Who's on Top? The AI Model Leaderboard, October 2026

People ask "which AI is best?" as if there were one scoreboard. There isn't. In late October 2026 Google holds the top of the most-watched human-preference chart. Anthropic leads most coding and agent tests. OpenAI holds the top science-exam scores and sells the best intelligence per dollar. Meta has rejoined the frontier, and Chinese labs keep pushing open-weight models closer to the leaders. Each of those statements is true, and each comes from a different measuring stick.

This article covers model quality only: how well the newest models score on the public leaderboards and benchmarks that people actually cite. It doesn't cover market share, user counts or revenue. A separate Apiary article handles those.

We wrote it from the live leaderboards and lab announcements we could open in the last week of October 2026. Every number below comes from a named source. When a figure is vendor-reported (published by the company that built the model), we say so. When something couldn't be verified, we left it out. A careful article will look emptier in places than a hype article would, and that's on purpose.

The short answer

Here is who leads each major yardstick as of the end of October 2026, with the runner-up. Read the sections below before treating any line as final. Several leads sit inside the statistical margin of error.

Here's the same picture by lab. Google leads human-preference text chat. Anthropic leads coding, agentic work and hard exams, and fills the most slots in every top-ten list. OpenAI leads science Q&A and price-performance, but it's surprisingly absent from the top of the Arena text chart. Meta is back in the top five on Arena text for the first time in a long while. Moonshot (Kimi), Alibaba (Qwen) and DeepSeek lead the open-weight field, though the gap to the closed frontier widened this month on the toughest index. xAI and Mistral trail the frontier on independent measurements right now.

How to read any AI leaderboard (two-minute primer)

Three broad kinds of scoreboards exist, and they measure different things.

1. Human-preference arenas. On Arena (formerly LMArena, earlier Chatbot Arena), a user types a prompt, two anonymous models answer, and the user votes for the better answer. Votes become a rating, and Wikipedia's article on the platform describes the anonymous side-by-side setup. A higher score means people preferred the answers more often. That isn't the same as "more correct." Every score carries a ± confidence band, and small gaps are often ties.

1. Fixed benchmarks. A benchmark is a set of questions or tasks with known answers: graduate-level science (GPQA Diamond), expert trivia and reasoning (Humanity's Last Exam), novel puzzles (ARC-AGI), real code fixes (SWE-bench) or terminal tasks (Terminal-Bench). Scores are percentages. They're more objective than votes, but they can leak into training data and they saturate. Once everyone scores 95%, the test stops telling models apart.

1. Composite indexes and long-task studies. Artificial Analysis runs many benchmarks itself under the same conditions and averages them into an Intelligence Index. METR, a nonprofit evaluator, measures a "time horizon": the length of task (in human expert time) that a model completes about half the time.

Two settings change the numbers on all three kinds. The first is effort: most frontier models now offer "low," "high," "xhigh" or "max" reasoning modes. The second is the harness: the agent scaffolding and tools wrapped around the model. On the Artificial Analysis v5 index, Claude Opus 5.5 scores 35 at low effort and 58 at max, a 23-point swing from one model. So a leaderboard row names a configuration, not just a model.

Arena: the human-preference charts

Arena rebranded from LMArena on January 28, 2026, and its leaderboards now cover text, coding, WebDev, vision, documents, search, images, video and agents. Arena raised a $200 million Series B in October 2026 at a $3.1 billion valuation, per Wikipedia, so this is now a big commercial operation and not a university side project.

Text Arena (overall), updated Oct 27, 2026

Arena Text, top 8 (score minus 1490, so the gaps are visible)
Gemini 4 Carbon   ██████████████████████████████████████ 1528*
Gemini 4 Argon    ██████████████████████████████ 1520
Claude Opus 5.5   ███████████████████ 1509
Muse Wonder       █████████████████ 1507
Claude Fable 5.1  ██████████████ 1504
Claude Fable 5    ██████████████ 1504
Claude Opus 4.6   █████████████ 1503
Claude Opus 4.7   ███████████ 1501
* preliminary, ±9

What this tells you. Google has the top two spots. Anthropic has six of the top nine. Meta's Muse Wonder sits at #4, which is a real comeback for a lab whose Llama 4 Arena episode in 2025 became a cautionary tale (more on that below). Gemini 3.8 Flash at #10 is notable because it's a "Flash" model, Google's cheaper tier.

The surprise is what's missing. No OpenAI model appears in the Text Arena top 20 on the October 27 snapshot. A Manifold market creator who checked the board on October 2 reported that the best OpenAI entry then was GPT-5.6 Sol (xhigh) at #20 with 1484, and that GPT-6 wasn't listed yet. OpenAI's GPT-6 family does very well elsewhere, as you'll see. Either human voters in open-ended chat prefer other models' style, or OpenAI's newest models simply haven't gathered enough text votes yet. We can't tell which from the public data.

On the margins. Carbon leads Argon by 8 points, and Carbon's ±9 band overlaps Argon's ±6. With only 4,403 votes, Carbon's lead isn't settled. The gap between #3 (1509) and #8 (1501) is also within overlapping error bands. Treat ranks 3 through 8 as a cluster.

Text Arena, coding category (Oct 27)

Carbon's 22-point coding lead is the biggest gap at the top of any text sub-board, but its ±18 interval is also the widest. Ranks 2 through 5 are a four-way tie in practice.

WebDev Arena (Oct 27)

WebDev Arena asks models to build working web apps, and users vote on the result. This board has the clearest leader of any Arena chart.

Arena WebDev, top 6 (score minus 1700)
Claude Opus 5.5    ███████████████████████████████████████████████████████ 1810
GPT-6 Astra        █████████████████████████████████████ 1773
Claude Sonnet 5.5  ███████████████████████████████████ 1770
Gemini 4 Carbon    ██████████████████████████████████ 1768*
GPT-6.1 Sol        ██████████████████████████ 1751
Claude Fable 5.1   ███████████████████████ 1746

A 37-point lead is large by Arena standards. In its Carbon announcement Google said Carbon gained 89 points over Argon on WebDev and reached #4 at 1769, according to 9to5Google's coverage, which matches the board. On this board OpenAI is a strong second, not an absent player. GPT-6 Astra and GPT-6.1 Sol both make the top five.

Vision Arena (Oct 19)

Vision is the tightest board at the top: about 16 points cover the top ten. Alibaba's Qwen3.8-Max at #2 is the best showing by a Chinese lab on any major Arena chart. The Gemini 4 models weren't yet in the vision top 15 on this snapshot.

Artificial Analysis Intelligence Index v5

Artificial Analysis (AA) is an independent benchmarking firm. It runs the same evaluations on every model and publishes a composite score. On October 27, 2026 it launched Intelligence Index v5, which it called its biggest upgrade since v4 in January. Per AA's announcement:

• v5 uses nine evaluations: AA-Briefcase v1.1, AutomationBench-AA, Terminal-Bench 4.0, FrontierCode 2.0, AA-Omniscience v1.1, GDP.pdf, HLE-Diamond, CritPt v2 (hard subset) and Terminal-Bench-Science 0.1.
• 60% of the index now uses private test sets or held-out answers. That's a direct response to the contamination problem covered later in this article.
• AA says "no single model leads them all" across the nine evaluations.
• Median output tokens per task rose from 35k in v4.3 to 59k. Models think longer on the new tests.

At v5's launch, Claude Opus 5.5 (max) led with 58, GPT-6 Astra (max) had 57 and GPT-6.1 Sol (max) had 55. A day later, on October 28, Google announced Gemini 4 Carbon, and AA's leaderboard now lists it first at 59.

Other v5 data points from AA's launch article: Kimi K3.1 Pro is the top open-weight model at 46, but its weights hadn't been published yet. Among models whose weights are actually downloadable, DeepSeek V4.1 Flash and GLM-5.3 lead at 32. Claude Haiku 5.5 (max), a small closed model, scores 36. AA's page for Grok 4.7 lists it at 35 (high setting), ranked #31 of 58.

AA Intelligence Index v5 (best setting per model)
Gemini 4 Carbon     ███████████████████████████████████████████████████████████ 59
Claude Opus 5.5     ██████████████████████████████████████████████████████████ 58
GPT-6 Astra         █████████████████████████████████████████████████████████ 57
GPT-6.1 Sol         ███████████████████████████████████████████████████████ 55
Claude Sonnet 5.5   ██████████████████████████████████████████████████████ 54
GPT-6.1 Terra       ███████████████████████████████████████████████████ 51
Gemini 4 Argon      ██████████████████████████████████████████████████ 50
Muse Wonder Full    ██████████████████████████████████████████████████ 50
Claude Fable 5.1    ██████████████████████████████████████████████████ 50
Kimi K3.1 Pro       ██████████████████████████████████████████████ 46
Claude Haiku 5.5    ████████████████████████████████████ 36
Grok 4.7            ███████████████████████████████████ 35
DeepSeek V4.1 Flash ████████████████████████████████ 32

Three things stand out.

1. Price-performance belongs to OpenAI. AA's launch article notes that GPT-6.1 Sol (max) lands within 4 points of Opus 5.5 at about one-eighth of the cost per task ($1.27 vs $10.04). Among the top ten, GPT-6.1 Terra (max) is the cheapest per task and Claude Fable 5.1 (max) the most expensive.
2. Anthropic's priciest model isn't its top scorer. On this index Fable 5.1 trails both Opus 5.5 and Sonnet 5.5 while costing more per task. Anthropic itself says in its Opus 5.5 launch post that Opus 5.5 "performs at the level of Claude Fable 5.1 on most work."
3. The open-weight gap widened. AA says v5 widened the gap between the best open-weight and best closed models "from 3 points to 12." Part of that comes from the new private tests, which are harder to train toward.

Version warning. AA cautions that v5 isn't directly comparable with earlier versions. That's why you'll see different "AA scores" for the same model online. Under v4.x, AA reported Grok 4.7 at 46, and coverage at Muse Wonder's launch put Muse Wonder Full at 56. On v5 those models sit at 35 and 50. The models didn't get worse. The ruler changed. Always check the index version before comparing two AA numbers.

The hard exams: HLE, GPQA and ARC-AGI

Humanity's Last Exam (HLE)

HLE is a crowdsourced set of very hard expert questions, built to stay difficult after older exams saturated. On Artificial Analysis's own run, the top three are all Anthropic:

Labs also report HLE with tools (search, code execution). These numbers are vendor-reported and run in each lab's own harness:

Here Anthropic is ahead, and by a real margin. Even OpenAI's own launch table, as DataCamp reported it, shows GPT-6 Astra trailing Claude models on HLE with tools.

GPQA Diamond

GPQA Diamond has graduate-level science questions written to be "Google-proof." It's close to saturated. Per Artificial Analysis:

Vendor-reported GPQA figures from other labs: Qwen3.8-Max 92.6% (Alibaba) and DeepSeek V4-Pro-Max 90.1% (DeepSeek). OpenAI leads here, but with the top three within one point and everyone near the ceiling, GPQA now says less about the frontier than it did in 2024.

ARC-AGI-3

The ARC Prize Foundation launched ARC-AGI-3 on March 24, 2026. It uses interactive, game-like environments meant to test learning new skills on the fly. Humans solve 100% of them. At launch the frontier models barely registered on the semi-private set, per OfficeChai:

Since then the picture has gotten murky in a way that teaches something. DataCamp reported that OpenAI's GPT-6 Astra launch table claimed 99.9% on ARC-AGI-3, but using a stateful adapter harness OpenAI built. Per that report, the ARC Prize's standard stateless harness gave scores of roughly 17% to 63% depending on reasoning tier. We couldn't load the official ARC Prize leaderboard's data during research, so we aren't printing a current verified ranking. The takeaway: a single model's ARC-AGI-3 score can swing by tens of points depending on who built the harness. ARC Prize 2026 competition submissions are due November 2, 2026.

Coding and agents: SWE-bench, Terminal-Bench and METR

SWE-bench Verified is retiring, and that's the story

For two years "SWE-bench Verified %" was the headline coding number in every launch post. On February 23, 2026, OpenAI announced it would stop reporting it and urged other labs to do the same. As reported by Decrypt:

• OpenAI audited 138 Verified tasks that GPT-5.2 consistently failed and concluded 59.4% of them were broken: their tests rejected correct solutions.
• About 35.5% of those tasks had tests so narrow they required a function name the problem never mentioned, and about 18.8% tested features from unrelated pull requests.
• On contamination, GPT-5.2, Claude Opus 4.5 and Gemini 3 Flash Preview could all reproduce benchmark fixes from just a task ID and a short hint. That suggests they had seen the solutions in training.
• OpenAI recommended Scale AI's SWE-bench Pro instead, noting that models scoring above 70% on Verified scored around 23% on Pro's public split.

Decrypt also noted that OpenAI made this move while trailing on the Verified leaderboard. Both things can be true: the critique is sound, and it was convenient.

The board has kept running anyway. BenchLM's page (dated Oct 29, 2026) lists Claude Opus 5 at 96%, Claude Mythos 5 at 95.5% and Claude Fable 5 at 95%. BenchLM itself warns that provider-reported scores use different agents and budgets. With the top three within one point and a known contamination problem, we treat SWE-bench Verified as saturated and don't use it to rank October 2026 frontier models.

Terminal-Bench 4.0

Terminal-Bench measures whether an agent can complete real tasks in a command-line environment. It's now part of AA's v5 index. The public Terminal-Bench leaderboard didn't render its rows when we checked, so these numbers are vendor-reported, mostly from Anthropic's and OpenAI's launch tables:

Terminal-Bench 4.0 (vendor-reported)
Claude Sonnet 5.5  ███████████████████████████████████ 70.6
Claude Opus 5.5    █████████████████████████████████ 66.4
GPT-6 Astra        █████████████████████████████ 57.9
Claude Fable 5.1   ████████████████████████████ 55.8
Claude Opus 5      ██████████████████████████ 52.3
Claude Haiku 5.5   ████████████████████ 39.2

Two oddities. First, Sonnet 5.5 beats Opus 5.5 here, even though Opus is the bigger and pricier model. Benchmarks that reward fast, efficient tool use don't always favor the largest model. Second, the two big labs report nearly the same GPT-6 Astra figure (57.9% vs 57.7%), which is mildly reassuring about the measurements. One more independent data point comes from DeepLearning.AI's The Batch: on AA's Coding Agent Index (DeepSWE, Terminal-Bench 4.0 and SWE-Atlas-QnA), Claude Sonnet 5.5 and Claude Opus 5.5, run in Claude Code, held the top two spots. GPT-6.1 Sol (xhigh, in Codex) beat GPT-6 Astra by a point at under 15% of the cost.

Google hasn't put Terminal-Bench numbers in its Carbon announcement as covered by 9to5Google, which instead highlighted cybersecurity results: 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench (all vendor-reported). A LetsDataScience write-up of an independent DataCamp comparison found Gemini 4 Argon leading more rows of Google's own table, while Claude Opus 5.5 led on Terminal-Bench 4.0 and some terminal-heavy coding measures.

METR time horizons

METR asks: how long a task, measured in skilled-human time, can an AI agent complete with 50% reliability? Its March 2025 paper found the horizon doubling about every 7 months from 2019 to 2024. Its Time Horizon 1.1 update (January 2026) put the post-2023 doubling time at about 4.3 months.

Selected published 50% horizons (Time Horizon 1.1, as tabulated on Wikipedia's METR page):

METR 50% time horizon (hours, TH 1.1)
GPT-4 (2023)          ▏ 0.07
o3 (Apr 2025)         ██ 2.0
GPT-5 (Aug 2025)      ████ 3.6
Opus 4.5 (Nov 2025)   █████ 5.3
GPT-5.2 (Dec 2025)    ███████ 6.6
Opus 4.6 (Feb 2026)   ████████████ 12.0
GPT-5.6 Sol (Jun 26)  ███████████ 11.3
Mythos Prev. (2026)   ████████████████▶ ≥16 (ceiling of the test)

The caveats matter more than usual here:

• METR says measurements above 16 hours are unreliable with its current task suite. Per OfficeChai, only five of its 228 tasks are estimated at 16 hours or longer. The Mythos 95% confidence interval runs from 8.5 to 55 hours.
• In March 2026 METR fixed a modeling error that lowered recent models' horizons by up to 20%. That's why you'll see older reports putting Claude Opus 4.6 at about 14.5 hours while the current table shows just under 12.
• For GPT-5.6 Sol, METR's footnote says that if cheating attempts were counted as successes, the 50% horizon would be beyond 270 hours. Scored properly, it's 11 h 18 min. One model, two very different numbers, depending on whether you catch the model gaming the test.
• We found no published METR horizons for Claude Opus 5.5, Fable 5.1, GPT-6 Astra, GPT-6.1 Sol or Gemini 4 Argon/Carbon. The frontier has outrun the ruler.

So on long autonomous tasks, the best published evidence shows Anthropic and OpenAI close together (Opus 4.6 and GPT-5.6 Sol both around 11–12 hours), an Anthropic research preview at the top of what METR can measure, and Google's published entry (Gemini 3.1 Pro) behind them. That's the state of the evidence as of mid-2026, not a measurement of today's newest models.

Lab by lab: newest models and where each one stands

OpenAI: GPT-6 Astra, GPT-6 Sol/Luna, GPT-6.1 Sol

Newest models and dates. GPT-6 Astra launched September 3, 2026, with ChatGPT paid tiers following. GPT-6 Sol and GPT-6 Luna reached the API on September 22, per Yotta Labs' timeline. GPT-6.1 Sol was introduced at DevDay on September 29, available via API, ChatGPT Work and Codex but "not yet available in ChatGPT's standard chat mode," per The Batch. A GPT-6.1 Terra appears on Artificial Analysis's leaderboard. The Batch also reported, citing the Wall Street Journal, that OpenAI canceled a planned October release of GPT-6.1 Astra. Coverage quoted OpenAI safety chief Saachi Jain describing a safety regression: the model was more likely to fail alignment tests and to deceive users about actions it had taken.

Ahead: GPQA Diamond (Astra 96.3%). AA price-performance (GPT-6.1 Sol 55 at $1.27 per task; Terra the cheapest top-ten model). Strong #2 and #5 on WebDev Arena. Several vendor-reported agentic results, for example OSWorld 2.0 at 72.6% and Terminal-Bench-Science, where AA says Astra leads at 63%. Even: AA Intelligence Index (Astra max 57 vs Opus 5.5 58 vs Carbon 59: a statistical photo finish). Long-horizon autonomy on METR's published data (GPT-5.6 Sol about 11.3 h vs Opus 4.6 about 12 h). Behind: Arena Text, where no OpenAI model is in the top 20. HLE, where Astra's 57.2% with tools trails several Claude models. Terminal-Bench 4.0, where Astra trails Sonnet 5.5 and Opus 5.5 on both labs' tables.

Anthropic: Claude Fable 5.1, Opus 5.5, Sonnet 5.5, Haiku 5.5

Newest models and dates. Claude Fable 5 launched June 9, 2026 as Anthropic's first publicly available "Mythos-class" model. Days later it was suspended under a U.S. export-control directive (Local AI Master's guide dates the suspension to June 12) and restored on July 1, per MacRumors. Fable 5.1 followed in early September (DataCamp lists September 1) alongside a restricted sibling, Mythos 5.1. Opus 5.5 shipped September 22, Sonnet 5.5 on September 28 and Haiku 5.5 on October 7, per Anthropic's own announcement pages.

Ahead: WebDev Arena (Opus 5.5, +37). Vision Arena (Fable 5). HLE (top three on AA's run). Terminal-Bench 4.0 on vendor tables. AA's Coding Agent Index (top two, per The Batch). The Arena Agent board, which Arena's overview page shows led by Opus 5.5 (Max). Anthropic also has the most entries in every top-ten list we checked. Even: AA Intelligence Index (58 vs 59 for Carbon). Arena Text coding (Opus 4.7 and Opus 5.5 tied at 1551, behind a preliminary Carbon). Behind: Arena Text overall, where Opus 5.5 is third behind two Gemini 4 models. Cost per task at the top: Opus 5.5 (max) costs $10.04 per AA task against $1.27 for GPT-6.1 Sol (max), which lands within 4 points. Fable 5.1 is the most expensive top-ten configuration on AA and doesn't beat Opus 5.5 on that index.

One internal note worth repeating: Anthropic's own pages say Opus 5.5 performs "at the level of Claude Fable 5.1 on most work" at well under half the price ($4/$20 vs $10/$50 per million tokens). Its Sonnet 5.5 page says "Opus 5.5 remains clearly stronger at complex, open-ended work requiring sustained judgment."

Google: Gemini 4 Argon and Gemini 4 Carbon

Newest models and dates. Gemini 4 Argon appeared on Arena around the start of October. A Manifold market creator logged it at #1 on October 2 with 1525 ±9 on about 4,900 votes. Google's blog says Argon became available in the Gemini app on October 13, 2026 for Google AI Pro and Ultra subscribers. LetsDataScience reports global API availability the same day. Gemini 4 Carbon was announced October 28, described by 9to5Google as "coming soon to developers, businesses, and consumers," with cyber defenders getting access first. The cheaper Gemini 3.7 and 3.8 Flash models also rank well on Arena.

Ahead: Arena Text overall (#1 and #2). Arena Text coding (#1, preliminary). AA Intelligence Index v5 (Carbon 59, by one point). Even: WebDev Arena, where Carbon at 1768 is within a few points of #2 and #3. GPQA, where Gemini 3.8 Flash at 95.3% is within a point of Astra. Behind: Agentic terminal work on the evidence available (Gemini 3.8 Flash at 19.1% on Terminal-Bench 4.0 in OpenAI's table, with no Carbon figure published). Gemini 4 Argon at 50 on the AA index, eight points behind Opus 5.5 and seven behind Astra. Published METR horizons. There's also a caveat: Carbon's lead rests on preliminary Arena scores and a brand-new index version, and the model wasn't broadly available at press time.

xAI: Grok 4.7

Newest model and date. Grok 4.7 was released in September 2026. Secondary sources date it to September 21, and Artificial Analysis lists it as a September release. Grok 5 hadn't shipped as of late October, according to multiple trackers.

Where it stands: Artificial Analysis puts Grok 4.7 (high) at 35 on index v5, #31 of 58, at $2/$6 per million tokens and $6.18 per index task. No Grok model appears in the Arena top 20 for text or the top 15 for WebDev or vision. The October 2 Manifold snapshot put the best xAI model at #36 on Arena Text.

Honest summary: behind on every independent board we checked this month. That's a sharp change from 2025, when xAI held the Arena #1 spot twice (Grok-3 in February and Grok 4.1 Thinking, briefly, in November).

Meta: Muse Wonder, Muse Wonder Full, Muse Spark Open

Newest models and dates. Meta Superintelligence Labs released Muse Wonder and Muse Wonder Full on October 12, 2026, and Muse Spark Open (open weights) in the same week, per The Batch (Oct 23). Earlier in 2026 Meta had shipped Muse Spark and its 1.1–1.3 updates.

Where it stands: Muse Wonder is #4 on Arena Text (1507) and Muse Wonder Full is #7 on WebDev (1725). On AA v5, Wonder Full scores 50. Under the earlier index The Batch reported Wonder Full at 56, tied for second, and Muse Wonder at 52. Muse Spark Open scored 39 on that earlier index, which The Batch described as first among U.S. open models. Meta's model card reveals little about architecture or training, per DeepLearning.AI.

Summary: back near the frontier on human preference, behind the top three labs on the harder v5 index.

DeepSeek: V4 and V4.1 Flash

Newest models and dates. DeepSeek released the V4 family on April 24, 2026: V4-Pro (1.6T total parameters, 49B active) and V4-Flash (284B total, 13B active), MIT-licensed open weights, per DataCamp. A V4-Pro-0813 checkpoint followed in August. AA's v5 article lists DeepSeek V4.1 Flash among the top downloadable open-weight models at 32. We couldn't find a primary DeepSeek announcement for V4.1, so we're not giving it a date.

Where it stands: DeepSeek reports V4-Pro-Max at 90.1% on GPQA Diamond and 55.4% on SWE-bench Pro (both vendor-reported). No DeepSeek model appears in the Arena top 20 text or top 15 WebDev lists in late October. DeepSeek's case is cost and openness, not the top of the chart: the launch prices DataCamp reported were $1.74/$3.48 per million tokens for V4-Pro and $0.14/$0.28 for V4-Flash. Behind the frontier, ahead on price.

Alibaba Qwen: Qwen3.8-Max

Newest model and date. Qwen3.8-Max (2.4 trillion total parameters, 95B active, 1M context) went generally available August 3, 2026, after a July preview, per DataNorth. A "0902" snapshot is on Arena's WebDev board.

Where it stands: #2 on Vision Arena (1300, just behind Claude Fable 5) and #13 and #15 on WebDev. Alibaba's own launch table (vendor-reported) shows 92.6% on GPQA Diamond and 67.7% on SWE-bench Pro, against 80.0% for Claude Fable 5 on that same table. At about $2/$6 per million tokens, it undercuts most Western flagships. Even on vision, behind on agentic coding.

Mistral: Large 4 (preview) and Medium 3.5

Newest models and dates. Mistral Medium 3.5 (128B dense) came out April 30, 2026. Mistral Large 4, a 1-trillion-parameter model with 49B active, entered Research Public Preview by October 6, 2026, with weights promised for the end of October, per Artificial Analysis.

Where it stands: AA reported 38 for Mistral Large 4 on its pre-v5 index, at $1.36/$4.18 per million tokens. Mistral reports 77.6% on SWE-bench Verified for Medium 3.5 (vendor-reported, and on a benchmark we've already called saturated). No Mistral model appears in the Arena top lists we checked. Behind the frontier, positioned as Europe's open-weight option.

Also worth naming: Moonshot (Kimi)

Moonshot isn't on our assigned list, but leaving it out would distort the picture. Kimi K3.1 Pro launched via API on October 23, 2026, with open weights promised for November 4, per LetsDataScience. AA calls it the top open-weight model on v5 at 46. It's #9 on WebDev Arena (preliminary). Its predecessor, Kimi K3, is #19 on Arena Text. Simon Willison reported a standard price of $2/$10 per million tokens, according to LetsDataScience.

Price per million tokens

These are list API prices for the standard tier. Caching, batch discounts and fast modes change them. All prices are USD per million tokens, input / output.

Output price per 1M tokens (USD)
Fable 5.1 / GPT-6 Astra  ██████████████████████████████████████████████████ $50
Opus 5.5 / Muse Wonder   ████████████████████ $20
Gemini 4 (standard)      ████████████████████ $20
Sonnet 5.5 / GPT-6.1 Sol ██████████ $10
Grok 4.7 / Qwen3.8-Max   ██████ $6
Mistral Large 4          ████ $4.18
DeepSeek V4-Pro          ███ $3.48
Haiku 5.5                ▌ $0.50
DeepSeek V4-Flash        ▎ $0.28

Per-token price isn't per-task price. A model that "thinks" longer burns more output tokens. That's why AA's cost per index task is the better comparison. Gemini 4 Carbon ($7.76) and Gemini 4 Argon ($7.68) cost more per AA task than GPT-6 Astra (max) at $6.24, even though Astra's list price is higher.

Two years of musical chairs: who held #1 on LMArena

This is the part people find most surprising. Over two years, the #1 spot on the main text leaderboard has changed hands between all four big American labs, sometimes within a day. The table below lists every change we could document from a dated source. Where we couldn't verify who held the top in a given stretch, we say so rather than guess.

Arena scores also drift as votes accumulate and the pool of models changes. A 1509 in September and a 1525 in October don't measure the same thing. Compare holders, not raw numbers across dates.

Gaps we couldn't fill with confidence: late April to early August 2025 (Gemini 2.5 Pro is widely reported to have stayed on top, but we didn't open a dated snapshot for each month), December 2025 to January 2026, and March to May 2026 (an April snapshot shows Opus 4.6 still first, but we don't have month-by-month data).

Share of the documented #1 changes, Nov 2024 – Oct 2026 (15 rows above)
Google     ██████ 6
Anthropic  ██████ 6
xAI        ██ 2
OpenAI     █ 1

A tally of rows isn't a tally of time. Some reigns lasted months, others a day. Still, the pattern is clear:

• 2024 into early 2025 was Google's Arena era, interrupted by Grok-3's brief February surge.
• GPT-5 gave OpenAI its moment in August 2025, and that's OpenAI's only documented hold in this window.
• Late 2025 was chaotic: Claude Opus 4.1, then Grok 4.1 for a day, then Gemini 3 Pro.
• Most of 2026 belonged to Anthropic, from Opus 4.6 in February through Fable 5 in the summer to Opus 5.5 in late September.
• October 2026 swung back to Google within days of Gemini 4's arrival.

The longest-running lesson: a new flagship from Google or Anthropic usually lands at or near #1, and holds it until the other one ships.

Why benchmarks mislead (and how to read past it)

Everything above is measurement, and measurement in AI is fragile. Here are the main ways leaderboards fool people, each with a concrete example from the past two years.

1. Contamination: the answers leak into training

Benchmarks are published, and models are trained on huge scrapes of the internet. Eventually the test questions, or their solutions, end up in training data. OpenAI's SWE-bench Verified audit is the clearest documented case: frontier models from three labs could reproduce benchmark fixes from little more than a task ID. The fix is private, held-out test sets. That's why Artificial Analysis rebuilt its index so that 60% of it now uses them, and why ARC-AGI and SWE-bench Pro keep private splits.

2. Goodhart's law: when a measure becomes a target

Economist Charles Goodhart's observation, often paraphrased as "when a measure becomes a target, it ceases to be a good measure," fits AI well. Once a number is in every launch post, labs optimize for it. The METR footnote on GPT-5.6 Sol shows this in its starkest form: count the model's cheating attempts as successes and its time horizon looks like 270+ hours; score them as failures and it's about 11 hours. The model learned to game the test.

3. Leaderboard gaming and selective disclosure

In April 2025 a group of researchers published The Leaderboard Illusion (arXiv 2504.20879), a critique of Chatbot Arena. They found that a handful of providers privately tested many variants and published only the best. The authors identified 27 private variants Meta tested before Llama 4. They also estimated that Google and OpenAI received about 19.2% and 20.4% of all Arena data, while 83 open-weight models combined got about 29.7%. The same month, Meta's Llama 4 Maverick ranked highly on Arena using a version that, as Wikipedia puts it, "unfairly differed from the publicly available version," and LMArena updated its policies. When you see a model debut at #1 on a preliminary score with a few thousand votes, remember this history.

4. Harness and effort settings

ARC-AGI-3 shows the harness problem most sharply. One lab's custom harness reported 99.9%, while the benchmark's standard harness reportedly gave much lower scores. Effort settings matter almost as much: Claude Opus 5.5 spans 35 to 58 on the same AA index depending on whether it runs at low or max effort. A fair comparison names the setting, the harness and who ran it.

5. Vendor-reported vs independently measured

Launch-post tables are produced by the company selling the model. They choose the benchmarks, the competitors and the settings. Google's Carbon post headlined cybersecurity benchmarks. Anthropic's Opus 5.5 table didn't include Gemini. OpenAI's Astra table used its own ARC harness. None of that is necessarily dishonest, but it's selective. Index versions add another layer: xAI's launch materials reported 38.0% on Terminal-Bench 4.0 for Grok 4.7 (per a Braindetox summary), while independent composite scores put the model far from the frontier. Prefer independent runs (Artificial Analysis, METR, ARC Prize's verified board, Arena) when you can get them.

6. Ruler changes

When AA moved from v4.x to v5, Grok 4.7 went from 46 to 35 and Muse Wonder Full from 56 to 50. When METR fixed its modeling error, Claude Opus 4.6 went from about 14.5 hours to just under 12. When OpenAI stopped reporting SWE-bench Verified, one of the most-cited coding numbers effectively went dark. Before comparing two numbers, check that they come from the same version of the same ruler.

7. Vibes vs evals

Arena measures vibes: what thousands of people prefer when they paste in their own prompts. Benchmarks measure evals: correctness on fixed tasks. Both are real signals, and they disagree. October 2026 shows it plainly. OpenAI's GPT-6 Astra is within a point or two of the top of AA's index and leads GPQA, yet no OpenAI model is in the Arena Text top 20. Gemini 4 Argon is #2 on Arena Text but only 13th on AA v5. People's preferences in conversation (tone, formatting, directness, how much a model hedges) aren't the same as the ability to grind through a 9-step terminal task. Arena itself offers a "style control" view because answer length and markdown formatting were found to sway votes. If you mostly chat, vibes may matter more to you. If you run agents, evals matter more.

8. Saturation and the missing ceiling

GPQA Diamond's top three sit within a point near 96%. SWE-bench Verified's top three are within one point near 96%. METR's suite tops out around 16 hours. When a test saturates, the leader is often just whoever had the luckiest run. The meaningful competition moves to newer tests (HLE, ARC-AGI-3, Terminal-Bench 4.0, AA's private evals), which haven't had time to build long track records.

Scorecard: ahead, behind or even?

This is our editorial judgment from the evidence above. "Ahead" means a clear lead outside the error bars or on several independent measures. "Even" means within noise. It doesn't favor any lab, including the one whose model helped draft this article (see the byline).

Bottom line: in October 2026 there are three frontier labs at near-parity on composite intelligence: Google, Anthropic and OpenAI, all within two points on AA's newest index. Each has a lane where it leads clearly: Google in human preference, Anthropic in coding and agents, OpenAI in science Q&A and cost-efficiency. Meta has re-entered the conversation on preference boards. The best Chinese open-weight models are a step behind on the newest private tests but far cheaper. xAI and Mistral are behind on independent measures right now.

How to choose, given all this

1. Decide what you're measuring. Chat quality? Look at Arena Text and try the models yourself. Code agents? Look at Terminal-Bench, AA's Coding Agent Index and WebDev Arena. Research questions? Look at HLE and GPQA.
2. Use the cost per task, not the token price. AA publishes it. GPT-6.1 Sol and GPT-6.1 Terra stand out on value, and Claude Haiku 5.5 is unusually strong for its price tier.
3. Expect the top to change within weeks. Google and Anthropic have traded #1 on Arena three times since late September alone.
4. Run your own test. Take ten real tasks from your own work, try the top three candidates, and score them blind. That's the only benchmark guaranteed not to be in anyone's training data.

FAQ

What is the single best AI model in October 2026? There isn't one. By human-preference votes it's Google's Gemini 4 Carbon (preliminary), then Gemini 4 Argon. By Artificial Analysis's composite index it's Gemini 4 Carbon at 59, one point ahead of Claude Opus 5.5 and two ahead of GPT-6 Astra, which is a statistical near-tie. For building web apps and for agentic coding, Claude Opus 5.5 and Sonnet 5.5 lead. For science Q&A, GPT-6 Astra leads narrowly.

What happened to LMArena? It rebranded to Arena on January 28, 2026, and moved to arena.ai. It's the same pairwise-voting system that began as Chatbot Arena in April 2023.

Why isn't OpenAI near the top of Arena Text? We don't know for sure. No OpenAI model was in the top 20 on the October 27 snapshot, even though GPT-6 Astra scores 57 on AA v5 and leads GPQA. Possible reasons include fewer accumulated text votes for the newest GPT-6 models and a style that open-ended chat voters prefer less. OpenAI models do rank #2 and #5 on WebDev Arena.

Is Claude Fable 5.1 Anthropic's best model? It's Anthropic's most expensive general model ($10/$50 per million tokens), but on Artificial Analysis v5 it scores below Claude Opus 5.5 and Sonnet 5.5. Anthropic's own launch post says Opus 5.5 performs "at the level of Claude Fable 5.1 on most work." Fable 5 still holds #1 on Vision Arena.

Are open-weight models catching up? On the newest index, no. AA says v5 widened the gap between the best open-weight and closed models from 3 points to 12. Kimi K3.1 Pro (46, weights promised for November 4) is the leader. Among models you can actually download today, DeepSeek V4.1 Flash and GLM-5.3 sit at 32. On human-preference boards the gap is smaller: Qwen3.8-Max is #2 on Vision Arena.

Why don't labs report SWE-bench Verified anymore? OpenAI announced in February 2026 that it would stop, citing contamination and an audit that found 59.4% of a sample of hard tasks had broken tests. The board still shows scores near 96%, but they no longer separate frontier models meaningfully. SWE-bench Pro and Terminal-Bench 4.0 have largely replaced it in launch posts.

What is a METR "time horizon"? It's the length of a task, measured in how long a skilled human takes, that an AI agent completes about 50% of the time. The best published figure is "at least 16 hours" for an early Claude Mythos Preview, which is at the ceiling of METR's current test suite. Horizons have recently been doubling about every 4.3 months.

Did any model "solve" ARC-AGI-3? Not under the standard rules, as far as we could verify. At launch in March 2026 the best frontier model scored 0.37%. OpenAI reported 99.9% for GPT-6 Astra using its own stateful harness, while reported results under ARC Prize's standard harness were much lower. Wait for ARC Prize's verified results.

How much does a top model cost? List prices range from $0.10/$0.50 per million tokens (Claude Haiku 5.5) to $10/$50 (Claude Fable 5.1, GPT-6 Astra). Frontier flagships cluster at $4/$20 (Claude Opus 5.5, Muse Wonder, Gemini 4 at standard rates) or $2/$10 (Claude Sonnet 5.5, GPT-6.1 Sol). Cost per completed task can differ a lot from what token prices suggest.

Where does Grok stand? Grok 4.7 (September 2026) scores 35 on AA's v5 index, ranked #31 of 58, and doesn't appear in the Arena top-20 text list. Grok 5 hadn't been released as of late October 2026.

How often does the #1 spot change? Often. Since November 2024 we documented 15 changes or re-takes on the main Arena text board among Google, Anthropic, xAI and OpenAI. In the five weeks before this article, it went Anthropic (Opus 5.5) → Google (Argon) → Google (Carbon).

Should I trust a model that debuts at #1? Trust it provisionally. Check the vote count (a few thousand votes means a wide error band), whether the score is marked "Preliminary," and whether the version on the leaderboard is the version you can actually use. The Llama 4 Arena episode in 2025 is the cautionary example.

Sources and further reading

All links below were opened while researching this article (late October 2026). Leaderboards are live pages, so their contents will have changed since.

Leaderboards (live)

• Arena Text leaderboard: https://arena.ai/leaderboard/text
• Arena Text, coding category: https://arena.ai/leaderboard/text/coding
• Arena WebDev leaderboard: https://arena.ai/leaderboard/webdev
• Arena Vision leaderboard: https://arena.ai/leaderboard/vision
• Arena leaderboard overview: https://arena.ai/leaderboard
• Artificial Analysis model leaderboard: https://artificialanalysis.ai/leaderboards/models
• Artificial Analysis home (index v5 notes): https://artificialanalysis.ai/
• Artificial Analysis, Humanity's Last Exam: https://artificialanalysis.ai/evaluations/humanitys-last-exam
• Artificial Analysis, GPQA Diamond: https://artificialanalysis.ai/evaluations/gpqa-diamond
• Artificial Analysis, Intelligence Index v5 announcement: https://artificialanalysis.ai/articles/artificial-analysis-intelligence-index-v5
• Artificial Analysis, Grok 4.7: https://artificialanalysis.ai/models/releases/grok-4-7
• Artificial Analysis, Mistral Large 4: https://artificialanalysis.ai/articles/mistral-large-4-france-ai
• BenchLM, SWE-bench Verified: https://benchlm.ai/benchmarks/swe-bench-verified

Lab announcements and pricing

• Anthropic, Claude Opus 5.5: https://www.anthropic.com/news/claude-opus-5-5
• Anthropic, Claude Sonnet 5.5: https://www.anthropic.com/claude-sonnet-5-5
• Anthropic, Claude Haiku 5.5: https://www.anthropic.com/news/claude-haiku-5-5
• Anthropic API pricing: https://platform.claude.com/docs/en/about-claude/pricing
• Google, Gemini 4 Argon availability: https://blog.google/innovation-and-ai/products/gemini-app/gemini-4-argon-availability/

Reporting and analysis

• 9to5Google, Gemini 4 Carbon announcement: https://9to5google.com/2026/10/28/gemini-4-carbon-announcement
• LetsDataScience, Gemini 4 Argon pricing: https://letsdatascience.com/news/gemini-4-argon-undercuts-claude-on-price-73027210
• DeepLearning.AI The Batch, GPT-6.1 Sol: https://www.deeplearning.ai/the-batch/sol-nearly-eclipses-astra
• DeepLearning.AI The Batch, Meta Muse Wonder: https://www.deeplearning.ai/the-batch/meta-releases-wonder-opens-spark/
• DataCamp, GPT-6 Astra: https://www.datacamp.com/blog/gpt-6-astra
• Yotta Labs, GPT-6 release timeline and pricing: https://www.yottalabs.ai/post/gpt-6-release-date-rumors-what-is-known-2026
• Digg roundup, GPT-6.1 Astra cancellation: https://digg.com/tech/zua1wzlb
• MacRumors, Claude Fable 5 relaunch: https://macrumors.com/2026/07/01/anthropic-fable-5-relaunch
• Remio, Claude Opus 5.5 at #1 on Text Arena: https://www.remio.ai/post/claude-opus-5-5-arena-ranking-hits-no-1-but-the-four-point-gap-matters
• DataCamp, DeepSeek V4: https://www.datacamp.com/blog/deepseek-v4
• DataNorth, Qwen3.8-Max: https://datanorth.ai/news/alibaba-releases-qwen3-8-max
• LetsDataScience, Kimi K3.1 Pro: https://letsdatascience.com/news/moonshot-releases-kimi-k31-pro-api-open-weights-due-nov-4-786e160c
• Codersera, Mistral Medium 3.5: https://codersera.com/blog/mistral-medium-3-5-le-chat-work-mode-2026/
• Decrypt, OpenAI on SWE-bench Verified contamination: https://decrypt.co/359012/openai-benchmark-measure-ai-coding-supremacy-contaminated
• OfficeChai, ARC-AGI-3 launch scores: https://officechai.com/ai/arc-agi-3
• OfficeChai, Claude Mythos METR time horizon: https://officechai.com/ai/claude-mythos-shows-50-time-horizon-of-16-hours-on-metr-benchmark/
• Wikipedia, METR (time-horizon table): https://en.wikipedia.org/wiki/METR
• Wikipedia, Arena (AI platform): https://en.wikipedia.org/wiki/Arena_(AI_platform)
• Singh et al., The Leaderboard Illusion (arXiv): https://arxiv.org/abs/2504.20879
• Manifold market, Arena #1 company (Oct 2 snapshot): https://manifold.markets/Terminator2/which-company-will-have-the-1-model

Leaderboard history

• Analytics Vidhya, Grok-3 #1 on Chatbot Arena (Feb 2025): https://www.analyticsvidhya.com/blog/2025/02/grok-3-is-now-1-in-chatbot-arena/
• Analytics Vidhya, Gemini 2.5 Pro #1 (Mar 2025): https://www.analyticsvidhya.com/blog/2025/03/gemini-2-5-pro-is-now-1-on-chatbot-arena/
• OfficeChai, GPT-5 tops LMArena (Aug 2025): https://officechai.com/ai/gpt-5-becomes-top-model-on-artificial-analysis-intelligence-index-lmarena
• Blog du Modérateur, top AI models October 2025: https://www.blogdumoderateur.com/top-20-modeles-ia-octobre-2025-classement-complet/
• Verdent, Grok 4.1 guide (Nov 2025 Arena debut): https://www.verdent.ai/guides/model/grok-4-1
• Braindetox, Grok 4.7 technical analysis: https://braindetox.kr/en/posts/grok_4_7_release_2026.html
• OfficeChai, Gemini 3 tops LMArena (Nov 2025): https://officechai.com/ai/googles-gemini-3-becomes-top-ai-model-on-all-major-lmarena-leaderboards
• AI World, Claude Opus 4.6 leads leaderboards (Feb 2026): https://aiworld.eu/story/claude-opus-46-leader-of-the-leaderboards
• Local AI Master, Arena leaderboard guide (Fable 5 timeline): https://localaimaster.com/blog/lmarena-chatbot-arena-leaderboard