Data snapshot · October 2026

There Is No Single
No. 1 in AI Anymore

As of October 2026, the global AI race has fragmented. Token volume, agent reliability, hard reasoning, coding, and enterprise revenue now point to different leaders — and every model's strengths are highly context-dependent. For US buyers, the question is no longer “who tops the board,” but which trade-off between task risk, cost sensitivity, and ecosystem lock-in actually fits the job.

188T Weekly tokens across all global AI models
93.5T Chinese models' weekly tokens — 26 straight weeks above US models
33.6T DeepSeek V4.1-Flash weekly tokens, +801% week over week
70% Kimi K3.1 Pro's top score on Humanity's Last Exam
Before & Now

From Transformer to a Fragmented Leaderboard

Everything traces back to a 2017 Google paper. In under a decade, the race moved from parameter counts to reasoning and tool use — and by 2026, to leadership that depends entirely on which dimension you measure.

2017
The Transformer architecture

Google publishes “Attention Is All You Need,” the foundation for every mainstream model that followed.

2018
GPT-1 and the pretrain–finetune paradigm

OpenAI brings the pretrain-then-finetune approach to language models.

2020
GPT-3: scale produces emergent capability

175B parameters show what scaling can do — and set off the modern model race.

2023
GPT-4 goes multimodal

Models move from text-only generation into multimodal understanding and output.

2025
DeepSeek-R1 and the GPT-5 line

DeepSeek-R1 pushes the reasoning-model approach into the frontier; GPT-5 ships in August 2025, as leading labs shift from single-shot generation toward reasoning plus tool use.

H2 2025
Claude wins the developer mindshare

Claude builds a reputation among developers for coding and enterprise workflows — the groundwork for agentic coding's commercial payoff.

2026
Rotating crowns become the norm

Grok, Gemini, GPT, and Claude take turns topping different boards. Capability divergence has outgrown any single leaderboard.

Leaderboards · October 2026

Who Leads Which Board

“No. 1” means nothing without a dimension. Here is who leads on token volume, agent reliability, general and coding ability, and the hardest reasoning benchmark.

Token Volume DeepSeek V4.1-Flash 33.6T tokens per week
Agents Claude Opus 5.5 Debuts at No. 1
General / Coding Gemini 4 Argon ELO 1525 / 1562
Hard Reasoning (HLE) Kimi K3.1 Pro 70% — top score

Token Volume: DeepSeek Dominates as Chinese Models Sweep the Top Five

OpenRouter, week of October 19–25, 2026

Weekly tokens: Chinese vs. US models
Chinese models
93.54T
US models
22.47T
Weekly tokens per model
DeepSeek V4.1-Flash-1017
33.6T
GPT-6 Luna (No. 6)
8.35T
Claude Opus 5.5 (No. 7)
8.27T
Global weekly volume hit 188 trillion tokens. The top five were all Chinese models — DeepSeek (an earlier version), MiniMax M3.1 Flash, Zhipu GLM 5.3 Flash, and Tencent Hunyuan Hy4 preview.
Volume is not revenue. On OpenRouter's spend-share board for general tasks such as classification, Claude Opus 5.5 holds the top spot at 10.3%, with GPT-5.6 Sol and Gemini 3.1 Pro second and third. The only Chinese model on that board is DeepSeek V4.1 Flash. Chinese labs are winning on volume and price per token; US labs still hold the high ground on unit value.

Agents: Claude Opus 5.5 Debuts at No. 1

LMArena agent board, week 41 (October 5–11)

No. 1Claude Opus 5.5 (Max)
15.53%Net improvement
17.29%Task confirmation rate
0.37%Tool hallucination rate
Agents are graded on multi-step plan coherence, tool-call reliability, and error recovery — not single-turn answer quality. That is precisely the dimension US enterprise buyers care about when they move AI from pilots into production workflows.

General & Coding: Gemini 4 Argon Leads, Claude Close Behind

LMArena general and coding boards, ELO scores

General board
Gemini 4 Argon High
1525
Claude Opus 5.5 High
1507
Coding board
Gemini 4 Argon High
1562
Claude Fable 5 High
1551
Claude Opus 5.5 High
1550
On the general board, Gemini 4 Argon and Claude Opus 5.5 sit within the margin of error. On coding, Claude Opus 5.5 jumped seven spots into the top five. Released in October 2026, Gemini 4 Argon ranks first on Google's Vals index — which covers finance, coding, legal, and tax workflows — with a 14% accuracy gain over Gemini 3.1.

Hard Reasoning (HLE): Kimi K3.1 Pro Tops at 70%

Humanity's Last Exam leaderboard, October 23, 2026

Kimi K3.1 Pro
70%
Claude Fable 5.1
65%
Claude Opus 5
64.7%
GPT-5.4 Pro
58.7%
Kimi K3 uses a 2.8-trillion-parameter sparse MoE architecture, strong on agentic tasks and long-context retrieval. Leading HLE signals it has reached the first tier on knowledge-intensive reasoning — a notable milestone for a Chinese lab on a benchmark designed in the West.
Strengths & Weaknesses

The Model Scorecard

There is no all-around champion. Each model's strengths and blind spots follow directly from the technical and product bets its maker chose.

Claude

Anthropic · US

The benchmark for agent reliability and enterprise workflows

Strengths

  • Instruction fidelity: given a 2,000-word system prompt with 15 constraints, it follows all of them
  • Long-form writing that reads naturally — varied sentence structure, real transitions, tonal control
  • Cross-file reasoning over large codebases; 1M context can hold an entire repo
  • Well-calibrated on uncertainty: admits what it doesn't know instead of confidently making things up

Weaknesses

  • Slower than GPT-5.4 on long-generation tasks
  • Tool calls less finely tuned than GPT's
  • No native image generation

GPT-5 Series

OpenAI · US

Speed and structured output, with the most mature tool ecosystem

Strengths

  • Faster response than Claude and Gemini at comparable quality
  • Well suited to real-time chat, autocomplete, and inline suggestions
  • Fastest function calling and the deepest tool-calling ecosystem

Weaknesses

  • Model quality is not the problem — product conversion lags Anthropic
  • Less precise than Claude on complex, constraint-heavy prompts
  • Still catching up on product with Codex and ChatGPT Work in 2026

Gemini

Google · US

Multimodal, long context, and unmatched distribution

Strengths

  • Best-in-class combination of multimodality and ultra-long context
  • Costs roughly 1/7.5 of Claude — the price-performance benchmark
  • Embedded in Search, Android, Workspace, and Google Cloud — a distribution network no rival can match
  • Gemini 4 Argon ranks first on the Vals index for high-stakes professional work

Weaknesses

  • Instruction fidelity on complex prompts still trails Claude
  • Slower than GPT in some workloads
  • If frontier capability stays half a step behind, it will struggle to win the most profitable core enterprise workflows

DeepSeek

DeepSeek · China

Price-performance and math reasoning, taken to the extreme

Strengths

  • Blended API cost roughly 1% of GPT-5.5
  • Cached input as low as ¥0.2 per million tokens (about $0.03)
  • Math reasoning score of 78.95 — the highest of any model tested
  • No. 1 in global token volume; V4 API gross margins at 70–80%

Weaknesses

  • Complex reasoning slightly behind GPT
  • Thin agent capabilities — can't autonomously generate mind maps or slide decks
  • Severe capacity crunches at peak hours
  • Shifting from aggressive low pricing toward price hikes and peak/off-peak tiers

Qwen3.8-Max

Alibaba · China

The standard-bearer for overall quality and open-weight ecosystems

Strengths

  • No. 1 on SuperCLUE overall (71.48), a Chinese-language benchmark
  • Clear lead on hallucination control and agentic task planning
  • 2.4T-parameter MoE with up to 1M context
  • Flagship Max model open-sourced for the first time; inference cost down to roughly 1/8 of industry norms

Weaknesses

  • Long latency per request — sits in the low-efficiency tier

Kimi

Moonshot AI · China

Technical credibility in agentic coding and frontier reasoning

Strengths

  • No. 1 on SuperCLUE's agentic coding sub-score (75.79)
  • Strong long-document parsing and cross-domain research reports
  • Kimi K3.1 Pro tops HLE at 70%
  • 2.8T-parameter sparse MoE with strong long-context retrieval

Weaknesses

  • Slowest per request — up to 3× the fastest model
  • Some technical claims criticized as detached from engineering reality

Doubao

ByteDance · China

Distribution is the moat

Strengths

  • 382M monthly active users as of June 2026 — No. 1 among China's AI-native apps
  • Deeply integrated into Douyin (TikTok's Chinese sibling) and CapCut
  • Effectively a content-pipeline machine

Weaknesses

  • Very weak on deep logic and coding
  • Limited openness and tooling breadth
  • Whether scale translates into deep engagement and durable paid conversion remains unproven

Zhipu GLM

Zhipu AI · China

Long-horizon agents and government/enterprise compliance

Strengths

  • Among China's leaders on autonomous long-horizon agent task completion
  • Compliance posture is its most durable moat for state and enterprise buyers
  • GLM 5.3 Flash ranked No. 4 on the token-volume board

Weaknesses

  • Consumer-facing experience feels flat
  • High price floor (Lite ¥49/month, ~$7; Max ¥469/month, ~$66)
Roadmap & Outlook

Where Each Lab Is Heading

The next phase will be decided by who runs the full chain — capability → productization → durable willingness to pay — more completely.

OpenAI: From model leader to product reset

US

All-in on agents and Codex.

  • Openly acknowledges it lost ground on product — ChatGPT's explosive growth let the company underweight coding and the enterprise market. “The model led the leaderboard, but it did not convert into good enough real products.”
  • Codex replaces ChatGPT as the product spine. Sora and the standalone Atlas browser are scaled back; compute and teams consolidate around Codex, whose agent capabilities are folded into ChatGPT Work.
  • The GPT-5.6 line splits into three tiers — Sol (complex agent workflows), Terra (balanced), and Luna (cost-first) — a shift from one model for everything to task-tiered offerings.
  • OpenAI's annualized revenue is around $40B, versus more than $65B for Anthropic.
Key variable: not whether the next model wins back the benchmark crown, but whether OpenAI can turn capability into a workflow product enterprises keep paying for. The “personal AGI” vision — chips, data centers, wearable hardware, humanoid robots — is expansive, but near-term position depends on whether Codex and the agent line can stop the enterprise slide.

Anthropic: Agentic coding has been proven out

US

The narrowest bet, the most complete execution — but capex looms.

  • Focused from the architecture level on “controllable, logical, strictly rule-following” behavior, with a bet on coding. RLVR (reinforcement learning with verifiable rewards) delivered its sharpest gains exactly there, and Opus 4.5 produced the step-change in agentic coding.
  • H1 revenue of $16.29B, up 1,300% year over year and roughly 3.6× all of 2025; Q2 operating profit turned positive with a 16.4% adjusted net margin.
  • Opus 5.5 cut per-task cost about 40% versus its predecessor; enterprise adoption hit 34.4%, overtaking OpenAI's 32.3% for the first time.
  • Cloud commitments in the $518B–$600B range against roughly $80.7B in H1 net losses.
Key variable: Anthropic has shown the unit economics can work — lower cost per task plus deeper enterprise spend. But whether revenue can grow fast enough to cover an unprecedented compute buildout is the core question facing every frontier lab, not just this one.

Google Gemini: Distribution is the moat

US

Gemini 4 Argon closes the high-value professional gap.

  • Google isn't trying to win every benchmark. It's making Gemini the default AI layer across every surface it owns: Search, Android, Workspace, Gmail, Docs, Chrome, YouTube, and Google Cloud.
  • Gemini 3.1 Pro established the price-performance benchmark with top-tier long-context capability at roughly 1/7.5 of Claude's cost.
  • Gemini 4 Argon pushes into the highest-stakes, highest-priced professional work — finance, legal, coding, cybersecurity, and scientific research — striking directly at the enterprise revenue that Anthropic and OpenAI depend on.
Central tension: breadth of distribution versus sharpness of frontier capability. If Argon keeps proving reliable in high-risk professional settings, Google's distribution advantage converts into enterprise API and subscription revenue. If frontier capability stays half a step behind, distribution still guarantees reach — but likely not the largest share of the most profitable core enterprise workflows.

Chinese labs: volume lead secured, the pricing leap is the hard part

China

The volume milestone is real. Converting it into revenue is not.

  • 2026 marked a shift from chasing version numbers to defining distinct paths: DeepSeek wins volume on price and math, Qwen bridges open weights and cloud services, Kimi builds credibility in agentic coding and frontier reasoning, Doubao owns the consumer entry point, and Zhipu anchors government and enterprise compliance.
  • Training costs run roughly 1/10 of US frontier labs, and API prices sit at 10–20% of comparable US models — a structural cost advantage rather than loss-leader subsidy.
  • But volume has not converted into pricing power. Average Chinese token prices rose from ¥6.8 in H2 2025 to ¥13.46 in 2026 — nearly double — while typical US models average about ¥117.8, still roughly 8.8× higher.
  • Raising prices improves margins (DeepSeek's V4 API already runs at 70–80% gross margin) but tests user retention and the durability of volume.
Three paths that will decide the outcome: whether DeepSeek can raise prices without losing its volume base and close its agent gap; whether Qwen can convert open weights and cloud depth into enterprise revenue; whether Doubao can turn 382M monthly actives into deep engagement and paid conversion; and whether Kimi and Zhipu can build sustainable businesses in their technical and compliance niches.
Conclusion

The Bottom Line

As of October 2026, global AI competition has entered a phase of multi-dimensional leadership and full-stack execution. No model dominates every board, and the question for buyers is no longer “who is No. 1” but which trade-off between task risk, cost sensitivity, and ecosystem lock-in fits the job.

US frontier labs retain a clear edge in enterprise revenue, agent reliability, and high-value professional work. Chinese labs have built genuine structural strength in token volume, price-per-token, and open-weight ecosystems — but the gap remains structural on three axes: revenue per token, penetration into core enterprise workflows, and premium pricing for frontier capability.

The next phase belongs to whoever runs the full chain more completely: capability → productization → durable willingness to pay.