OpenAI
US- Flagship
- GPT-6 Astra
- Weights
- Closed
- Chat app
- ChatGPT
- AA Index
- 53
Reads websites with GPTBot, OAI-SearchBot, ChatGPT-User
Checked
- Kimi K3.1 Pro Kimi
- Decisions API, general availability OpenAI
- mammoth-1 and elephant-1 Base Labs
- DeepSeek V4.1 Flash 1017 DeepSeek
- Perplexity Decider v1.2 Perplexity
- d1-ultra Liquid AI
- Muse Spark Open Meta
- Gemini 4 Argon, open to all Google
- Muse Wonder Meta
- Muse Wonder Full Meta
- Microsoft-Decision-1 Microsoft
- Claude Haiku 5.5 Anthropic
- Karotte (RL environment framework) Preference Model
- Decisions API beta OpenAI
- Mistral Large 4 Mistral
- Nano Banana 2.1 Google
- Clef and Clef-flash Cloudflare
- Strands Decider 2B Amazon
- Gemini 4 Argon Google
- GPT-6.1 Sol OpenAI
- Claude Sonnet 5.5 Anthropic
- Claude Opus 5.5 Anthropic
- GPT-6 Luna OpenAI
- Grok 4.7 xAI
- MiMo-V2.6-Flash MiMo
- MiMo-V2.6-Pro MiMo
- Jev TypeSafe AI
- Kimi K2.8 Preview Kimi
- GPT-6 Astra OpenAI
- Gemini 3.8 Flash Google
- Muse Spark 1.3 Meta
- Claude Fable 5.1 Anthropic
- Claude Mythos 5.1 Anthropic
Open research · Large language models
LLM Watch: the AI models that write, read and cite your content
We track the 11 major AI companies, their current models, and how well each model writes. Use it to pick a model for content work. Use it to see which AI tools read your website.
| Rank | Company | Bar | Elo |
|---|---|---|---|
| OpenAIGPT-6 Astra | 2,174.0 | ||
| AnthropicClaude Fable 5.1 | 2,162.7 | ||
| MoonshotKimi K3 | 2,083.5 | ||
| Z.aiGLM-5.3 | 2,076.7 | ||
| xAIGrok 4.7 | 2,012.2 | ||
| MetaMuse Wonder | 2,007.0 | ||
| QwenQwen3.8 (2.4T-A95B) | 1,843.9 | ||
| GoogleGemini 4 Argon | 1,758.0 | ||
| DeepSeekDeepSeek V4 Flash | 1,559.4 | ||
| XiaomiMiMo-V2.5-Pro | 1,493.1 | ||
| MistralMistral Medium 3.1 | 1,475.4 |
Elo score. Scale starts at 1,400. Gemini 4 Argon is not tested.
Releases · Sep 1 to Oct 23
33
Notable model releases since September 1, from the companies on this page. Last month's best model is often not this month's.
Context window
1M tokens
Most flagship models now read about 1 million tokens in one request. That is about 750,000 words. Grok 4.7 reads 500,000.
Price spread · input
100×
GPT-6 Astra costs $10 for each 1M input tokens. GPT-6 Luna costs $0.10. Pick the tier for the task.
Clicks with an AI summary
8% vs 15%
Google visits with an AI summary got half the result clicks. Pew Research Center
Release log
33 since Sep 1- Kimi K3.1 Pro Kimi Available
- Decisions API, general availability OpenAI Available
- mammoth-1 and elephant-1Base Labs Available
- DeepSeek V4.1 Flash 1017 DeepSeek Available
- Perplexity Decider v1.2 Perplexity Available
- d1-ultra Liquid AI Available
- Muse Spark Open Meta Available
- Gemini 4 Argon, open to all Google Available
- Muse Wonder Meta Available
- Muse Wonder Full Meta Limited access
- Microsoft-Decision-1Microsoft Available
- Claude Haiku 5.5 Anthropic Available
- Karotte (RL environment framework)Preference Model Available
- Decisions API betaOpenAI Preview
- Mistral Large 4 Mistral Preview
- Nano Banana 2.1Google Available
- Clef and Clef-flashCloudflare Available
- Strands Decider 2BAmazon Available
- Gemini 4 Argon Google Available
- GPT-6.1 Sol OpenAI Available
- Claude Sonnet 5.5 Anthropic Available
- Claude Opus 5.5 Anthropic Available
- GPT-6 Luna OpenAI Available
- Grok 4.7 xAI Available
- MiMo-V2.6-Flash MiMo Available
- MiMo-V2.6-Pro MiMo Available
- JevTypeSafe AI Preview
- Kimi K2.8 Preview Kimi Preview
- GPT-6 Astra OpenAI Available
- Gemini 3.8 Flash Google Available
- Muse Spark 1.3 Meta Available
- Claude Fable 5.1 Anthropic Available
- Claude Mythos 5.1 Anthropic Limited access
General intelligence · best model
AA Index| Rank | Company | Bar | Index |
|---|---|---|---|
| AnthropicClaude Opus 5.5 (max) | 58 | ||
| MetaMuse Wonder Full (max) | 56 | ||
| KimiKimi K3.1 Pro | 55 | ||
| OpenAIGPT-6 Astra (max) | 53 | ||
| GoogleGemini 4 Argon (high) | 53 | ||
| xAIGrok 4.7 (xhigh) | 46 | ||
| MiMoMiMo-V2.6-Pro | 46 | ||
| DeepSeekDeepSeek V4.1 Flash 1017 (max) | 45 | ||
| QwenQwen3.8 Max (0902) | 45 | ||
| GLMGLM-5.3 (max) | 45 | ||
| MistralMistral Large 4 | 38 |
Source: Artificial Analysis Intelligence Index (checked October 26, 2026). Moonshot, Qwen and Z.ai have no model in its top 25.
Why this matters for a business. These are the models behind ChatGPT, Claude, Gemini, Meta AI and Grok. When a customer asks one of them for a business like yours, the model writes the answer. It names the businesses that it can read clear facts about.
Get a free AI visibility check01 · Players
The 11 companies that make the major models
Five are in the United States, five are in China, and one is in France. Closed models run only on the company's own service. Open-weight models are free to download and run on your own servers.
Anthropic
US- Flagship
- Claude Opus 5.5
- Weights
- Closed
- Chat app
- Claude
- AA Index
- 58
Reads websites with ClaudeBot, Claude-SearchBot, Claude-User
- Flagship
- Gemini 4 Argon
- Weights
- Closed
- Chat app
- Gemini
- AA Index
- 53
Reads websites with Googlebot (AI Overviews use the Search index); Google-Extended is a control token
xAI (SpaceXAI)
US- Flagship
- Grok 4.7
- Weights
- Closed
- Chat app
- Grok
- AA Index
- 46
Reads websites with GrokBot
Meta
US- Flagship
- Muse Wonder Full
- Weights
- Mixed
- Chat app
- Meta AI
- AA Index
- 56
Reads websites with Meta-ExternalAgent
DeepSeek
China- Flagship
- DeepSeek V4-Pro
- Weights
- Open
- Chat app
- DeepSeek
- AA Index
- 45
Alibaba (Qwen)
China- Flagship
- Qwen3.8-Max
- Weights
- Mixed
- Chat app
- Qwen Chat
- AA Index
- 45
Moonshot AI (Kimi)
China- Flagship
- Kimi K3.1 Pro
- Weights
- Open
- Chat app
- Kimi
- AA Index
- 55
Z.ai (GLM)
China- Flagship
- GLM-5.3
- Weights
- Open
- Chat app
- Z.ai chat
- AA Index
- 45
Mistral AI
France- Flagship
- Mistral Large 4
- Weights
- Open
- Chat app
- Le Chat
- AA Index
- 38
Reads websites with MistralAI-User
Xiaomi (MiMo)
China- Flagship
- MiMo-V2.6-Pro
- Weights
- Open
- Chat app
- Not listed
- AA Index
- 46
Do these models name your business?
We ask leading AI answer engines what your customers ask. Then we show you who they name.
Get a free AI visibility check →02 · Models
Current models, by company
Select a company. Prices are list prices in US dollars for each 1 million tokens, input then output. A token is about three quarters of a word.
OpenAI
Three GPT-6 tiers share a 1.05M-token context window. Astra is the most capable. GPT-6.1 Sol gives near-Astra results at one fifth of the price. Luna is the low-cost tier.
Sources: OpenAI API models · LLM Releases tracker (dates)
Anthropic
Four public Claude models, all with a 1M-token context window. Anthropic recommends Opus 5.5 for most work and Fable 5.1 for the most demanding work. Claude Mythos 5.1 is not open to the public.
Sources: Claude models overview · Anthropic news
Gemini 4 Argon arrived on September 30 for a small group. Since October 13, everyone can use it in the paid Gemini API, AI Studio and the Gemini app (Pro and Ultra). The low price ends on November 30. Gemini 3.1 Pro is still a preview.
Sources: Gemini API models · Gemini API changelog · Google: Gemini 4 Argon · Gemini 4 Argon broad availability (VentureBeat, Oct 13)
xAI (SpaceXAI)
Grok 4.7 is the current API model, with a 500K-token context window. Cursor co-develops Grok models with SpaceXAI from Grok 4.5 onward.
Sources: xAI models · Grok on Wikipedia
Meta
Meta moved from open Llama models to the Muse line. Muse Wonder (October 12) is on the Meta Model API. Muse Wonder Full is only in Muse Code. Muse Spark Open (October 15) and Muse Glimmer 30B have open weights.
Sources: Meta Model API models · Muse Wonder launch (VentureBeat, Oct 12) · Meta model versions list · LLM Releases tracker
DeepSeek
Two API models, both with a 1M-token context window. Prices drop by half outside peak hours (peak: 01:00 to 04:00 and 06:00 to 10:00 UTC, weekdays).
Sources: DeepSeek API pricing · DeepSeek V4.1 Flash 1017 release (Oct 17)
Alibaba (Qwen)
Qwen3.8-Max has 2.4 trillion parameters, with about 95 billion active for each token. Alibaba previewed Qwen 4 in September. It is still in training.
Sources: Alibaba: Qwen3.8-Max · LLM Releases tracker
Moonshot AI (Kimi)
Kimi K3.1 Pro (October 23) is the new flagship, with a 1M-token context window. Moonshot calls it an open-weight model, but the weights are not posted yet. Kimi K3 (2.8 trillion parameters) has had open weights since July 26, 2026.
Sources: Kimi K3.1 Pro launch (Moonshot, Oct 23) · Kimi K3.1 Pro on OpenRouter · Pure AI: Kimi K3 launch · LLM Gateway: Kimi K3 weights
Z.ai (GLM)
GLM-5.3 uses the GLM-5.2 base model with new post-training. GLM-5.3-Flash is the low-cost open model under the MIT license.
Sources: DataNorth: GLM-5.3 · LLM Releases tracker
Mistral AI
Mistral Large 4 is a preview with about 1 trillion parameters, 49 billion active. Mistral says the open weights come at the end of October 2026.
Sources: Artificial Analysis: Mistral Large 4 · Simon Willison on Large 4
Xiaomi (MiMo)
MiMo-V2.6-Pro is one of the strongest open-weight models on the Artificial Analysis index, at about $0.13 for each index task. Xiaomi retired MiMo-V2.5 on October 21.
Sources: LLM Releases tracker · Artificial Analysis leaderboard
03 · Benchmarks for content creation
Which models write, draw, research and decide best
Eight public boards in three groups. Writing: four tests of text quality. Media: human votes on images and video. Agents: web research reports, and small models that make fast decisions.
EQ-Bench Creative Writing v3
Short creative pieces from a fixed set of prompts. An AI judge scores each piece, then models compete in pairs for an Elo rating.
| Board rank | Model | Bar | Elo |
|---|---|---|---|
| #1 | GPT-6 AstraOpenAI | 2,174.0 | |
| #2 | Claude Fable 5.1Anthropic | 2,162.7 | |
| #3 | Claude Opus 5Anthropic | 2,133.5 | |
| #4 | GPT-6 SolOpenAI | 2,124.6 | |
| #5 | Kimi K3Moonshot | 2,083.5 | |
| #6 | GLM-5.3Z.ai | 2,076.7 | |
| #7 | Claude Opus 5.5Anthropic | 2,050.8 | |
| #8 | Grok 4.7xAI | 2,012.2 | |
| #9 | Muse WonderMeta | 2,007.0 | |
| #11 | GPT-5.6 SolOpenAI | 1,973.5 | |
| #12 | Claude Fable 5Anthropic | 1,943.6 | |
| #14 | Muse Spark 1.1Meta | 1,928.2 |
Read this: Muse Wonder (Meta) is new, in 9th place (2007.0). Gemini 4 Argon is now on the board, in 31st place (1758.0), the top Google model. Two places (10 and 13) belong to models with no named company, so they are not shown.
Sources: EQ-Bench Creative Writing v3 · Board rank is the place on the full board.
EQ-Bench Longform Creative Writing
The model plans a story, revises the plan, then writes eight chapters of about 1,000 words. The test also checks if quality drops in later chapters.
| Board rank | Model | Bar | /100 |
|---|---|---|---|
| #1 | Claude Opus 5Anthropic | 86.3 | |
| #2 | Claude Fable 5.1Anthropic | 85.3 | |
| #3 | Claude Opus 5.5Anthropic | 84.6 | |
| #4 | Claude Fable 5Anthropic | 83.0 | |
| #5 | Grok 4.7xAI | 83.0 | |
| #6 | GPT-6 AstraOpenAI | 82.8 | |
| #7 | Muse Spark 1.3Meta | 82.8 | |
| #8 | GPT-6 SolOpenAI | 82.8 | |
| #10 | Claude Opus 4.7Anthropic | 81.8 | |
| #11 | GLM-5.3Z.ai | 81.8 | |
| #12 | GPT-5.6 SolOpenAI | 81.7 | |
| #13 | Muse Spark 1.2Meta | 81.5 |
Read this: This is the test closest to a long blog post or report. Claude models take the top four places. Kimi K3 is in 18th place (79.6). Gemini 4 Argon is new, in 14th place (80.9).
Sources: EQ-Bench Longform Creative Writing · Board rank is the place on the full board.
Arena Text: Creative Writing
Real users write a prompt, read two anonymous answers, and pick the better one. The board had 1,393,561 votes on October 8, 2026.
| Board rank | Model | Bar | score |
|---|---|---|---|
| #1 | Claude Opus 5.5 (high)Anthropic | 1,520±14 | |
| #2 | Gemini 4 Argon (high)Google | 1,512±14 | |
| #3 | Claude Fable 5 (high)Anthropic | 1,501±7 | |
| #4 | Claude Opus 4.6 (high)Anthropic | 1,499±6 | |
| #5 | Muse Wonder (max)Meta | 1,492±17 | |
| #6 | Gemini 3.7 Flash (high)Google | 1,492±8, early | |
| #7 | Claude Opus 4.7 (high)Anthropic | 1,489±7 | |
| #8 | Claude Fable 5.1 (max)Anthropic | 1,487±11 | |
| #9 | Gemini 3.8 Flash (high)Google | 1,485±7, early | |
| #10 | Gemini 3 ProGoogle | 1,485±8 | |
| #11 | Claude Opus 4.7Anthropic | 1,482±7 | |
| #12 | Gemini 3.1 Pro (preview)Google | 1,481±5 |
Read this: Claude Opus 5.5 is now first and Gemini 4 Argon second. The two scores are inside each other's margin of error (±14 each). Muse Wonder (Meta) is new, in 5th place. The top OpenAI entry is GPT-5.6 Sol (xhigh), in 17th place (1470). Board of October 22, 2026, with 1,441,226 votes.
Sources: Arena Text: Creative Writing · Board rank is the place on the full board.
EQ-Bench Slop Score
How often a model uses words and phrases that AI models overuse. EQ-Bench counts matches against a list built from many models. Lower is better.
| Writing rank | Model | Bar | slop |
|---|---|---|---|
| #3 | Claude Opus 5Anthropic | 0.9 | |
| #2 | Claude Fable 5.1Anthropic | 1.1 | |
| #1 | GPT-6 AstraOpenAI | 1.2 | |
| #4 | GPT-6 SolOpenAI | 1.2 | |
| #6 | GLM-5.3Z.ai | 1.2 | |
| #8 | Grok 4.7xAI | 1.2 | |
| #9 | Muse WonderMeta | 1.2 | |
| #5 | Kimi K3Moonshot | 1.3 | |
| #7 | Claude Opus 5.5Anthropic | 1.4 | |
| #11 | GPT-5.6 SolOpenAI | 1.6 | |
| #31 | Gemini 4 ArgonGoogle | 3.2 | |
| #61 | DeepSeek V4.1 FlashDeepSeek | 3.1 |
Read this: The top writers use the fewest AI phrases. Gemini 4 Argon scores 3.2, and DeepSeek V4.1 Flash and Gemini 3.8 Flash score 3.1, more than twice the rate of the top models. Edit their drafts with more care.
Sources: EQ-Bench Slop Score · Writing rank is the place on the Creative Writing v3 board.
Arena Text-to-Image
Users write a prompt, see two anonymous images, and pick the better one. The board had 6,604,125 votes for 82 models on October 7, 2026.
| Board rank | Model | Bar | score |
|---|---|---|---|
| #1 | GPT Image 2.5 SunburstOpenAI | 1,425±7, early | |
| #2 | GPT Image 2.5 FlareOpenAI | 1,398±7, early | |
| #3 | GPT Image 2 (medium)OpenAI | 1,383±4 | |
| #4 | Grok Imagine Image 2.0 (canvas)xAI | 1,334±12, early | |
| #5 | MAI-Image-2.6Microsoft AI | 1,332±5 | |
| #6 | Nano Banana 2.1Google | 1,328±9, early | |
| #7 | Reve 2.1Reve | 1,302±8 | |
| #8 | Grok Imagine Image 2.0xAI | 1,294±6 | |
| #9 | Muse ImageMeta | 1,274±5 | |
| #10 | Reve 2.0Reve | 1,269±6 | |
| #11 | Nano Banana 2 (web search)Google | 1,262±4 | |
| #12 | Seedream 5.0 ProByteDance | 1,255±4 |
Read this: OpenAI has the top three places. Microsoft AI, Google and xAI are next, close together. "Early" marks a score that can change with more votes.
Sources: Arena Text-to-Image · Board rank is the place on the full board.
Arena Text-to-Video
Users write a prompt, watch two anonymous videos, and pick the better one. The board had 718,577 votes for 48 models on September 21, 2026.
| Board rank | Model | Bar | score |
|---|---|---|---|
| #1 | Gemini Omni 1.1 FlashGoogle | 1,516±15 | |
| #2 | Gemini Omni FlashGoogle | 1,513±9 | |
| #3 | FLUX 3 VideoBlack Forest Labs | 1,493±17, early | |
| #4 | Grok Imagine Video 1.5 (agent)xAI | 1,492±18, early | |
| #5 | Seedance 2.0 (720p)ByteDance | 1,479±8 | |
| #6 | Wan 3.0Alibaba | 1,476±13 | |
| #7 | Seedance 2.5 (720p)ByteDance | 1,474±9 | |
| #8 | MiniMax H3MiniMax | 1,460±9 | |
| #9 | Muse VideoMeta | 1,456±15 | |
| #10 | HappyHorse 1.0Alibaba-ATH | 1,427±13 | |
| #11 | Sora 2 ProOpenAI | 1,368±7 | |
| #12 | Veo 3.1 (audio)Google | 1,364±14 |
Read this: Google's Gemini Omni models are first and second. Sora 2 Pro is the top OpenAI entry, in 11th place. New models have wide margins (±15 to ±18), so their places can change.
Sources: Arena Text-to-Video · Board rank is the place on the full board.
DeepResearch Bench II (Mercor)
130 open research briefs in 22 fields and two languages, graded against 9,287 criteria that experts wrote. Each brief comes from a real review article that the model cannot read. The model must find the same facts on the web.
| Board rank | Model | Bar | % |
|---|---|---|---|
| #1 | Claude Opus 5 (high)Anthropic | 56.1 | |
| #2 | Claude Fable 5.1 (max)Anthropic | 55.0 | |
| #3 | Claude Sonnet 4.6 (high)Anthropic | 53.3 | |
| #4 | Claude Opus 5.5 (max)Anthropic | 53.3 | |
| #5 | DeepSeek V4.1 Flash (max)DeepSeek | 53.2 | |
| #6 | GLM-5.3Z.ai | 52.0 | |
| #7 | Claude Sonnet 5.5Anthropic | 51.7 | |
| #8 | GPT-5.6 SolOpenAI | 51.5 | |
| #9 | Qwen3.8-MaxAlibaba | 51.0 | |
| #10 | GPT-6 AstraOpenAI | 50.1 | |
| #11 | GPT-5.6 LunaOpenAI | 49.7 | |
| #12 | Grok 4.7xAI | 49.0 |
Read this: This test is closest to a research report or a long article. The top three are within 2.8 points. On BrowseComp, a web search test with scores that the providers report, the leaders are Atria Dawn Preview (92.5%), GPT-5.6 Sol (92.2%), GPT-6 Astra (91.5%) and Kimi K3 (91.2%).
Sources: DeepResearch Bench II (Mercor) · BenchLM mirror (places 7 to 12) · BrowseComp scores · Board rank is the place on the full board.
Decision models
A decision model does not write text. It reads an input and returns one typed answer with a probability: yes or no, one choice from a list, or a score. Agents use them to route requests, flag problems and check work, fast and at low cost.
| Model | Company | Released | Input price per 1M (output free) | Weights | Published score |
|---|---|---|---|---|---|
| Jev 1.13TypeSafe AI | TypeSafe AI | $0.042 | Closed | JevBench 72.13 (3rd of 112) | |
| d1-ultraLiquid AI | Liquid AI | $0.042 | Not stated | Beats Jev 1.14 on 112 evals (self-reported) | |
| mammoth-1 (27B)Base Labs | Base Labs | Free until Oct 27 | Open (MIT) | Jev Decision Index 71.24 (self-reported) | |
| Microsoft-Decision-1Microsoft | Microsoft | $0.042 | Closed (built on Qwen3.5-9B) | No numbers published | |
| Decisions API (GPT-6 Luna)OpenAI | OpenAI | $0.10 | Closed | No numbers published (generally available since Oct 22) | |
| Clef (27B)Cloudflare | Cloudflare | $0.24 | Open (Apache 2.0) | Decision Index 61.2 (self-reported) | |
| Clef-flash (9B)Cloudflare | Cloudflare | $0.038 | Open (Apache 2.0) | Decision Index 57.1 (self-reported) | |
| Perplexity Decider v1.2 27BPerplexity | Perplexity | $0.02 | Open (Apache 2.0) | 85.71% on its own 11-task panel | |
| Strands Decider 2BAmazon (Strands Labs) | Amazon (Strands Labs) | Self-host | Open (Apache 2.0) | JevBench 0.762 (own run) | |
| d1-3BLiquid AI | Liquid AI | Self-host | Open (LFM Open License) | Decision Index 48.57 (self-reported) |
Read this: The scores come from different boards and versions. Do not compare them directly. Microsoft and OpenAI have not published scores that you can compare with other models. New since October 12: d1-ultra, mammoth-1 and elephant-1, and Perplexity Decider v1.2. Jev 1.14 is in preview.
Sources: Decision models · TypeSafe AI Jev launch · Microsoft-Decision-1 · OpenAI Decisions API · Cloudflare Clef · Decision model overview · OpenAI Decisions API GA (Oct 22) · Liquid AI d1-ultra (Oct 15) · Base Labs mammoth-1 and elephant-1 (Oct 22) · Perplexity Decider v1.2
BEAM: Beyond a Million Tokens
BEAM tests how well an AI system remembers a long conversation. It has 100 conversations and 2,000 questions, at 100K, 500K, 1M and 10M tokens. The questions cover ten memory abilities, such as updated facts, contradictions, event order, preferences and summaries. A memory system stores the conversation and finds the right parts for each question.
| System | 100K | 500K | 1M | 10M | Run by |
|---|---|---|---|---|---|
| past.devMemory API | 92.1 | 89.6 | 90.7 | 85.0 | Self-reported, Sep 29, 2026. GPT-5.6 Luna answers and judges. |
| Exabase M-1Exabase | 76.9 | — | 75.0 | 68.0 | Self-reported, Jul 28, 2026. Gemini 3 Flash. |
| HindsightVectorize | 75.0 | 71.1 | 73.9 | 64.1 | Self-reported on the Agent Memory Benchmark site. |
| Mem0 PlatformMem0 | — | — | 64.1 | 48.6 | Self-reported, 2026. About 6,900 tokens per question. |
| HonchoPlastic Labs | 63.0 | 64.9 | 63.1 | 40.6 | Plastic Labs figures, as listed by Hindsight and past.dev. |
| LIGHT (BEAM paper)Llama 4 Maverick | 35.8 | 35.9 | 33.6 | 26.6 | BEAM paper, ICLR 2026. The paper's own memory method. |
| Long context only (BEAM paper)Llama 4 Maverick | 24.0 | 28.3 | 25.9 | 10.4 | BEAM paper. The model reads the whole history, with no memory system. |
Read this: Each company ran its own test with its own answer model and judge, so the rows are not a fair head-to-head. No one has reproduced the past.dev scores yet. past.dev also lists higher Hindsight scores (86.2, 80.1 and 79.1) that we could not confirm on the Hindsight site. The two paper rows use the original BEAM setup and older models, so they are much lower. A dash means that no score is published at that size.
Sources: BEAM: Beyond a Million Tokens · past.dev BEAM leaderboard · past.dev results and harness · Hindsight Agent Memory Benchmark · Hindsight BEAM post (Honcho figures) · Exabase M-1 announcement · Mem0 BEAM results · Checked October 26, 2026.
What we see
The judges disagree
- The AI judge on EQ-Bench puts OpenAI and Anthropic first. People who vote on the Arena put Google and Anthropic first.
- Claude models are in the top four on the longform test. For long blog posts and reports, start there.
- The best writers also use the fewest overused AI phrases. A low slop score means less editing for you.
- Open-weight Kimi K3 and GLM-5.3 are in the creative writing top six, near the closed flagships.
- OpenAI leads images, with the top three places. Google leads video. One company rarely leads text, images and video together.
- Decision models cost a fraction of a chat model. Jev and Microsoft-Decision-1 charge $0.042 for each 1M input tokens, and output is free.
04 · Research · 16 studies
Research on AI search and AI writing
Studies and audits that show how AI tools pick sources, and how AI models write. Each card says what the finding means for your content.
9.3%
AI Mode Cut 70% of Its Sources. Here's What It Kept.
Maeva Cifuentes, The Search Spot (Scrunch data)
Google AI Mode went from about 16 sources per answer to about 4 at the end of August. Only 9.3% of the pages that AI Mode cited in the summer kept the same citation. The data covers 8 B2B software brands and 407 prompts.
For your content: Citations can disappear overnight. Check your AI visibility each month, not once.
88%
Google AI Overviews nearly double on commercial keywords
Semrush, reported by Search Engine Land
Semrush tracked 400 US desktop keywords in banking, ecommerce and software. AI Overviews showed for 88% of these commercial keywords in August 2026, up from 45.75% in September 2025.
For your content: People who are ready to buy now see an AI answer first. Your business must be in that answer.
0.62%
ChatGPT sends less than 1% of Google's traffic: Data
Marie Haynes, reported by Search Engine Land
In GA4 data from 100 websites over 12 months, ChatGPT sent 0.62% as many sessions as Google organic search. On the 55 sites that track conversions, 13.8% of ChatGPT visitors converted, against 3.02% from Google.
For your content: Google's AI features still send most visits. ChatGPT sends few visitors, but they convert well.
31,200 answers
Location, Location, Generation: Where You Ask Matters for Generative Engine Optimization
Mahe Chen and Nick Koudas (arXiv)
The authors asked ChatGPT and Gemini 300 questions from different locations. ChatGPT changed its sources and recommended businesses based on the declared city. Gemini followed the country of the network connection.
For your content: AI answers change by location. Test your visibility from your customers' city.
71.2%
The Effects of Incremental Instruction Delivery on Language-Model Creative Writing
Singh, Eyasir, Yaqoob, Manavalan
When writers gave their requirements over 5 to 9 turns, models kept only 71.2% of the quality that they got from one full brief.
For your content: Write one complete brief at the start. Do not add rules one at a time.
3 years
Are LLMs becoming similarly creative? Evidence from three years of models
Patel, Crossman, Aggarwal, Wenger
The authors compared answers from three years of models. Output diversity went down over time, so models now give more similar answers.
For your content: Many businesses use the same models. Your own facts, stories and voice make your content different.
45 studies
Optimizing Visibility in Generative Engines: A Critical Survey of Generative Engine Optimization (2023-2026)
Olivier Martinez
A review of 45 GEO studies. Gains are clear when AI tools already found the page. The evidence for gains in getting found at all is weaker.
For your content: First make AI tools able to find and read your pages. Then improve the content.
73% / 44% / 11%
Generative Engine Optimization at Scale: Measuring Brand Visibility Across AI Search Engines
Pratyush Kumar
In more than 100,000 AI answers, global brands appeared 73% of the time, regional brands 44%, and niche brands 11%. About 78% of citations went to company websites.
For your content: Your own website is the main source. Small brands start far behind, so clear facts matter more.
16%
Synthetic Sources?: Auditing Generative Search Engine Citations for Evidence of AI-Generated Sources
Allaham, Diakopoulos
The authors sent 712 real queries to ChatGPT, Copilot, Gemini and Perplexity. About 16% of the cited sources were AI-generated.
For your content: Citations go to a small set of domains. Human-checked content is a way to stand out.
Book-length
Towards Human-Level Book-Writing Capability
Zierstek, Batelic, Medjad, Schönenberger
A model trained to expand a plan into a full book scored above GPT-5.5 and Claude Opus 4.8 in the authors' writing test.
For your content: Plan first, then write. A clear outline gives better long content.
21,143 citations
From Citation Selection to Citation Absorption: A Measurement Framework for Generative Engine Optimization Across AI Search Platforms
Zhang, He, Yao
The study looked at 21,143 citations from ChatGPT, Google and Perplexity. Pages that shaped the answer most were longer, more structured and had more facts to quote.
For your content: A citation is not enough. Give AI tools facts that they can use in the answer.
+17.3%
Structural Feature Engineering for Generative Engine Optimization: How Content Structure Shapes Citation Behavior
Yu, Yang, Ding, Sato
The authors changed only page structure: layout, chunks and emphasis. Across six AI engines, the citation rate went up by 17.3%.
For your content: Structure counts, not only words. Headings, short sections and clear lists help.
82.7% → 62%
Can Good Writing Be Generative? Expert-Level AI Writing Emerges through Fine-Tuning on High-Quality Books
Tuhin Chakrabarty, Paramveer S. Dhillon (CHI 2026)
Expert judges picked human writing 82.7% of the time over prompted AI. After fine-tuning on each author's books, experts picked the AI writing 62% of the time.
For your content: Models trained on good examples write better. Give your AI tool samples of your best content.
8% vs 15%
Google users are less likely to click on links when an AI summary appears in the results
Pew Research Center
Pew tracked the browsing of 900 US adults. People clicked a result link on 8% of visits with an AI summary, and on 15% of visits without one.
For your content: Fewer people click. The AI answer must name your business, or people do not see you.
60%+
AI Search Has a Citation Problem
Columbia Journalism Review, Tow Center
The Tow Center sent 1,600 queries about news articles to eight AI search tools. Together, the tools gave wrong answers to more than 60% of the queries.
For your content: AI tools often get sources wrong. Clear facts on your own site lower the risk that they get you wrong.
Up to 40%
GEO: Generative Engine Optimization
Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan, Deshpande
The paper that named GEO. It reports that content changes can increase visibility in AI answers by up to 40%.
For your content: The start of the field. Later studies test where these gains hold.
Want our own analysis? Read A/GEO insights or see the 50-check audit.
05 · Method
How we collect the data
We do not run the benchmarks. We collect published results and check them against the source.
- 01
Primary sources first. Model names, prices and context windows come from each company's API documents. When a company does not publish a value, we use a release tracker and name it.
- 02
Scores as published. Benchmark scores are copied from each board on the date at the top of the page. We do not adjust them.
- 03
Named models only. The boards also list test models with code names. On the writing boards, we show only models from the 11 companies on this page. On the media and agent boards, we show every named company. Each model keeps its place on the full board.
- 04
Access is marked. "Limited access" means that the public cannot use the model now. "Preview" means that the model can change.
- 05
Unknown stays unknown. "Not listed" means that we did not find a reliable value. We do not estimate.
Questions
LLM Watch FAQ
What is LLM Watch?
LLM Watch is a free research page from A/GEO Agency. It tracks the major companies that make large language models (LLMs), their current models, their prices, and their scores on writing benchmarks. Each figure links to its source, and the page shows the date that we checked the data.
Which AI model is best for writing content in October 2026?
No single model wins every test. On EQ-Bench Creative Writing v3, GPT-6 Astra has the top score (2174.0 Elo) and Claude Fable 5.1 is second (2162.7). On the EQ-Bench Longform Writing test, Claude Opus 5 is first (86.3 of 100). In human votes on the Arena creative writing board, Claude Opus 5.5 (1520) and Gemini 4 Argon (1512) are first and second, inside each other's margin of error.
Who are the major LLM companies?
In the United States: OpenAI, Anthropic, Google, xAI and Meta. In China: DeepSeek, Alibaba (Qwen), Moonshot AI (Kimi), Z.ai (GLM) and Xiaomi (MiMo). In Europe: Mistral AI, in France.
Why do writing benchmarks give different results?
Each test uses a different judge. EQ-Bench uses an AI model as the judge, with a scoring rubric. The Arena uses votes from people who compare two answers. AI judges and people do not always prefer the same writing, so use more than one board.
Which models can I download and run myself?
These models have open weights: DeepSeek V4.1 Flash 1017 (MIT), Kimi K3 (Kimi K3 license), GLM-5.3-Flash (MIT), MiMo-V2.6-Pro (MIT), Meta Muse Spark Open and Muse Glimmer 30B, and Mistral Large 3 (Apache 2.0). Moonshot calls Kimi K3.1 Pro an open-weight model, but the weights are not posted yet. Mistral says that the Large 4 weights come at the end of October 2026. The OpenAI, Anthropic and Google models on this page are closed.
What does this mean for my business website?
These models write the answers in ChatGPT, Claude, Gemini, Meta AI and Grok. In a Pew Research Center study, people clicked a result link on 8% of Google visits with an AI summary, and on 15% of visits without one. If AI tools cannot find clear facts about your business, they cannot recommend you. A/GEO Agency gives AI tools those facts.
Which AI models make the best images and videos?
In human votes on the Arena boards, OpenAI has the top three image models: GPT Image 2.5 Sunburst (1425), GPT Image 2.5 Flare (1398) and GPT Image 2 (1383). For video, Google's Gemini Omni 1.1 Flash (1516) and Gemini Omni Flash (1513) are first and second. Sora 2 Pro is the top OpenAI video model, in 11th place.
What is a decision model?
A decision model does not write text. It reads an input and returns one typed answer with a probability: yes or no, one choice from a list, or a score. Agents use decision models to route requests, flag problems and check work. Examples are Jev from TypeSafe AI, Microsoft-Decision-1, the OpenAI Decisions API, Cloudflare Clef and Liquid AI d1-ultra. Jev and Microsoft-Decision-1 cost $0.042 for each 1M input tokens, and output is free.
Which AI model is best for research reports?
On DeepResearch Bench II from Mercor, which grades long web research against expert rubrics, Claude Opus 5 is first (56.1%) and Claude Fable 5.1 is second (55.0%). The top three are within 2.8 points, so several models give similar results.
Which AI memory system scores highest on BEAM?
On BEAM, a test of memory over long conversations, past.dev reports the highest scores: 92.08% at 100K tokens and 85.03% at 10M tokens. Exabase M-1 (68.0%) and Hindsight (64.1%) follow at 10M tokens. Each company ran its own test with its own models and judge, and no one has reproduced the past.dev scores yet.
How current is the data?
The date at the top of the page shows when we last checked the data. AI companies release new models often, so each figure links to its source. Use the source to check the current value.
Find out what AI says about you.
We ask leading AI answer engines the questions your customers ask. Then we show you who they name, and what to fix first. The check is free.