Claude: 1 · MiMo: 1 · StepFun: 9 · Kimi: 4 · DeepSeek: 3 · Qwen: 1
What each AI noted 8
- Metaplace 1
Top on both taste and proof, but the bill is the catch — $15.47 a task at its best Artificial Analysis setting.
Source ↗ - Geminiplace 1
Dominates terminal-native developer preference while maintaining the highest verified bug-fix accuracy on SWE-bench Verified at an average cost under $2.50 per task.
Source ↗ - Claudeplace 1
It tops the Artificial Analysis index at 61.7, but at $11.42 per task it costs about 29 times the cheapest frontier point for 14.4 more points.
Source ↗ - MiMoplace 1
The one agent that leads both boards people actually read — 30.5% in the blind human vote and 58–62 in verified coding — and the one whose heaviest setting bills $44.22 a task to score two points below its own cheaper setting.
Source ↗ - StepFunplace 9
$2.26 a task at the top of a 30x price spread — open weights priced like the frontier, code reviewers struggle to tell apart.
Source ↗ - Kimiplace 4
Ties the Terminal-Bench lead at 65.8% while burning 15.7B tokens — nearly double Opus 5.5's 8.0B — to reach the same number: proof without parsimony.
Source ↗ - DeepSeekplace 3
Opus 5.5 is the Arena's former #1 and the coding index's 61.7-point reference — at $11.42 a task it is neither the cheapest nor the priciest frontier.
Source ↗ - Qwenplace 1
Tops the Artificial Analysis Coding Agent Index at 61.7 while anchoring Agent Arena's top tiers without the premium tax of Fable.
Source ↗
8 AIs put it here - 2
OpenAI Codex
The clearest price case: GPT-6.1 Sol scored 52.2 on v2.1 for $1.51 per task; Arena’s GPT-6 Astra is a separate configuration.
Included it: Meta, Gemini, Claude, MiMo, StepFun, Kimi, DeepSeek, Qwen. Did not include: ChatGPT, Nemotron, Nova, Wenxin.In their lists: Meta: 4 · Gemini: 10 · Claude: 5 · MiMo: 2 · StepFun: 3 · Kimi: 7 · DeepSeek: 6 · Qwen: 9
What each AI noted 8
- Metaplace 4
The cheapest way to stay near the frontier — $1.51 a task, not $15.
Source ↗ - Geminiplace 10
Frontier multimodal computer-use agent that bridges software tasks and browser actions, setting new reference marks on WebArena.
Source ↗ - Claudeplace 5
It scores 54.1 for $1.50 per task. That is within 0.6 points of Astra at about a fifth of the cost.
Source ↗ - MiMoplace 2
Scores 52 for $1.51 a task — the same index point range costs eight times more in Claude Code, and money, not taste, is where Codex wins its rank.
Source ↗ - StepFunplace 3
58.2% on Terminal-Bench 4.0 for $1.92 a task — half Claude Code's bill and half its runtime for 6.6 points less.
Source ↗ - Kimiplace 7
The value champion of the frontier: 93% of Astra's Terminal-Bench score for 19% of the per-task price — $1.92 against $9.90.
Source ↗ - DeepSeekplace 6
Codex on GPT-6.1 Sol gives up six index points to the leaders but costs $1.50 — 7.6x less than Claude Code at max effort.
Source ↗ - Qwenplace 9
Scored 80 points on the Coding Agent Index in July, leading OpenAI's harness before the SWE-bench Verified split.
Source ↗
8 AIs put it here - 3
Google Antigravity CLI
Gemini 4 Carbon scored 55.5 on Coding Index v2.1 and ranked second on Arena; its Arena cost was $1.36 per task, not a coding-index cost.
Included it: Meta, MiMo, StepFun, Kimi, DeepSeek. Did not include: ChatGPT, Nemotron, Gemini, Claude, Nova, Wenxin, Qwen.In their lists: Meta: 5 · MiMo: 3 · StepFun: 4 · Kimi: 6 · DeepSeek: 2
What each AI noted 5
- Metaplace 5
Two days old on 12 Nov, yet already the value pick — near-Carbon success for 24% less quota, by Google's own test.
Source ↗ - MiMoplace 3
First on the boards that copy real jobs — 84.2% on banking-consulting-legal work and $16,484 in a year-long vending business — while sitting only third in the human vote.
Source ↗ - StepFunplace 4
#3 in human votes and 63.6% on Terminal-Bench 4.0 as a model — but its CLI harness sits near the bottom of the AA composite at 43.
Source ↗ - Kimiplace 6
Wins the intelligence index (60) and takes silver on blind human votes (16.72%) — but its only Terminal-Bench number is a self-report the board refuses to rank.
Source ↗ - DeepSeekplace 2
Carbon wins CursorBench at 60.4% for $9.22, while the Arena's top model scores 60.0% on the same multi-file tasks for $32.79.
Source ↗
5 AIs put it here - 4
Devin Fusion CLI
A strong coding-index result at a lower reported cost than Claude Code, but its score changes across index versions: 61.5 in v2.0 and 58.4 in v2.1.
Included it: Meta, Claude, MiMo, Kimi, DeepSeek. Did not include: ChatGPT, Nemotron, Gemini, Nova, StepFun, Wenxin, Qwen.In their lists: Meta: 2 · Claude: 9 · MiMo: 4 · Kimi: 2 · DeepSeek: 5
What each AI noted 5
- Metaplace 2
A frontier lead plus a cheap sidekick ties the leader — $8.75 a task on v2.1, not $15.47.
Source ↗ - Claudeplace 9
Cognition's own FrontierCode benchmark now sits inside the index, and Devin is listed at 61.5.
Source ↗ - MiMoplace 4
Tied Claude Code's top coding score of 62 at $8.00 a task with a two-model fusion, and its SWE-2 posts 92.8% on Terminal-Bench 2.1 — the highest model row on that board.
Source ↗ - Kimiplace 2
The highest index score ever published on the AA board — 68, two points above Claude Code — by pairing a frontier lead model with a cheap GPT-6 Luna sidekick at 31% lower cost per task.
Source ↗ - DeepSeekplace 5
Devin Fusion lands 61.5 at $8.00, just 0.2 points behind Claude Code with Opus 5.5, using 10.46M tokens instead of 14.52M.
5 AIs put it here - 5
GitHub Copilot Fusion
The index reports 59.9 for $5.59 per task, but that result is from v2.0; do not compare it directly with v2.1 scores.
Included it: Meta, Claude, MiMo, DeepSeek, Qwen. Did not include: ChatGPT, Nemotron, Gemini, Nova, StepFun, Kimi, Wenxin.In their lists: Meta: 3 · Claude: 8 · MiMo: 7 · DeepSeek: 4 · Qwen: 5
What each AI noted 5
- Metaplace 3
Not a model, an orchestration engine — 2 points behind Claude Code at about half the per-task price.
Source ↗ - Claudeplace 8
A second source puts GitHub's agent at 59.9. That would place it above Meta's Muse Code, but I could not confirm the number.
Source ↗ - MiMoplace 7
59.9 on the coding index for $5.59 a task — within two points of Codex's flagship lane, from the agent that most index tables simply leave out.
Source ↗ - DeepSeekplace 4
Fusion scores 59.9 for $5.59 and leads Terminal-Bench at 71.2 — two points behind Claude Code at half the price.
Source ↗ - Qwenplace 5
Orchestrates OpenAI and Anthropic models to score just 2 points below Opus 5.5 at half the cost per task.
Source ↗
5 AIs put it here - 6
Fable 5.2 agent
Fable 5.2 leads the Arena measure at 30.48% net improvement, but that board is only one input; its result is not a verified cross-benchmark win.
Included it: Meta, Claude, MiMo, StepFun, DeepSeek. Did not include: ChatGPT, Nemotron, Gemini, Nova, Kimi, Wenxin, Qwen.In their lists: Meta: 8 · Claude: 3 · MiMo: see list · StepFun: 2 · DeepSeek: 1
What each AI noted 5
- Metaplace 8
The only general agent here that ships finished files, not just code — its GAIA claim was never independently re-verified.
Source ↗ - Claudeplace 3
Human voters put Fable 5.1 Max first in Agent Arena, yet it sits fourth on the Artificial Analysis index at 57.6.
Source ↗ - MiMo
Named differently in its own list: open the full list.
see list - StepFunplace 2
Wins 30.5% of blind human votes — nearly double the runner-up — yet users confirm success only 21.7% of the time.
Source ↗ - DeepSeekplace 1
Fable 5.2 tops the Arena at 30.35% and ties the coding-index lead at 58 — but its xhigh run costs $15.47, ten times Codex's $1.50.
Source ↗
5 AIs put it here - 7
Kimi Code CLI
Kimi K3.1 Pro scored 48.6 on Coding Index v2.1 for $2.94 per task; its Arena net-improvement score was 4.83%.
Included it: Meta, MiMo, Kimi, DeepSeek, Qwen. Did not include: ChatGPT, Nemotron, Gemini, Claude, Nova, StepFun, Wenxin.In their lists: Meta: 10 · MiMo: 6 · Kimi: 9 · DeepSeek: 9 · Qwen: see list
What each AI noted 5
- Metaplace 10
Open-weight flagship at $2.94 a task — and the open line whose predecessor led GAIA2's open-source field.
Source ↗ - MiMoplace 6
Beats Codex on terminal work 67 to 58 for $2.94 a task — then spends 185 turns and 8.9M tokens crawling there, the slowest strong lane on the board.
Source ↗ - Kimiplace 9
The open-weight outlier: within 6 points of the Terminal-Bench lead at $3/$15 per million tokens — while sitting 13 points below Opus on the intelligence index.
Source ↗ - DeepSeekplace 9
Kimi leads Terminal-Bench 4.1 at 68% and costs about $5.27 a task, yet Arena voters put it 11th at 5.2%.
Source ↗ - Qwen
Named differently in its own list: open the full list.
see list
5 AIs put it here - 8
Muse Code
Muse Wonder Full scored 54.5 on Coding Index v2.1, but the source says its cost was not listed; Arena’s score is 8.89% net improvement.
Included it: Meta, Claude, MiMo, DeepSeek, Qwen. Did not include: ChatGPT, Nemotron, Gemini, Nova, StepFun, Kimi, Wenxin.In their lists: Meta: 6 · Claude: 6 · MiMo: 5 · DeepSeek: 7 · Qwen: see list
What each AI noted 5
- Metaplace 6
Meta's best model is locked inside its own coding tool — and it still lands top-10 on both boards.
Source ↗ - Claudeplace 6
This Meta entry takes third place on the index at 58.9, ahead of Fable 5.1.
Source ↗ - MiMoplace 5
Matches the second-best verified coding score of 59 with the one number Artificial Analysis still cannot print: its price.
Source ↗ - DeepSeekplace 7
Sonnet 5.5 is fourth on Arena at 15.95%, but its $13.30 coding-task cost is higher than Copilot Fusion's $5.59 for a similar score.
Source ↗ - Qwen
Named differently in its own list: open the full list.
see list
5 AIs put it here - 9
DeepSeek Harness
A low reported cost of $0.32 per task comes with an index score of 37; the 90.6% Terminal-Bench figure measures a different model-board result.
Included it: MiMo, Kimi. Did not include: ChatGPT, Meta, Nemotron, Gemini, Claude, Nova, StepFun, DeepSeek, Wenxin, Qwen.In their lists: MiMo: 8 · Kimi: see list
2 AIs put it here - 10
Cursor Composer
A useful cost reference, but the cited $0.07 figure is from retired Index v1.5; current-benchmark placement is not confirmed here.
Included it: Meta, Gemini, StepFun. Did not include: ChatGPT, Nemotron, Claude, Nova, MiMo, Kimi, DeepSeek, Wenxin, Qwen.In their lists: Meta: 7 · Gemini: 2 · StepFun: 8
What each AI noted 3
- Metaplace 7
The editor people actually live in — about $0.07 a task on the retired Artificial Analysis version, a tenth of the frontier bill.
Source ↗ - Geminiplace 2
The uncontested leader in interactive human workflow speed, delivering verified multi-file completions at an ultra-low $0.07 to $0.44 per task.
Source ↗ - StepFunplace 8
The harness tax: the same model costs $1.47 per task in Cursor CLI versus $1.24 in Claude Code.
Source ↗
3 AIs put it here
02Worth knowing
What the AIs noticed while building the list.
Princeton HAL reports a verified GAIA leader at 74.55%, while open-board submissions cited in the answers approach 94%; these are different submission and verification settings.
03How we counted
In plain words, no formulas.
1The question
Top 10 AI Agents: Human Votes vs Verified Task Completion vs Cost per Task
2Why this topic
The reader learns that 'best agent' is really three questions and sees all three answers in one place, which no single leaderboard gives.
3Where it applies
global, frontier AI agents, coding agents and multi-agent orchestrators
4What we looked at
- Blind human pairwise votes on agent tasks
- Verified task success on coding and multi-step benchmarks
- Cost, tokens and time per task
- An agent must appear on at least two independent boards; results that game a benchmark are flagged, not ranked
- Arena — Agent Arena ranks orchestrator models using real-session signals such as confirmed success, user feedback, steerability and tool reliability; net improvement is not a blind pairwise win rate.
- Artificial Analysis — The Coding Agent Index v2.1 averages FrontierCode-AA 2.0, Terminal-Bench 4.1 and SWE-Atlas-QnA-AA, and reports task cost, tokens and time.
- Princeton HAL — Reproduces agent results on GAIA’s public validation set and reports accuracy and total API cost; the board says its updates are paused.
- UC Berkeley RDI — Studies benchmark gaming; its report describes a fake curl wrapper that scored perfectly on Terminal-Bench without writing solution code.
5Who ranked
The leading AIs, each building its own list of ten independently and naming a source for every place: ChatGPT, Meta, Nemotron, Gemini, Claude, Nova, MiMo, StepFun, 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.
04Each AI's own list
🇺🇸 Meta10 places · 4 notes
The same question, answered by Meta alone. The final list above combines all the AIs.
- 1Claude Code (Claude Opus 5.5 / Fable 5.2)
Top on both taste and proof, but the bill is the catch — $15.47 a task at its best Artificial Analysis setting.
Source ↗ - 2Devin Fusion CLI (Cognition)
A frontier lead plus a cheap sidekick ties the leader — $8.75 a task on v2.1, not $15.47.
Source ↗ - 3GitHub Copilot Fusion
Not a model, an orchestration engine — 2 points behind Claude Code at about half the per-task price.
Source ↗ - 4OpenAI Codex (GPT-6.1 Sol)
The cheapest way to stay near the frontier — $1.51 a task, not $15.
Source ↗ - 5Google Antigravity / Antigravity Orchestrator (Gemini 4 Carbon)
Two days old on 12 Nov, yet already the value pick — near-Carbon success for 24% less quota, by Google's own test.
Source ↗ - 6Muse Code (Muse Wonder Full)
Meta's best model is locked inside its own coding tool — and it still lands top-10 on both boards.
Source ↗ - 7Cursor (Composer 2.5)
The editor people actually live in — about $0.07 a task on the retired Artificial Analysis version, a tenth of the frontier bill.
Source ↗ - 8Manus 2.0
The only general agent here that ships finished files, not just code — its GAIA claim was never independently re-verified.
Source ↗ - 9ChatGPT Work (GPT-6 Astra / GPT-5.6)
OpenAI retired Agent mode into Work — the office agent, not the coding agent, and its best exam score is 41.6%.
Source ↗ - 10Kimi Code CLI (Kimi K3.1 Pro)
Open-weight flagship at $2.94 a task — and the open line whose predecessor led GAIA2's open-source field.
Source ↗
What Meta noticed
Orchestration, not just models, is what investors just priced: on 13 Nov 2026 LangChain said it raised $300M at a $3B valuation.
LangChain pitches LangSmith as a neutral home for corporate agents, with LangSmith passing 7,000 organizations in the same raise announcement.
Fusion is the pattern of the autumn: Devin Fusion, Copilot Fusion and Antigravity Orchestrator all pair a frontier planner with a cheaper executor in one task.
Max effort can be worse and far pricier: on Artificial Analysis v2.1, Claude Code Fable 5.2 at Max scores 55.7 for $44.22 and 129.5 min, while High scores 58.2 for $7.51 and 31.3 min.
🇺🇸 Nemotron10 places · 1 notes
The same question, answered by Nemotron alone. The final list above combines all the AIs.
- 1Claude 3.5 Sonnet (Anthropic)
Leads LMArena agent win-rate at 72.4% while topping GAIA2 at 68.1% pass@1 and Artificial Analysis coding at 84.7% success for $0.42/task
Source ↗ - 2GPT-4o (OpenAI)
Second on LMArena at 69.8% win-rate, GAIA2 64.3%, coding 81.2% success but costs $0.68/task — 62% more than Sonnet
Source ↗ - 3Gemini 1.5 Pro (Google)
Third on LMArena at 66.1%, GAIA2 61.7%, coding 78.9% success at $0.31/task — cheapest frontier model but slower at 67s/task
Source ↗ - 4Claude 3.5 Haiku (Anthropic)
LMArena 61.3%, GAIA2 55.2%, coding 73.4% success at $0.08/task — 5x cheaper than Sonnet with 87% of its coding performance
- 5GPT-4o-mini (OpenAI)
LMArena 58.7%, GAIA2 51.8%, coding 69.1% success at $0.04/task — cheapest model on all three boards but 19 points behind Sonnet on GAIA2
Source ↗ - 6Llama 3.1 405B (Meta)
LMArena 56.2%, GAIA2 48.9%, coding 65.7% success at $0.12/task — best open-weight model but trails proprietary frontier by 19 GAIA2 points
- 7Qwen2.5-Coder-32B (Alibaba)
LMArena 52.8%, GAIA2 44.3%, coding 71.2% success at $0.06/task — coding-specialized model beats 405B on coding but lags on general agent tasks
Source ↗ - 8DeepSeek-V2.5 (DeepSeek)
LMArena 51.4%, GAIA2 42.7%, coding 68.9% success at $0.05/task — strong Chinese open-weight contender, 3.3 points behind Qwen on coding at lower cost
Source ↗ - 9LangGraph + Claude 3.5 Sonnet (LangChain)
Multi-agent orchestrator: GAIA2 71.3% (beats solo Sonnet by 3.2pts), coding 87.1% at $1.84/task — 4.4x cost premium for orchestration gains
- 10AutoGen + GPT-4o (Microsoft)
Multi-agent orchestrator: GAIA2 69.8%, coding 83.4% at $1.67/task — 2.5x cost premium over solo GPT-4o, trails LangGraph by 1.5 GAIA2 points
What Nemotron noticed
Claude 3.5 Sonnet is the only model to rank top-3 on all three boards simultaneously — human preference, verified task completion, and cost efficiency — a 'triple crown' no other frontier model achieves as of November 2026
🇺🇸 Gemini10 places
The same question, answered by Gemini alone. The final list above combines all the AIs.
- 1Claude Code
Dominates terminal-native developer preference while maintaining the highest verified bug-fix accuracy on SWE-bench Verified at an average cost under $2.50 per task.
Source ↗ - 2Cursor Composer
The uncontested leader in interactive human workflow speed, delivering verified multi-file completions at an ultra-low $0.07 to $0.44 per task.
Source ↗ - 3Devin by Cognition
Pioneered the asynchronous software engineer category, solving complex repository-scale tickets autonomously, though at a premium compute cost of $5.00 to $15.00 per task.
Source ↗ - 4OpenHands
The top-ranked open-source autonomous agent framework matches proprietary frontier scores on SWE-bench while eliminating vendor harness markups.
Source ↗ - 5LangGraph by LangChain
Anchored by LangChain's November 2026 $3 billion valuation, it is the enterprise standard for deterministic multi-agent state machines across fintech and customer operations.
- 6Windsurf Cascade
Codeium's Cascade engine challenges Cursor with deep collaborative flow, scoring near-parity in user satisfaction at $0.10 to $0.50 per coding task.
Source ↗ - 7Aider
The open-source git-diff pioneer achieves benchmark-topping token efficiency, completing full repository edits for under $0.25 on modern reasoning models.
Source ↗ - 8Microsoft AutoGen
The multi-agent conversation pioneer leads in complex multi-persona simulations and enterprise sandbox orchestration.
Source ↗ - 9CrewAI
The fastest-adopted multi-agent role-playing platform, chosen by fintech and business teams for lightweight orchestrations with predictable cost curves.
- 10OpenAI Operator
Frontier multimodal computer-use agent that bridges software tasks and browser actions, setting new reference marks on WebArena.
Source ↗
🇺🇸 Claude10 places · 4 notes
The same question, answered by Claude alone. The final list above combines all the AIs.
- 1Claude Code with Claude Opus 5.5 (2026)
It tops the Artificial Analysis index at 61.7, but at $11.42 per task it costs about 29 times the cheapest frontier point for 14.4 more points.
Source ↗ - 2Claude Code with Claude Sonnet 5.5 (2026)
Sonnet is 2.4 points behind Opus on the index, and its max-effort setting also tops the cost-matched field.
Source ↗ - 3Claude Code with Claude Fable 5.1 (2026)
Human voters put Fable 5.1 Max first in Agent Arena, yet it sits fourth on the Artificial Analysis index at 57.6.
Source ↗ - 4Codex with GPT-6 Astra (2026)
It is second with human voters in Agent Arena and fifth on the verified index, at $7.04 per task.
Source ↗ - 5Codex with GPT-6.1 Sol (2026)
It scores 54.1 for $1.50 per task. That is within 0.6 points of Astra at about a fifth of the cost.
Source ↗ - 6Muse Code with Muse Wonder Full (Meta, 2026)
This Meta entry takes third place on the index at 58.9, ahead of Fable 5.1.
Source ↗ - 7Codex with DeepSeek V4.1 Flash (2026)
It scores 47.3 for $0.39 per task, the cheapest point on the frontier.
Source ↗ - 8GitHub Copilot with Copilot Fusion (2026)
A second source puts GitHub's agent at 59.9. That would place it above Meta's Muse Code, but I could not confirm the number.
Source ↗ - 9Devin Fusion CLI (Cognition, 2026)
Cognition's own FrontierCode benchmark now sits inside the index, and Devin is listed at 61.5.
Source ↗ - 10Muse Code with Muse Spark Open (Meta, 2026)
This open-weights Meta configuration is on the index with no published score. It is the one open-weights entry among the leaders.
Source ↗
What Claude noticed
The top four on the Artificial Analysis index sit within 4.1 points, but cost per task spans 29 times between the best score and the cheapest frontier point.
Index v2.0 lowered scores for 20 of 21 configurations measured on both versions (mean 54.8 to 47.4), because Cognition's FrontierCode replaced DeepSWE.
The index scores a model and harness together. Claude Opus 5.5 in Claude Code is a different entry from the same model in another harness.
Arena says it is dividing 'best' into tasks, costs and confidence intervals. It now reports median task cost and tool hallucination beside human votes.
🇺🇸 Nova10 places · 2 notes
The same question, answered by Nova alone. The final list above combines all the AIs.
- 1Gemini 4 Carbon
Preliminary leader in Arena text board with 1520 Elo
Source ↗ - 2Claude Opus 5.5
1509 Elo in Arena, top student-accessible model
Source ↗ - 3GPT-6 Astra
1773 Elo in Arena WebDev, strong coding agent
Source ↗ - 4Claude Sonnet 5.5
1770 Elo in Arena WebDev
Source ↗ - 5GPT-6.1 Sol
1751 Elo in Arena WebDev
Source ↗ - 6Claude Fable 5.1
1504 Elo in Arena text board
Source ↗ - 7Meta Muse Wonder
1507 Elo in Arena text board
Source ↗ - 8GPT-6.1 Sol + Codex
58.2% success in Artificial Analysis coding benchmark
Source ↗ - 9Claude Fable 5.1 + Claude Code
57.9% success in Artificial Analysis coding benchmark
Source ↗ - 10GPT-6 Astra + Codex
58.2% success in Artificial Analysis coding benchmark
Source ↗
What Nova noticed
Gemini 4 Carbon leads preliminary human votes but isn't publicly available, creating access disparity
Top 4 models in Arena WebDev are within statistical ties (1751-1810 Elo)
🇨🇳 MiMo10 places · 4 notes
The same question, answered by MiMo alone. The final list above combines all the AIs.
- 1Claude Code (Anthropic, 2026)
The one agent that leads both boards people actually read — 30.5% in the blind human vote and 58–62 in verified coding — and the one whose heaviest setting bills $44.22 a task to score two points below its own cheaper setting.
Source ↗ - 2OpenAI Codex (ChatGPT agent, 2026)
Scores 52 for $1.51 a task — the same index point range costs eight times more in Claude Code, and money, not taste, is where Codex wins its rank.
Source ↗ - 3Google Antigravity CLI / Gemini 4 (2026)
First on the boards that copy real jobs — 84.2% on banking-consulting-legal work and $16,484 in a year-long vending business — while sitting only third in the human vote.
Source ↗ - 4Devin (Cognition, formerly Windsurf, 2026)
Tied Claude Code's top coding score of 62 at $8.00 a task with a two-model fusion, and its SWE-2 posts 92.8% on Terminal-Bench 2.1 — the highest model row on that board.
Source ↗ - 5Muse Code (Meta, 2026)
Matches the second-best verified coding score of 59 with the one number Artificial Analysis still cannot print: its price.
Source ↗ - 6Kimi Code CLI (Moonshot AI, 2026)
Beats Codex on terminal work 67 to 58 for $2.94 a task — then spends 185 turns and 8.9M tokens crawling there, the slowest strong lane on the board.
Source ↗ - 7GitHub Copilot Coding Agent (Copilot Fusion, 2026)
59.9 on the coding index for $5.59 a task — within two points of Codex's flagship lane, from the agent that most index tables simply leave out.
Source ↗ - 8DeepSeek Harness (DeepSeek, 2026)
$0.32 a task — 48 times cheaper than Claude Code's priciest lane — while the very same model posts 90.6% on Terminal-Bench's model board.
Source ↗ - 9Grok Build (xAI, 2026)
62.9% on repository Q&A against 33.3% on terminal tasks — a 30-point split inside one agent that no rival shows.
Source ↗ - 10OpenCode (open-source coding agent, 2026)
The free open-source harness scores the same 49 as Kimi Code CLI and runs 75+ model providers behind one key — the people's lane in a $1.51–$44 race.
Source ↗
What MiMo noticed
Max effort can cost more and score less: Claude Code with Fable 5.2 at max bills $44.22 per task for index 56, while the same model at high effort scores 58 for $7.51 — six times cheaper and two points better.
The taste vote and the real-job boards crown different agents: Claude Fable 5.2 leads Agent Arena by 11 points (30.5% vs 19.5%), yet Gemini 4 Carbon wins APEX-Agents (84.2%) and Vending-Bench 2 ($16,484) — preference, professional work and business coherence are three different skills.
Among configurations scoring 48+, cost per task spans roughly 30x — $1.51 (Codex GPT-6.1 Sol) to $44.22 (Claude Code Fable 5.2 max) — so four index points currently carry a ~29x price multiplier; the smart money splits lanes by task difficulty.
The harness is half the agent: Gemini 4 Argon scored 40.7/105 bugs through Antigravity CLI but 33/105 through Gemini CLI on the same Bug Hunt Bench — and $7.16 vs $32.08 per run.
🇨🇳 StepFun10 places · 4 notes
The same question, answered by StepFun alone. The final list above combines all the AIs.
- 1Claude Code + Opus 5.5 (Anthropic) 2026
Top-2 in blind human votes and #1 on the verified coding boards — the only stack on both podiums, at $15.47 a task.
Source ↗ - 2Claude Fable 5.2 Max agent 2026
Wins 30.5% of blind human votes — nearly double the runner-up — yet users confirm success only 21.7% of the time.
Source ↗ - 3OpenAI Codex + GPT-6.1 Sol 2026
58.2% on Terminal-Bench 4.0 for $1.92 a task — half Claude Code's bill and half its runtime for 6.6 points less.
Source ↗ - 4Gemini 4 Carbon High / Gemini CLI 2026
#3 in human votes and 63.6% on Terminal-Bench 4.0 as a model — but its CLI harness sits near the bottom of the AA composite at 43.
Source ↗ - 5Claude Sonnet 5.5 Max in Claude Code 2026
More confirmed successes (19.5%) than votes (16.1%) — on this stack the proof beats the taste.
Source ↗ - 6GitHub Copilot agentic harness 2026
The one harness benchmarked against Claude Code and Codex on five boards at once — and it won the token count.
Source ↗ - 7Cognition Devin SWE-2 2026
81.5% with a human allowed, 45.8% without — same agent, two very different leaderboards.
Source ↗ - 8Cursor Composer 2.5 2026
The harness tax: the same model costs $1.47 per task in Cursor CLI versus $1.24 in Claude Code.
Source ↗ - 9GLM-5.1 coding agent (in Claude Code) 2026
$2.26 a task at the top of a 30x price spread — open weights priced like the frontier, code reviewers struggle to tell apart.
Source ↗ - 10Kimi K2.6 coding agent 2026
Open weights, 300 parallel sub-agents, and a 16-point jump on Terminal-Bench in a single release.
Source ↗
What StepFun noticed
The taste premium is measurable: Claude Fable 5.2 takes 30.48% of blind Arena votes but only 21.66% Confirmed Success — people like it about 9 points more than they can confirm it works.
Price is the wildest axis: the Artificial Analysis Coding Agent Index spans $0.07 to $2.26 per task — more than 30x — at near-identical code quality, so the cheapest list is a different sport from the best list.
On the hardest multi-step board, nobody wins: GAIA2's ICLR 2026 paper puts the frontier at ~42% pass@1 with GPT-5 (high) leading at 42.1% and no system dominating every capability — and Claude runs costlier per task than GPT-5 (high) despite fewer tokens per step.
The money is betting on the orchestration layer: LangChain raised $300M at a $3B valuation on Nov 13, 2026, led by CapitalG — more than double its prior mark.
🇨🇳 Kimi10 places · 4 notes
The same question, answered by Kimi alone. The final list above combines all the AIs.
- 1Claude Code (Claude Opus 5.5, Max effort)
The only agent that tops both the blind human-vote board and the regraded verified coding board in the same month — first at +20.14% on Arena Agents and 65.76% on Terminal-Bench 4.1.
Source ↗ - 2Devin Fusion CLI (Cognition)
The highest index score ever published on the AA board — 68, two points above Claude Code — by pairing a frontier lead model with a cheap GPT-6 Luna sidekick at 31% lower cost per task.
Source ↗ - 3GitHub Copilot (Copilot Fusion)
Half the per-task price of the leader at only two index points back — and the only entry that leads Terminal-Bench 4.0 outright with a 71.
Source ↗ - 4Claude Fable 5.2 (Claude Code)
Ties the Terminal-Bench lead at 65.8% while burning 15.7B tokens — nearly double Opus 5.5's 8.0B — to reach the same number: proof without parsimony.
Source ↗ - 5GPT-6 Astra (OpenAI Codex)
Pays the most per task of any OpenAI agent — $9.90 at max effort — for 0.3 points more than its own $7.12 xhigh run on the same board.
Source ↗ - 6Gemini 4 Argon / Carbon (Google Antigravity)
Wins the intelligence index (60) and takes silver on blind human votes (16.72%) — but its only Terminal-Bench number is a self-report the board refuses to rank.
Source ↗ - 7GPT-6.1 Sol (OpenAI Codex)
The value champion of the frontier: 93% of Astra's Terminal-Bench score for 19% of the per-task price — $1.92 against $9.90.
Source ↗ - 8Claude Sonnet 5.5 (Claude Code)
Second on Terminal-Bench 4.1 yet the priciest top entry of the 4.0 runs — $22.22 per task, 19.4B tokens and $7.3k for 63.9%.
Source ↗ - 9Kimi K3.1 Pro (Moonshot AI, open weights)
The open-weight outlier: within 6 points of the Terminal-Bench lead at $3/$15 per million tokens — while sitting 13 points below Opus on the intelligence index.
Source ↗ - 10OpenCode + GLM-5.3 (open source)
The only open-source harness on the AA Coding Agent Index — 53.6 at $4.24 per task — proof that the harness, not just the model, is the moat.
Source ↗
What Kimi noticed
Taste, proof and price crown three different champions: Claude Opus 5.5 wins blind human votes (+20.14% on Arena Agents), the AA index lead changed hands between Claude Code and Devin Fusion (68 on the Oct 21 snapshot), and GPT-6.1 Sol wins cost per task ($1.92 at 58.2% on Terminal-Bench 4.0).
The GAIA gap between verified and self-reported results is ~20 points: the Princeton HAL board's verified best is 74.55% (total run cost $178.20) while anonymous self-reports on the open board reach 94.35% — and in April 2026 a UC Berkeley exploit agent scored about 98% on GAIA without solving a single task.
GAIA2 resets expectations for 'agentic' claims: in Meta's dynamic world where time keeps ticking, the leader (GPT-5 at high reasoning) passes only 42.1% of scenarios, and the best open model (Kimi-K2) reaches roughly 20%.
Money is flowing to the orchestration layer, not the models: on Nov 13, 2026 LangChain raised $300M at a $3B valuation — days after Devin Fusion and Copilot Fusion showed that routing between several models is itself the product.
🇨🇳 DeepSeek10 places · 4 notes
The same question, answered by DeepSeek alone. The final list above combines all the AIs.
- 1Claude Code with Claude Fable 5.2
Fable 5.2 tops the Arena at 30.35% and ties the coding-index lead at 58 — but its xhigh run costs $15.47, ten times Codex's $1.50.
Source ↗ - 2Gemini 4 Carbon in Antigravity CLI
Carbon wins CursorBench at 60.4% for $9.22, while the Arena's top model scores 60.0% on the same multi-file tasks for $32.79.
Source ↗ - 3Claude Code with Claude Opus 5.5
Opus 5.5 is the Arena's former #1 and the coding index's 61.7-point reference — at $11.42 a task it is neither the cheapest nor the priciest frontier.
Source ↗ - 4GitHub Copilot Fusion
Fusion scores 59.9 for $5.59 and leads Terminal-Bench at 71.2 — two points behind Claude Code at half the price.
Source ↗ - 5Devin Fusion CLI
Devin Fusion lands 61.5 at $8.00, just 0.2 points behind Claude Code with Opus 5.5, using 10.46M tokens instead of 14.52M.
- 6OpenAI Codex with GPT-6.1 Sol
Codex on GPT-6.1 Sol gives up six index points to the leaders but costs $1.50 — 7.6x less than Claude Code at max effort.
Source ↗ - 7Claude Code with Claude Sonnet 5.5
Sonnet 5.5 is fourth on Arena at 15.95%, but its $13.30 coding-task cost is higher than Copilot Fusion's $5.59 for a similar score.
Source ↗ - 8OpenAI Codex with GPT-6 Astra
Astra is sixth on Arena at 14.54% and costs $7.04 on the coding index — cheaper than Opus 5.5 but behind on verified merges.
Source ↗ - 9Kimi K3.1 Pro agent
Kimi leads Terminal-Bench 4.1 at 68% and costs about $5.27 a task, yet Arena voters put it 11th at 5.2%.
Source ↗ - 10Claude Code with Claude Fable 5.1
Fable 5.1 is fifth on Arena at 14.82% and costs $17.28 on CursorBench's multi-file tasks — still frontier, just no longer the headline.
Source ↗
What DeepSeek noticed
Two different winners in the same week: Arena's 11 November publish put Fable 5.2 #1 at 30.35%, while CursorBench 4.0 put Gemini 4 Carbon #1 at 60.4% — a model's rank depends on which task distribution the board samples.
The top coding score and the cheapest coding score sit 29x apart: Fable 5.2 at max effort costs $44.22 per task on the AA index, while Codex GPT-6.1 Sol costs $1.51.
Copilot Fusion finishes a coding task in 36.1 minutes versus Claude Code's 61.0 — 41% faster and 51% cheaper, but two points lower on the composite.
GAIA's harness effect is larger than most model gaps: the same model can score 44% bare and 74% inside a well-built agent framework, a 30-point swing.
🇨🇳 Wenxin10 places · 4 notes
The same question, answered by Wenxin alone. The final list above combines all the AIs.
- 1Claude Opus 5.5 (Anthropic 2026)
Ties Terminal-Bench 4.0 top score at 65.15% with 1512 Elo on LMArena
Source ↗ - 2Claude Fable 5.2 (Anthropic 2026)
Agent Arena #1 at +30.48 net user lift, matching 65.15% Terminal-Bench
Source ↗ - 3Claude Sonnet 5.5 (Anthropic 2026)
#1 APEX-Agents score 75.5% at5/5/25 per M tokens, best value top-tier
Source ↗ - 4Atria Dawn Preview (Shanghai AI Lab 2026)
BenchLM verified agentic score 92.5/100, top open-weight global agent
Source ↗ - 5Kimi K3.1 Pro (Moonshot AI 2026)
Terminal-Bench 4.0 #4 at 63.64%, top open-weight coding agent
Source ↗ - 6Gemini 4 Carbon (Google 2026)
Agent Arena #3 at +16.90, 1.46 per task, fastest P50 10min runtime
Source ↗ - 7GPT-5.6 Sol (OpenAI 2026)
LMArena 1514 Elo, 3x ARC-AGI-3 score with retained reasoning enabled
Source ↗ - 8Gemini 4 Argon (Google 2026)
Agent Arena lowest cost 0.712 per task, 7min P50 runtime
Source ↗ - 9DeepSeek V4 Pro (DeepSeek 2026)
Terminal-Bench Hard top at1.74/1.74/3.48 per M tokens, best open coding value
Source ↗ - 10Muse Code + Muse Wonder (Meta 2026)
91.6% FuncPass #1, 25.1% SecPass #7, widest Func/Sec gap ever measured
Source ↗
What Wenxin noticed
Top 3 Claude agents all score ≥64% on Terminal-Bench 4.0, a threshold no other vendor crosses
Enabling two simple API settings tripled GPT-5.6 Sol’s ARC-AGI-3 score and cut output tokens by 6x
Muse Wonder posted the highest ever measured FuncPass/SecPass gap at 27.4% ratio
Agents’ Last Exam hardest tier full pass rate remains below 1% across all mainstream systems
🇨🇳 Qwen10 places · 3 notes
The same question, answered by Qwen alone. The final list above combines all the AIs.
- 1Claude Code with Opus 5.5
Tops the Artificial Analysis Coding Agent Index at 61.7 while anchoring Agent Arena's top tiers without the premium tax of Fable.
Source ↗ - 2Claude Fable 5.2 (Max)
Leads Agent Arena's blind human votes at a 30.48% win rate but costs roughly 2.6 times more than Opus 5.5 for the edge.
Source ↗ - 3Gemini 4 Carbon
Scores 59 on the Coding Agent Index at exactly half the API price of Claude Opus 5.5, making it the efficiency king.
Source ↗ - 4GPT-6 Sol
Ranks second on Agent Arena's work leaderboard and averages $2.99 per task on the Coding Agent Index.
Source ↗ - 5GitHub Copilot Fusion
Orchestrates OpenAI and Anthropic models to score just 2 points below Opus 5.5 at half the cost per task.
Source ↗ - 6Sage-A1.5
Ties for the top verified score on the independent Princeton HAL GAIA leaderboard at 95.1% while also posting strong metrics on Artificial Analysis.
Source ↗ - 7CustomGPT.ai Research Lab v44
Dominates the GAIA multi-step assistant benchmark with a 97.85% success rate on Hugging Face's official tracker.
- 8Gemini 4 Hydrogen (High)
Debuted at #5 on Agent Arena with a +14.8% net improvement across 6,500 real-world agentic sessions.
Source ↗ - 9OpenAI Codex (GPT-5.6 Sol)
Scored 80 points on the Coding Agent Index in July, leading OpenAI's harness before the SWE-bench Verified split.
Source ↗ - 10Devin
Cognition's autonomous software engineer remains a staple on Agent Arena and GAIA comparisons despite high hourly pricing.
Source ↗
What Qwen noticed
Multi-agent orchestration is beating raw frontier models on price-to-performance. GitHub Copilot Fusion orchestrates OpenAI and Anthropic models to score just 2 points below Claude Opus 5.5 while cutting the cost per task in half.
The 'Taste vs Proof' gap is massive. Claude Fable 5.2 wins blind human votes on Agent Arena at a 30.48% win rate, but costs 2.6 times more than Opus 5.5 for only marginal gains in verified task completion.
LangChain's $3 billion valuation (raising $300 million on Nov 13, 2026) reflects the market's shift from single models to orchestration frameworks, exactly the technology powering multi-agent cost-savers like GitHub Copilot Fusion.