Research brief · AI models · October 2026 · Artificial Analysis, lab pricing pages, filings

Best AI models
of 2026

Fourteen labs now sell serious AI: ChatGPT, Claude, Gemini, Grok, Meta AI, Mistral, DeepSeek, Qwen, Copilot, Perplexity, Cohere, Kimi, Amazon Nova and GLM. We compared every one on the independent Artificial Analysis index, on price, on what it is actually good at, and on whether the company behind it is built to last. Every product links to its official sign-up page. Research date: 8 October 2026.

See the comparison
1.2BPeople using ChatGPT each week (OpenAI, Oct 2026)
1B+Monthly Gemini app users (Google, Aug 2026)
58Top Artificial Analysis Intelligence Index score, Claude Opus 5.5
$0.10Per 1M input tokens for GPT-6 Luna, vs $10 for GPT-6 Astra
Server racks and network cabling in a data center
The race · February → October 2026 · frontier AI

Money, open weights, a new frontier

2026 was the year the war chests got absurd, Meta closed its flagship while Chinese labs opened theirs, and the newest frontier models shipped first to vetted partners. By October the top of the market is a three-way race between Anthropic, OpenAI and Google, with a fast, cheap open-weights pack close behind.

Feb → Jun 2026 · Capital

The war chests

$965B

SpaceX takes over xAI (Feb). OpenAI raises $122B at $852B (Mar). Anthropic raises $65B at $965B (May). SpaceX lists on Nasdaq (Jun).

Apr → Aug 2026 · Weights

Meta closes, China opens

2.8T

Meta's Muse Spark ships proprietary (Apr). DeepSeek V4, Kimi K3 (2.8T parameters), Qwen3.8 and GLM-5.3 publish their weights.

Sep 2026 · Frontier reset

The gated frontier

58

Claude Fable 5.1, GPT-6 Astra, Grok 4.7 and Claude Opus 5.5, which tops the index at 58. Gemini 4 Argon goes to cyber defenders only.

Oct 2026 · Mass market

Frontier for everyone

1.2B

GPT-6 reaches every ChatGPT tier. Mistral Large 4 opens in preview with weights promised for the end of October.

Capability against price, flagship models

Artificial Analysis Intelligence Index

Cross-check · Artificial Analysis, lab announcements, pricing pages

Where the evidence agrees

  • All
    Anthropic leads the independent index

    Claude Opus 5.5 scores 58 and Sonnet 5.5 scores 56 on the Artificial Analysis Intelligence Index. GPT-6 Astra, Gemini 4 Argon and Claude Fable 5.1 sit at 53.

  • All
    Cheap AI has never been cheaper

    GPT-6 Luna costs $0.10 in and $0.50 out per million tokens. DeepSeek V4.1 Flash costs $0.15 and $0.60 off-peak. The frontier tier still charges $20 to $50 per million output tokens.

  • All
    A million tokens of context is standard

    GPT-6.1 Sol (1.05M), Claude Opus 5.5 (1M), Mistral Large 4 (1M), DeepSeek V4.1 Flash (1M) and Kimi K3 (1M) all read a book-length input in one go.

  • All
    Distribution wins consumers

    ChatGPT has 1.2 billion weekly users, the Gemini app passed 1 billion monthly users, and Qwen's app 300 million. Meta AI sits inside apps used by 3.6 billion people a day.

  • Most
    The newest frontier ships gated

    Claude Mythos 5.1 goes only to vetted cyber and life-science organisations, and Gemini 4 Argon only to trusted cyber defenders. Capability is outrunning general release.

  • Most
    Open weights moved to China and Europe

    Meta's flagship Muse Spark is proprietary. The leading open models come from Xiaomi, Z.ai, Moonshot, Alibaba and DeepSeek, with Mistral and Cohere carrying Europe's and Canada's open releases.

Where they disagree

Will open weights catch the closed frontier?

Today the best open model scores 46 against 58 for the best closed one. Placed by where each camp expects the gap to settle.

Closed lead growsParity

How many frontier labs survive the capex race?

Hyperscale compute deals run into gigawatts and hundreds of billions. Does that leave room for more than the giants?

ConsolidationFragmentation

Where does the value end up?

In the model itself, in the app that owns the user, or in the agent that finishes the work?

Model makersApplications

Artificial Analysis leaderboard, 8 Oct 2026 (Opus 5.5 58, Sonnet 5.5 56, Astra / Argon / Fable 53, MiMo-V2.6-Pro 46 top open-weights) · OpenAI pricing (Luna $0.10 / $0.50) · DeepSeek pricing (V4.1 Flash off-peak $0.15 / $0.60) · OpenAI (1.2B weekly) · Google (1B+ monthly) · Alibaba (Qwen app 300M+ MAU) · Meta Q2 2026 (3.60B daily active people) · Anthropic and Google model pages (Mythos 5.1, Argon access).

The field · 14 labs and assistants · prices and scores as of 8 October 2026

Every major AI, side by side

The flagship model, what it is best at, open or closed weights, API price per million tokens (input / output), consumer plans, its Artificial Analysis Intelligence Index score, our survival rating, and the official place to sign up. Scroll the table sideways on mobile.

Lab / assistantSign upFlagshipBest atWeightsAPI price (in / out)Consumer plansAA indexSurvival
OpenAIChatGPTchatgpt.com
API
GPT-6 Astra (API); GPT-6 Sol and Luna in ChatGPTAll-round assistant, coding with Codex, the widest product rangeClosed (gpt-oss is open)Astra $10 / $50 · GPT-6.1 Sol $2 / $10 · Luna $0.10 / $0.50Free, Go, Plus $20, Pro $100 / $200 / $50053Very strong
AnthropicClaudeclaude.ai
API
Claude Opus 5.5; Claude Fable 5.1 at the topAgentic coding, long multi-step tasks, knowledge workClosedOpus 5.5 $4 / $20 · Fable 5.1 $10 / $50 · Sonnet 5.5 $2 / $10Free, Pro $20 ($17/mo annual), Max from $100, Team, Enterprise58 (#1)Very strong
GoogleGeminigemini.google.com
API
Gemini 4 Argon (limited access); Gemini 3.8 FlashMultimodal work, Google apps, price-performanceClosed (Gemma is open)Argon $2 / $10 intro, then $4 / $20 · 3.8 Flash $0.75 / $3.75 introFree, AI Plus, AI Pro, AI Ultra $99.99 / $199.9953Very strong
SpaceXAI (xAI)Grokgrok.com
API
Grok 4.7Strong reasoning per dollar, live X dataClosed$2 / $6SuperGrok Lite $10, SuperGrok $30, Heavy $30046Strong
MetaMeta AImeta.ai
API
Muse Spark 1.3; Muse Glimmer (30B, open)An assistant inside WhatsApp, Instagram and Facebook; local agents with GlimmerClosed flagship; Glimmer open (Apache 2.0)Muse Spark contributor tier $0.10 / $0.20 (Meta uses the data)Meta AI app and inside Meta's apps48Strong
Mistral AIVibechat.mistral.ai
API
Mistral Large 4 (public preview)European data sovereignty, self-hosting, multilingual workOpen weights due end of Oct 2026$1.36 / $4.18 (launch price $0.68 / $2.09)Free, Pro $14.99, Team $24.99/user, Enterprise38Solid
DeepSeekDeepSeekchat.deepseek.com
API
DeepSeek V4.1 Flash; V4 ProLowest-cost capable API, 1M contextOpen (V4 weights on Hugging Face)V4.1 Flash $0.15 / $0.60 off-peak, double at peakFree web and app chat39Solid
AlibabaQwenchat.qwen.ai
API
Qwen3.8 Max; Qwen3.8 2.4T (open)Open base models to fine-tune, multilingual workOpen (Qwen3.8-27B under Apache 2.0)Via Alibaba Cloud Model StudioQwen app (300M+ monthly users)45Strong
MicrosoftCopilotcopilot.microsoft.com
Business
Microsoft 365 CopilotWork inside Word, Excel, Outlook and TeamsClosedNot compared (sold per seat)Copilot Chat included with eligible Microsoft 365; Microsoft 365 Copilot $30/user/mo, paid yearlyn/aVery strong
PerplexityPerplexity · Cometperplexity.ai
Pro
Multi-model answer engine; Computer agentCited web research, AI browsingUses partner modelsNot comparedFree, Pro $20, Max $200, Enterprise Pro $40/seatn/aUncertain
CohereNorth · Commanddashboard.cohere.comCommand A+ (218B MoE)Private enterprise deployment, retrieval, 48 languagesOpen (Apache 2.0)Free trial key; production pay-as-you-goEnterprise only13Uncertain
Moonshot AIKimikimi.com
API
Kimi K3 (2.8T parameters)Long-context agentic work, the largest open modelOpen (custom Kimi K3 licence)$3 / $15Moderato $19, Allegretto $39, Allegro $99, Vivace $19944Solid
AmazonNovanova.amazon.com
AWS
Nova 2 Lite; Nova 2.5 Sonic (voice)Low-cost AWS-native workloads, speech-to-speech agentsClosedVia Amazon Bedrocknova.amazon.com14 (Nova 2.0 Pro preview)Strong
Z.ai (Zhipu)GLMchat.z.ai
Plans
GLM-5.3Open-weights coding and agents at a low priceOpen$1.40 / $4.40GLM Coding Plan45Solid

API prices per 1M tokens in USD from each lab's pricing page, standard context, 8 Oct 2026. Introductory prices are labelled. Plan prices are monthly in USD and vary by country and tax. AA index is the Artificial Analysis Intelligence Index for the lab's highest-scoring listed model (Anthropic: Opus 5.5; OpenAI: GPT-6 Astra; Alibaba: Qwen3.8 Max). Survival is a House of 47 rating, explained below. Grok plan prices are from SpaceX's S-1 via Digital Applied. The open-weights leader on the index, Xiaomi MiMo-V2.6-Pro (46, aistudio.xiaomimimo.com), is noted but not profiled.

Model by model · strengths, weaknesses, best use, outlook

What each one is good at

The short version of every lab: what it does best, where it falls short, what it costs, and who should use it.

OpenAI · ChatGPT

Flagship. GPT-6 Astra in the API ($10 / $50 per 1M tokens, 1.05M context). GPT-6.1 Sol costs a fifth as much ($2 / $10) and scores 52 to Astra's 53. ChatGPT runs GPT-6 Sol on paid plans and GPT-6 Luna on Free and Go.

Strengths. The biggest audience in AI (1.2 billion weekly users), the broadest product (chat, Codex for coding, agents, voice) and $2 billion of revenue a month as of March 2026. Weaknesses. No longer number one on the independent index, and its top model is priced at the high end.

Best for. One assistant that does everything, and coding with Codex. Plans. Free, Go (price varies by country), Plus $20, Pro $100, $200 or $500. Sign up: chatgpt.com · API platform.openai.com

Anthropic · Claude

Flagship. Claude Opus 5.5 (22 Sep 2026, $4 / $20, 1M context) is #1 on the Artificial Analysis index at 58. Claude Fable 5.1 ($10 / $50) is Anthropic's most capable model. Its twin, Mythos 5.1, goes only to vetted cyber and life-science organisations.

Strengths. Best measured intelligence, long agentic coding runs and careful writing. Opus 5.5 costs less than GPT-6 Astra at a higher score. Weaknesses. Heavy users hit plan usage limits (Max sells 5x or 20x the usage of Pro), and Fable 5.1 shares the top price tier with GPT-6 Astra.

Best for. Software engineering, research and long documents. Plans. Free, Pro $20 ($17/mo billed yearly), Max from $100, Team, Enterprise. Sign up: claude.ai · API platform.claude.com

Google · Gemini

Flagship. Gemini 4 Argon (30 Sep 2026) scores 53 but is limited to trusted cyber defenders for now, at $2 / $10 introductory and $4 / $20 later. Gemini 3.8 Flash ($0.75 / $3.75 until 31 Dec 2026, then $1.50 / $7.50) is the workhorse.

Strengths. Over 1 billion monthly app users, deep integration with Gmail, Docs, Android and Chrome, and strong multimodal input. Weaknesses. The best model is not generally available, and the free tier mixes models (3.6 Flash with varying 3.1 Pro access).

Best for. People who already live in Google's apps, and cheap high-volume multimodal work. Plans. Free, AI Plus, AI Pro, AI Ultra $99.99 or $199.99. Sign up: gemini.google.com · API aistudio.google.com

SpaceXAI · Grok

Flagship. Grok 4.7 (21 Sep 2026, $2 / $6, 500K context), scoring 46. Grok 5 is in training with no date. xAI became part of SpaceX in February 2026 and now trades as SpaceXAI.

Strengths. A low output price for its score, live access to X posts, and SpaceX's balance sheet behind it. Weaknesses. xAI lost $6.4 billion on $3.2 billion of revenue in 2025 according to the S-1, and it trails the top three by 7 to 12 points.

Best for. Real-time news and social context, and budget reasoning. Plans. SuperGrok Lite $10, SuperGrok $30 ($300/yr), Heavy $300. Sign up: grok.com · API console.x.ai

Meta · Meta AI

Flagship. Muse Spark 1.3, scoring 48, from Meta Superintelligence Labs. The first Muse Spark (Apr 2026) broke with the open Llama line and is proprietary. Muse Glimmer (Aug 2026) is a 30B open model under Apache 2.0 that runs on one consumer GPU.

Strengths. Distribution through apps used by 3.6 billion people a day, a cheap API tier, and 2026 capex guidance of $130 to $145 billion. Weaknesses. A strategy that has swung from open to closed, and the cheapest API tier lets Meta use your data.

Best for. Quick help inside WhatsApp, Instagram and Facebook; Glimmer for on-device agents. Sign up: meta.ai · API dev.meta.ai

Mistral · Vibe

Flagship. Mistral Large 4, in public preview since 6 Oct 2026: about 1 trillion parameters with 52 billion active, a 1M context, and open weights due at the end of October. The list price is $1.36 / $4.18, half that at launch.

Strengths. Europe's frontier lab, built for governments and companies that want control over where data goes. It raised €3 billion at over €21 billion in September. Weaknesses. The preview scores 38 against 58 at the top.

Best for. EU-hosted or self-hosted deployments. Plans. Free, Pro $14.99, Team $24.99 per user, Enterprise. Sign up: chat.mistral.ai · API console.mistral.ai

DeepSeek

Flagship. DeepSeek V4.1 Flash (10 Sep 2026): 1M context, up to 384K output, $0.15 / $0.60 off-peak. V4 (Apr 2026) published V4-Pro (1.6T parameters, 49B active) and V4-Flash (284B, 13B active) on Hugging Face.

Strengths. The cheapest capable API among the big labs, open weights, efficient long context. Weaknesses. Scores 39, and the US weighed blacklisting it before holding off in June 2026.

Best for. High-volume, cost-sensitive API work and self-hosting. Plans. Free web and app chat; pay-as-you-go API. Sign up: chat.deepseek.com · API platform.deepseek.com

Alibaba · Qwen

Flagship. Qwen3.8 Max scores 45. Alibaba also released the weights of its Qwen3.8 flagship (2.4T parameters, 95B active, 1M context) and Qwen3.8-27B under Apache 2.0.

Strengths. The biggest open model family, with more than 460 models and over 3 billion downloads, plus a consumer app with 300 million monthly users. Weaknesses. Exposure to US curbs on Chinese AI, and a weaker brand in the West.

Best for. Fine-tuning your own model, multilingual products, running locally. Sign up: chat.qwen.ai · API Alibaba Cloud Model Studio

Microsoft · Copilot

Product. Copilot is a product layer rather than a leaderboard model. Microsoft 365 Copilot costs $30 per user per month paid yearly, and Copilot Chat comes at no extra cost with eligible Microsoft 365 plans.

Strengths. Over 30 million paid Microsoft 365 Copilot seats as of July 2026, and it sits inside the Office apps where companies already work. Weaknesses. The value depends on how much of your work lives in Microsoft 365.

Best for. Companies on Microsoft 365. Plans. Consumer Copilot app; Business Standard with Copilot $23.50 and Business Premium with Copilot $32 per user. Sign up: copilot.microsoft.com · business plans

Perplexity

Product. An answer engine that cites its sources and routes queries across partner models, plus the Comet browser and the Computer agent (Feb 2026).

Strengths. The best-known cited-search product, with an estimated $750 million in annualized revenue (Sacra, Aug 2026) and 100 million+ monthly users. Weaknesses. No frontier model of its own, and Amazon is suing over Comet.

Best for. Research where every claim needs a source. Plans. Free, Pro $20, Max $200, Enterprise Pro $40 per seat. Sign up: perplexity.ai

Cohere · Command

Flagship. Command A+ (May 2026): a 218B mixture-of-experts model with 25B active, open under Apache 2.0, 48 languages, running on two H100 GPUs.

Strengths. Built for private, on-premises enterprise use, and it is merging with Germany's Aleph Alpha, a deal Handelsblatt valued at about $20 billion. Weaknesses. It scores 13 on the general intelligence index.

Best for. Regulated companies that need retrieval and agents on their own hardware. Sign up: dashboard.cohere.com

Moonshot · Kimi

Flagship. Kimi K3 (Jul 2026): 2.8 trillion parameters, 1M context, native vision, with full weights released under a custom Kimi K3 licence. It scores 44.

Strengths. The largest open-weights model, priced at $3 / $15, with a confidential Hong Kong IPO filing behind the company. Weaknesses. The licence is not MIT or Apache, and Kimi K3 renewed a US push to restrict Chinese models.

Best for. Long-context agent work where you want the option to self-host. Plans. Moderato $19, Allegretto $39, Allegro $99, Vivace $199. Sign up: kimi.com · API platform.kimi.ai

Amazon · Nova

Flagship. Nova 2 Lite (1M context, 64K output) and Nova 2.5 Sonic for speech-to-speech, sold through Amazon Bedrock.

Strengths. More than 50,000 customers, and it lives inside AWS billing, security and tooling. Weaknesses. Nova 2.0 Pro Preview scores 14, and Amazon also supplies Anthropic with compute under a 5GW deal.

Best for. AWS shops that want cheap, governed models and voice agents. Sign up: nova.amazon.com · aws.amazon.com/nova

Z.ai · GLM

Flagship. GLM-5.3 (Aug 2026), open weights two weeks after launch, $1.40 / $4.40 in the API, scoring 45.

Strengths. Near the top of the open pack at a low price, with a dedicated GLM Coding Plan. Zhipu has been listed in Hong Kong since January 2026. Weaknesses. The same geopolitical exposure as every Chinese lab.

Best for. Low-cost coding agents and open-weights deployments. Sign up: chat.z.ai · GLM Coding Plan

Model facts from each lab's announcement and documentation, Artificial Analysis scores as of 8 Oct 2026, and company figures from the filings and reporting listed in Sources.

Built to last? · money, traction, risk

Who survives

Frontier AI now costs hundreds of billions of dollars in compute. Survival depends on three things: who pays for it, how many people use it, and what could cut it off. Here is each lab on those three, with our rating.

LabMoney behind itTractionMain riskOutlook
OpenAI$122B round at $852B post-money (Mar 2026)1.2B weekly ChatGPT users; $2B revenue a month (Mar 2026)Compute bills; lost the #1 score to AnthropicVery strong
Anthropic$65B Series H at $965B (May 2026); gigawatt compute deals with Amazon, Google and BroadcomRun-rate above $47B (May); above $65B per a Reuters source (Aug); #1 on the indexDependence on partners' computeVery strong
GoogleAlphabet's own cash flow and TPUs1B+ monthly Gemini app usersIts best model is still gatedVery strong
MicrosoftAzure passed $100B in annual revenue30M+ paid Microsoft 365 Copilot seatsSeat growth has to justify $30 per userVery strong
Meta2026 capex guidance $130-145B; Q2 revenue $60.8B3.60B daily active people across its appsStrategy swings between open and closedStrong
SpaceXAI (xAI)Part of SpaceX, whose June 2026 IPO raised about $85.7B117M Grok monthly users; 1.9M SuperGrok subscribers$6.4B operating loss on $3.2B revenue (2025)Strong
Alibaba QwenAlibaba; cloud revenue up 36%3B+ downloads; app 300M+ monthly usersUS curbs on Chinese modelsStrong
Amazon NovaAWS50,000+ customersLags on capability; Amazon also backs AnthropicStrong as AWS product
Mistral€3B Series D at over €21B, led by Samsung (Sep 2026)Operates in 20 countries; demand for sovereign AIScale gap: 38 vs 58 on the indexSolid
DeepSeekFirst outside round, over $7B (Jun 2026); a filing implies about $52BThe cheapest capable APIUS blacklisting, considered and deferredSolid
Moonshot (Kimi)$50B valuation in an ongoing round; confidential HK IPO filingThe largest open-weights modelA US push to restrict Chinese modelsSolid
Z.ai (Zhipu)Listed in Hong Kong (Jan 2026), raising HK$4.35BGLM-5.3 near the top of the open packPrice war and geopoliticsSolid
Perplexity$20B valuation (Sep 2025 round)~$750M annualized revenue (Sacra estimate, Aug 2026); 100M+ monthly usersNo model of its own; Amazon lawsuitUncertain, likely buyer target
CohereMerging with Aleph Alpha, about $20B combined (Handelsblatt)$240M annual recurring revenue (2025)Model quality: 13 on the indexUncertain, survives as a merger

Funding, users and revenue from company releases, filings and the reporting in Sources. The outlook column is a House of 47 judgement: "Very strong" means a self-funding parent or a top-three model with mass distribution; "Uncertain" means no proprietary frontier model or a dependence on a deal still in progress.

Which to pick for what · October 2026

The right AI for the job

No single model wins everything. Pick by task, and keep a second option for when your first one is down, too slow or too expensive.

If you needFirst pickAlternativeWhy
One assistant for everythingChatGPT PlusGemini, Claude ProThe widest product, GPT-6 Sol on paid plans, 1.2B weekly users
Serious coding and agentsClaude Opus 5.5GPT-6.1 Sol, Codex#1 on the index (58) at $4 / $20; Sol scores 52 at $2 / $10
Maximum capability, cost no objectClaude Fable 5.1GPT-6 AstraBoth $10 / $50, the top tier of each lab
Life inside Gmail, Docs and AndroidGeminiChatGPTBuilt into the apps; 1B+ monthly users
A company on Microsoft 365Microsoft 365 CopilotClaude for Microsoft 365Works inside Word, Excel, Outlook and Teams; $30 per user
Cited research and sourcesPerplexity ProGemini, ChatGPTAn answer engine that cites every claim; $20 a month
Real-time news and XGrokPerplexityLive X data; Grok 4.7 at $2 / $6
Cheapest high-volume APIGPT-6 LunaDeepSeek V4.1 Flash, Gemini 3.8 Flash$0.10 / $0.50; DeepSeek $0.15 / $0.60 off-peak
EU data control or self-hostingMistral Large 4Cohere Command A+European lab, open weights due Oct 2026; Command A+ runs on two H100s
Best open-weights model to run yourselfGLM-5.3Kimi K3, Qwen3.8Scores 45, among the top open models; Kimi K3 44
A small model to fine-tune or run on a laptopQwen3.8-27BMeta Muse Glimmer (30B)Both Apache 2.0; Glimmer fits a 24 GB GPU when quantized
AWS-native workloads and voiceAmazon NovaClaude on BedrockInside AWS billing and security; Nova 2.5 Sonic for speech

Picks are House of 47 recommendations built on the scores and prices above. Claude in Microsoft 365 is listed as a Claude Pro feature on Anthropic's pricing page.

Probability of outcome · House of 47 synthesis

The odds

Synthesis estimates across the sources below, scenario estimates, not market forecasts. The through-line: the top three keep the frontier, open weights keep closing in on price rather than peak, and the weakest independents end up inside bigger companies.

Which model said what · one prompt, four AI models · 8 October 2026

Which model said what

We gave the same research prompt to four AI models on 8 October 2026: Grok 4.7, GPT-5.6 Sol, Claude Opus 5.5 and Gemini 3.1 Pro. Each wrote its own report on the same labs, with a survival score from 1 to 10, a top five and picks by job. Here is what each one said, where they agree and where they split.

Method: one prompt, run once per model on 8 Oct 2026. The prompt asked for a source link behind every number and for honesty about each model's own maker. We compared the four reports side by side and quote their scores and picks as written; we did not re-score them. Where a model's claim conflicts with the figures House of 47 checked for this brief, our checked figures stand. Gemini 3.1 Pro's report has no Microsoft Copilot section and no source links, although the prompt asked for both.

LabGrok 4.7GPT-5.6 SolClaude Opus 5.5Gemini 3.1 Pro
OpenAIGPT-5.6 Sol's maker910910
AnthropicClaude Opus 5.5's maker99910
GoogleGemini 3.1 Pro's maker10101010
SpaceXAI (xAI)Grok 4.7's maker8979
Microsoft9.5108Not covered
Meta9 Meta AI
6 Llama
10710
Alibaba Qwen8.51089
Amazon9 Bedrock
5 Nova
106 Nova
10 Amazon
10
Mistral7867
DeepSeek8879
Moonshot (Kimi)7867
Z.ai (Zhipu)7768
Perplexity6.5857
Cohere6758

Survival scores out of 10, each model's own opinion, as written in its report. Grok 4.7 and Claude Opus 5.5 scored some companies and their model brands separately; both numbers are shown.

Top five and picks, by model

RankGrok 4.7GPT-5.6 SolClaude Opus 5.5Gemini 3.1 Pro
1ClaudeChatGPTChatGPTClaude
2ChatGPTClaudeClaudeChatGPT
3GeminiGeminiGeminiGemini
4DeepSeek, with QwenMicrosoft CopilotQwenMeta
5GrokPerplexityMetaDeepSeek
First pick forGrok 4.7GPT-5.6 SolClaude Opus 5.5Gemini 3.1 Pro
WritingClaudeClaudeClaude Opus or Sonnet 5.5Claude Opus 5.5
CodingClaude Sonnet or Opus 5.5Claude Code, Opus or Sonnet 5.5Claude Opus or Sonnet 5.5Claude Sonnet or Opus 5.5
Research and searchPerplexityPerplexity; ChatGPT Deep Research for long reportsPerplexity; ChatGPT or Gemini deep research for long reportsPerplexity (Sonar Pro, Deep Research)
BusinessMicrosoft 365 Copilot in Office shopsMicrosoft Copilot in Microsoft 365 companiesMicrosoft 365 Copilot on M365; Claude or ChatGPT EnterpriseCohere Command A+
Privacy and self-hostingDeepSeek V4 (MIT), then QwenDeepSeek V4.1 Flash, GLM-5.3-Flash, Tencent Hy4, Command A+Qwen3.8-27B or GLM-5.3-FlashLlama 4 Scout or Qwen3.8-Max
Low budgetFree Gemini and DeepSeek; GPT-6 Luna for API volumeFree tiers; DeepSeek V4.1 Flash and others for APIFree tiers; DeepSeek V4.1-Flash, GPT-6 Luna for APIDeepSeek V4.1-Flash or Gemini 3.8 Flash

Rankings and picks as each model wrote them, shortened. GPT-5.6 Sol and Claude Opus 5.5 ranked complete products; Grok 4.7 ranked defaults that should still be sensible in 2028.

Where they agree

  • All 4
    Google survives

    All four models gave Google 10 out of 10, and all four put Gemini third in their top five.

  • All 4
    The top three are the same

    Every top three is Claude, ChatGPT and Gemini. They differ only on whether Claude or ChatGPT comes first.

  • All 4
    Claude for writing and coding

    Every model named Claude as its first pick for both writing and coding, Claude Opus 5.5 included.

  • All 4
    Perplexity for research

    All four named Perplexity first for quick, cited research. GPT-5.6 Sol and Claude Opus 5.5 added ChatGPT's deep research for longer reports.

  • 3 of 4
    Copilot for Microsoft shops

    Grok 4.7, GPT-5.6 Sol and Claude Opus 5.5 picked Microsoft 365 Copilot for companies on Microsoft 365. Gemini 3.1 Pro did not cover Copilot.

  • 3 of 4
    Claude leads the index

    Grok 4.7, GPT-5.6 Sol and Claude Opus 5.5 all cited Claude Opus 5.5 at 58 on the Artificial Analysis index. Gemini 3.1 Pro cited no benchmark scores.

Where they disagree

  • Split
    Who is number one

    Grok 4.7 and Gemini 3.1 Pro ranked Claude first. GPT-5.6 Sol and Claude Opus 5.5 ranked ChatGPT first, both for breadth rather than benchmark scores.

  • Split
    How generous to score

    Gemini 3.1 Pro gave 10 out of 10 to five labs (OpenAI, Anthropic, Google, Meta and Amazon) and nothing below 7. GPT-5.6 Sol gave 10 to six of the 14. Grok 4.7 and Claude Opus 5.5 gave 10 only to Google, and Claude Opus 5.5 scored Perplexity and Cohere 5.

  • Split
    Meta and Amazon

    GPT-5.6 Sol and Gemini 3.1 Pro gave both 10. Claude Opus 5.5 gave Meta 7 and Amazon's Nova models 6. Grok 4.7 scored the brands lower than the companies: Llama 6, Nova 5.

  • Split
    Cohere

    Scores ran from 5 (Claude Opus 5.5) to 8 (Gemini 3.1 Pro). Claude Opus 5.5 listed Command A+ as closed; the other three said it is open under Apache 2.0, which matches this brief. Gemini 3.1 Pro made Cohere its first business pick.

  • Split
    What to self-host

    Four models, four first picks: Grok 4.7 said DeepSeek V4, GPT-5.6 Sol listed DeepSeek V4.1 Flash among others, Claude Opus 5.5 said Qwen3.8-27B or GLM-5.3-Flash, and Gemini 3.1 Pro said Llama 4 Scout or Qwen3.8-Max.

Claims only one model made

  • Gemini
    "GPT-6 Pro" at $10 / $50

    Gemini 3.1 Pro called OpenAI's top API model "GPT-6 Pro (powered by GPT-6 Astra)". The other three reports and this brief name the $10 / $50 model GPT-6 Astra; Pro is a ChatGPT plan. Likely a naming mix-up.

  • Gemini
    GPT-5.6 Sol at $4 / $20

    Gemini 3.1 Pro listed GPT-5.6 Sol as OpenAI's general model at $4 / $20. This brief lists GPT-6.1 Sol at $2 / $10. Not verified.

  • Gemini
    Meta leads open weights

    Gemini 3.1 Pro called Meta "the absolute king of the open-weight ecosystem" and cited Muse Spark 1.1 and a 10M-token Llama 4 Scout. This brief has Muse Spark at 1.3, proprietary, with the leading open models coming from Chinese labs. The 10M figure is unverified.

  • Gemini
    DeepSeek V4-Pro retired

    Gemini 3.1 Pro said V4-Pro "was retired in late 2026". Grok 4.7 and Claude Opus 5.5 listed V4-Pro with current prices. Not verified.

  • Gemini
    Cohere and Perplexity prices

    Gemini 3.1 Pro gave Command A+ at $0.30 / $1.50 (Claude Opus 5.5 gave about the same) and Perplexity Sonar Pro at $3 / $15 plus search fees. GPT-5.6 Sol said Command A+ pricing could not be verified. Not checked for this brief.

  • Grok
    Gemini 4 unconfirmed

    Grok 4.7 treated Gemini 4 as unverified because Google's API pricing page did not list it, and read that page as showing Gemini 3.5 Flash. This brief, GPT-5.6 Sol and Claude Opus 5.5 all list Gemini 4 Argon in limited release and Gemini 3.8 Flash.

  • Claude
    A 2027 correction

    Claude Opus 5.5 said a 2027 market correction is "more likely than not", and that the top three labs would survive it. Labelled as opinion in its report.

Two Gemini 3.1 Pro figures we checked do hold up: Grok 4.7 at $2 / $6 and Cohere Command A+ under Apache 2.0 both match this brief's sources, and GPT-5.6 Sol reported the same.

Did each model favour its maker?

  • Grok
    Grok 4.7 scored its maker lower

    It gave SpaceXAI 8, below the 9 it gave OpenAI and Anthropic and the 10 it gave Google, and ranked Grok fifth. It disclosed the conflict: "I am that model; the table is still the table." GPT-5.6 Sol and Gemini 3.1 Pro gave SpaceXAI 9.

  • GPT
    GPT-5.6 Sol ranked its maker first

    It gave OpenAI 10 and Anthropic 9, and ranked ChatGPT first, while saying the evidence on raw capability favours Claude. Grok 4.7 and Claude Opus 5.5 gave OpenAI 9. Its report has no conflict-of-interest note.

  • Claude
    Claude Opus 5.5 did not rank its maker first

    It gave Anthropic 9, level with OpenAI and below Google, and ranked Claude second, adding "My maker, so apply a discount". It also disclosed that it ran inside Cursor, which SpaceX owns, and gave SpaceXAI 7, the lowest of the four.

  • Gemini
    Gemini 3.1 Pro: no sign of a lift

    It gave Google 10, as all four models did, and ranked Gemini third behind Claude and ChatGPT. It made no conflict-of-interest statement; it only noted it was "powered by Gemini 3.1 Pro".

Sources: the four model reports of 8 Oct 2026, compared by House of 47. Scores, rankings and claims are the models' own; checks are against this brief's Sources.

FAQ · the questions people ask about AI models

Questions, answered

  • Q
    What is the best AI model in October 2026?

    On the independent Artificial Analysis Intelligence Index, Claude Opus 5.5 leads with 58, followed by Claude Sonnet 5.5 at 56. GPT-6 Astra, Gemini 4 Argon and Claude Fable 5.1 tie at 53.

  • Q
    ChatGPT, Claude or Gemini: which should I pay for?

    ChatGPT Plus ($20) for the broadest all-round assistant, Claude Pro ($20, or $17 a month billed yearly) for coding and long documents, and Gemini if you live in Google's apps. All three have free tiers.

  • Q
    Which AI is best for coding?

    Claude Opus 5.5 is the top-scoring model at $4 / $20 per million tokens. GPT-6.1 Sol (52, $2 / $10) is the value pick, and GLM-5.3 and Kimi K3 are the strongest open-weights options.

  • Q
    What is the cheapest capable AI API?

    GPT-6 Luna costs $0.10 per million input tokens and $0.50 per million output tokens. DeepSeek V4.1 Flash costs $0.15 / $0.60 off-peak, with a 1M-token context.

  • Q
    What is the best open-weights AI model?

    On Artificial Analysis, Xiaomi MiMo-V2.6-Pro leads the open models at 46, then GLM-5.3 at 45 and Kimi K3 at 44. Alibaba's open Qwen3.8 flagship scores 40.

  • Q
    Which AI companies are most likely to survive?

    OpenAI, Anthropic, Google and Microsoft, which combine huge funding or cash flow with mass distribution. Perplexity and Cohere are the most exposed: Perplexity has no model of its own, and Cohere is merging with Aleph Alpha.

  • Q
    Are Chinese AI models safe to use?

    DeepSeek, Qwen, Kimi and GLM publish open weights you can run on your own servers, so no data leaves your infrastructure. Their hosted apps are a separate decision, and the US has weighed restrictions on Chinese models.

  • Q
    Is Meta's AI still open source?

    Partly. Its flagship Muse Spark is proprietary, but Muse Glimmer, a 30B model released in August 2026, is open under Apache 2.0.

House of 47 read · How to think about it

Pick by job, not by brand

The frontier is a three-way race

Anthropic, OpenAI and Google hold every score of 52 or above on the independent index, and they have the money to stay there. Anthropic leads today, but the gap between first and third is five points and changes with each release.

Below the frontier, price is collapsing

A model that scores 38 now costs $0.50 per million output tokens, and DeepSeek and GLM sell near-frontier work for a few dollars. Our numbers on how fast that cost is falling are in The collapsing cost of intelligence.

Distribution decides the consumer market

ChatGPT, Gemini and Meta AI reach billions because they sit where people already are. As these assistants replace the search box, being the source they quote becomes the new SEO. See AI Search vs Google, and the GEO work of Resocial.

Never depend on one model

Six frontier releases landed in five weeks this autumn, from Claude Fable 5.1 on 1 September to Mistral Large 4 on 6 October, and prices moved with each one. Route work across two or three providers and own the workflow, not the model. Building agents that do exactly that is the business of theagency47, and where it leads is in The AI Economy in 2030.

Base case

Plan for a frontier held by Anthropic, OpenAI and Google, with xAI and Meta close behind. A cheap, open Chinese pack sets the floor on price, and Mistral and Cohere serve buyers who need sovereignty. Perplexity-style apps and weaker labs consolidate into larger companies.

Tail

Stress-test for the swings. A US ban on Chinese models, a compute or funding crunch at a single lab, or an open-weights model taking first place would each reshape this table quickly. Keep a second provider ready and contracts short.

The best model changes every month. What lasts is never betting everything on one.

Sources · figures checked against primary and industry reporting · accessed 8 October 2026

Sources

Artificial Analysis 2026, LLM leaderboard & Intelligence Index (accessed 8 Oct), leaderboard

OpenAI 2026, GPT-6 for everyone, 1.2B weekly users, announcement

OpenAI 2026, API pricing (GPT-6 Astra, Luna), pricing

OpenAI 2026, GPT-6.1 Sol ($2 / $10, 1.05M context), announcement

OpenAI 2026, $122B at $852B, $2B monthly revenue, announcement

OpenAI Help Center 2026, ChatGPT Pro tiers, help article

TechCrunch 2026, the $100 ChatGPT Pro plan, report

Anthropic 2026, Claude Opus 5.5, model page

Anthropic 2026, Claude Fable and Mythos 5.1, announcement

Anthropic 2026, Claude plans & API prices, pricing

Anthropic 2026, Series H, $65B at $965B, announcement

Reuters 2026, Anthropic run-rate tops $65B (source), report

Google 2026, Gemini 4 Argon, announcement

Google 2026, Gemini 3.8 Flash, announcement

Google 2026, Gemini app passes 1B monthly users, announcement

Google 2026, Google AI plans, subscriptions and I/O update

xAI 2026, Grok 4.7, announcement and docs

Digital Applied 2026, SpaceX S-1: Grok users, xAI financials, plans, analysis

SpaceX 2026, IPO closing (about $85.7B gross), release

Ars Technica 2026, Meta's Muse Spark goes proprietary, report

Meta 2026, Muse Spark API tiers, developer page

Meta 2026, Muse Glimmer, open 30B model, announcement

Meta 2026, Q2 results (revenue, capex, 3.60B DAP), investor relations

Mistral AI 2026, Mistral Large 4, announcement and docs

Mistral AI 2026, Vibe plans, pricing

TechCrunch 2026, Mistral raises €3B at over €21B, report

DeepSeek 2026, API pricing (V4.1 Flash, V4 Pro), pricing

Hugging Face 2026, DeepSeek-V4 release, blog

Reuters 2026, DeepSeek closes over $7B, report

Reuters 2026, filing implies DeepSeek at about $52B, report

Reuters 2026, US holds off blacklisting DeepSeek, report

Alibaba Cloud 2026, Qwen3.8 weights, 460+ models, 3B+ downloads, blog

Alibaba Cloud 2026, Qwen app 300M+ MAU, cloud +36%, blog

Microsoft 2026, FY26 Q4 results (30M+ Copilot seats, Azure), release

Microsoft 2026, Microsoft 365 Copilot pricing, enterprise and business

Sacra 2026, Perplexity revenue, valuation, plans, profile

Cohere 2026, Command A+, announcement and pricing

The Next Web 2026, Cohere and Aleph Alpha merger (~$20B, $240M ARR), report

Moonshot AI 2026, Kimi K3, announcement, API pricing and memberships

Reuters 2026, Moonshot files confidentially for HK IPO, report

Tom's Hardware 2026, US push to restrict Chinese AI models after Kimi K3, report

Amazon 2026, Amazon Nova models, 50,000+ customers, overview

AWS, Nova 2 Lite model card, documentation

Z.ai 2026, GLM-5.3, announcement and pricing

Reuters 2026, Zhipu's Hong Kong debut, report

Changelog · what changed in this brief, newest first

Changelog