AI Index

Meta’s new reasoning and coding family

Meta launched Muse Wonder and Muse Wonder Full on 12 October. Wonder has a one-million-token context and a developer preview at $4 input and $20 output per million tokens. Wonder Full starts in the Power and Maximum Muse Code plans.

Anthropic's frontier family

Anthropic recommends Opus 5.5 as a starting point for most work at $4 input and $20 output per million tokens. Sonnet 5.5 costs $2/$10, with cache reads reduced to $0.10 on 7 October. Haiku 5.5 launched that day: $0.10/$0.50 for prompts up to 100,000 tokens, rising to $0.50/$2.50 above that threshold. Fable 5.1 remains the $10/$50 tier for demanding reasoning. The 5 October Artificial Analysis observation remains dated below.

OpenAI's new generation: Astra, Sol and Luna

GPT-6 Astra began rolling out on 3 September 2026 at $10 in and $50 out per million with a 1,050,000-token context, the first model OpenAI classifies Critical for cybersecurity. GPT-6 Sol ($2/$10) and GPT-6 Luna ($0.10/$0.50) followed on 22 September at half the GPT-5.6 prices, with Luna reaching free ChatGPT users in the desktop app. GPT-6.1 Sol, an upgrade to Sol, arrived on 29 September at the same $2 in and $10 out, with cached input at $0.10, a 1,050,000-token context and a 128,000-token output limit. GPT-5.6 Terra stays on sale at $2/$12. OpenAI began rolling out tuned Sol in Chat to paid plans on 7 October and Luna to Free and Go on 8 October. This Chat release leaves Work and Codex models unchanged.

Google DeepMind’s new frontier model

Google announced Gemini 4 Argon on 30 September 2026 with a one-million-token output limit and introductory API prices of $2 input and $10 output per million tokens. Access starts with trusted cyber defenders and testers, followed by a planned broader release beginning with paid API customers and Google AI Ultra subscribers.

Grok 4.7

xAI's flagship line

Grok 4.7, released 21 September 2026, holds the 500K context at $2 in and $6 out per million tokens, the same price as Grok 4.6. A prompt past 200,000 tokens re-prices the whole request at the long-context tier. On xAI's own seven-benchmark launch table it beats Grok 4.6 on every row and takes the best score on two, against four for Fable 5.1 Max at $10 and $50.

Moonshot AI's open flagship

Kimi K3, a 2.8-trillion-parameter flagship, launched on 16 July; its full weights landed on 27 July as the largest open-weight release yet, roughly 1.4TB even at four-bit precision, under a bespoke licence with a revenue clause for model-as-a-service hosts. On the Artificial Analysis index, read 2 October, it is third among open-weight models on 44, behind MiMo-V2.6-Pro on 46 and GLM-5.3 on 45. Kimi K2.8 Preview (11 September) rolled out across Kimi Code with the model ID unchanged, close to K3 on Moonshot's own description.

Alibaba's flagship line

Alibaba's flagship: 2.4 trillion parameters with 95 billion active, a 1M-token window, $2/$6 per million tokens, launched 3 August 2026. The promised open weights landed on the official Qwen Hugging Face pages around 13 August, in standard and FP8 form, with a 27B sibling alongside.

GLM-5 series

Z.ai's agent-endurance models

GLM-5.3 (14 August 2026) runs the GLM-5.2 base with heavier post-training, and its full 753B weights landed on Hugging Face on 28 August under a custom glm-5.3 licence rather than MIT. GLM-5.3-Flash (26 August, MIT, 320B with 18B active, 1M context) is the open workhorse, and the model Z.ai had been testing anonymously as ox-alpha.

The price-floor setter

DeepSeek-V4.1-Flash arrived on 10 September 2026 under an MIT licence: 552 billion backbone parameters, 8 billion active reading a prompt and 16 billion writing one, a 1M-token context and native image reading. Its global KV cache holds 890 bytes per token, and output runs $0.6 per million off-peak against $1.2 at peak. V4-Pro-0813 remains on the API at its own rates.

Meta's proprietary frontier

Meta Superintelligence Labs' first Muse model (8 April 2026) reasons with parallel agents in Contemplating mode; Muse Spark 1.1 (9 July) added a 1M context, tool and computer use, and a paid API. On 11 August 2026 Meta opened Muse Glimmer under Apache 2.0, dropping the Llama licence rules.

OpenAI's reasoning image line

OpenAI's image line holds the arena's top three seats, with the Images 2.5 Sunburst and Flare variants first and second and GPT Image 2 third. The first mainstream model family that reasons about a brief, checks references and self-corrects before rendering.

Google's editing powerhouse

Gemini 3 Pro's image engine: native 4K, best-in-class edits and character consistency. Nano Banana 2 covers the fast-and-cheap end of the family.

Midjourney V8.2

The art direction standard

Still the strongest pure art direction. V8.2 took the default seat on 24 July with Personalization profiles as the headline upgrade, on top of V8.1's native 2048px output; Niji 7 handles anime styles.

FLUX.2

Black Forest Labs' open family

The open-weight family to beat: Pro, Flex, Dev and Klein tiers, camera-accurate lighting and up to ten reference images for character consistency.

Ideogram

Text rendering specialist

The reliable choice when the image must carry words: posters, logos, packaging and layout-aware design work.

Recraft V4

Design-first generation

Built for designers rather than prompters: vector output, exact brand colours and print-ready renders straight from the model.

ByteDance's value line

ByteDance's image line delivers 4K output at some of the lowest per-image prices on the market, with strong bilingual text rendering.

MAI-Image-2.6

Microsoft AI's image model

Microsoft AI's in-house image model took second place on the arena text-to-image board at its 10 August launch, ahead of Google, Meta and xAI, with text rendering up 91 Elo on Microsoft's own figures and a focus on product, branding and photoreal work; on the board read 27 September it sits fourth on 1335.

Firefly Image 5

Adobe's licensed-data model

Trained on licensed data with indemnification for enterprise use, and now a hub that also fronts partner models inside Creative Cloud.

ByteDance's board contender

Seedance 2.5 sits second on the video-editing board, four points behind Wan 3.0, and fourth on image-to-video. One model generates picture and sound together in 30-second single passes, so lips, foley and music actually match the frame.

Google DeepMind's flagship

Native 48kHz dialogue, 4K output and SynthID provenance marks. The default choice where brand safety and clean licensing decide the choice. Its Gemini API previews shut down on 22 October 2026, with Gemini Omni 1.1 Flash the named replacement; the Vertex AI models run to 17 November 2026 or later.

Kuaishou's value flagship

Native 4K at 60fps, 15-second multi-shot storyboards and lip-sync in five languages, at roughly $0.10 a second.

OpenAI · retired

Retired: app and web access closed on 26 April and the API shut down on 24 September 2026. The page is a migration guide for the teams that built on it.

Runway Gen-4.5

The filmmaker's control surface

No longer top of the leaderboards but still the best control surface in the business: motion brushes, keyframes and a real editing timeline.

Ray3 · Dream Machine

Luma's HDR pioneer

First native 16-bit HDR video model, with Ray3 Modify for video-to-video work inside the Dream Machine 2.0 studio.

Alibaba's video family

Wan 3.0 (24 August 2026) generates 30-second single-pass clips, takes documents, slides and web pages as input, and tops the video-editing arena; its weights are not yet published. Wan 2.7 runs on hosted APIs, and the openly published weights of the family stop at Wan 2.2 under Apache 2.0.

LTX-2.3

Lightricks' open 4K model

A 22B open model producing native 4K at 50fps with stereo audio, under the LTX-2 Community License, free for organisations below $10m in annual revenue. Its successor LTX-2.5, also 22B, has had open weights since 23 July 2026.

MiniMax's motion specialist

MiniMax's line is prized for expressive, physical motion: the model animators reach for when movement has to feel alive. Its H3 model now tops the image-to-video arena.

FLUX 3 Video

Black Forest Labs' video debut

Black Forest Labs' first video family, rolled out in early August, sits third on the text-to-video arena on 1,493 Elo behind Google's two Omni Flash models and a point ahead of Grok Imagine Video 1.5's agent mode, with Wan 3.0 sixth (board read 27 September). FLUX 3 Action, the 7B robot-control model, followed on 23 September.

conversational video generation

Google's conversational video line. Omni 1.1 Flash (27 August 2026) adds end frames, scene extension to a cumulative 40 seconds, a 360p draft mode at a third of the cost, and 1080p or 4K export; it tops the text-to-video arena and sits second on image-to-video.

OpenClaw

Self-hosted agent gateway

The breakout self-hosted gateway: one daemon, twenty-plus messaging channels, skills and scheduled jobs, any model behind it. Since April it needs API billing rather than a Claude subscription, and it should be hardened before exposure.

Hermes Agent

Self-improving open agent

Nous Research's open-source, self-hosted agent, and the one that changes as you use it: finishing a task writes a reusable skill back into the agent. Version 0.20.6 (27 August) added consent-gated real-profile browsing from a managed copy of your Chrome profile, off by default; the Herald release (3 August) added real-time voice and an agent-to-agent protocol.

Merged into Claude

Folded into one Claude on 16 September 2026, rollout starting with Pro and Max: the file-and-app agentic work Cowork carried now runs in ordinary Claude chat, with Claude Docs and Slides launched in beta alongside. Its original launch is what markets nicknamed the SaaSpocalypse.

Google's always-on agent

Google's always-on agent living in a cloud VM: deep Gmail, Docs and Calendar access, MCP connectors from day one, and it keeps working while your laptop sleeps.

OpenAI's consumer agent

Agent mode inside ChatGPT with its own virtual computer for browsing, files and multi-step tasks under supervision.

Anthropic's builder stack

The loop behind Claude Code, embeddable in your own product, alongside a managed runtime (beta) where Anthropic hosts the agent for you.

Handoffs and guardrails

OpenAI's production framework with explicit handoffs, guardrails and tracing; MCP support landed early this year.

The connector standard

The USB-C of agent tooling: one protocol that lets any client use any tool or data server. Every major platform now speaks it.

One API, every model

One API over every major model with live rankings drawn from real usage: the quickest way to A/B models behind an agent without rewrites. Its Ori CLI (4 August 2026) tunes Claude Code, Codex, OpenCode and Hermes to run through the gateway. Stripe agreed to acquire it on 19 August; name, product and roadmap stay.

OpenAI's finished-work agent

OpenAI's agent for long-running jobs (9 July 2026): it works for hours and returns finished documents, spreadsheets, slides and web apps. It launched on GPT-5.6, and since 29 September Plus, Pro, Business, Enterprise and Edu users can run it on GPT-6.1 Sol.

Anthropic's terminal agent

Terminal-native and repo-aware, with Dynamic Workflows for long unattended runs; Anthropic's docs list Opus 5.5 as its default on Pro, Max, Team, Enterprise and the Anthropic API, with Fable 5.1 the tier above it, picked by hand. Included with a Claude Pro plan.

SpaceX's AI-native editor

The AI-native editor, a SpaceX company since 14 August 2026: Tab completions plus Composer cloud agents that work branches in parallel VMs, running Claude, Gemini and Grok, with OpenAI models due to leave on 12 November unless the wind-down is resolved.

The installed-base leader

Still the default inside VS Code and enterprises, with a free tier, IP indemnity and new flexible usage billing from 1 June.

OpenAI's cloud engineer

OpenAI's asynchronous cloud engineer: give it an issue and it returns a tested pull request. Since 29 September OpenAI's Codex docs point complex coding at GPT-6.1 Sol and focused, repeatable tasks at Luna, with GPT-5.6 Sol, Terra and Luna kept during the rollout.

Google's agent-first IDE

Google's agent-first IDE built around the Gemini 3.5 line, with a browser-testing loop and a generous free tier via a Google account.

Devin Desktop

Windsurf's successor

Windsurf's next chapter after the Cognition deal: the Devin autonomous engineer wrapped in the familiar editor, relaunched in June.

AWS's spec-driven agent IDE

Amazon's take: spec-first development where agents implement against written requirements, aimed at teams that want process over improvisation.

Prompt to deployed app

From prompt to deployed app in one place: environment, database and hosting handled for you. The fastest route for non-specialists.

Maker of Claude

Fable 5.1 and Mythos 5.1 (1 September) up top, Opus 5.5 (22 September) at $4/$20, which Anthropic says performs at Fable 5.1's level on most work, Sonnet 5.5 (28 September) at $2/$10, Claude Code and, since the 16 September merge, one Claude carrying the product story. A Life Sciences Verification Program opened on 17 September, swapping real-time blocking for usage monitoring at vetted organisations. May's round valued it at $965B.

Maker of ChatGPT

The GPT-6 generation now leads its range: Astra on 3 September 2026, Sol and Luna on 22 September and GPT-6.1 Sol on 29 September. ChatGPT Work, launched beside GPT-5.6 on 9 July 2026, runs for hours and returns finished documents, spreadsheets and web apps. Astra for Law followed on 17 September 2026: GPT-6 Astra over a 230-million-URL case-law index, with Sullivan & Cromwell, Ropes & Gray and Cooley named as early builders.

The full-stack player

Gemini 3.1 into 3.5 and now Gemini 4 Argon, announced on 30 September and opening with cyber defenders, Nano Banana for images, Veo and Omni Flash for video, Spark for agents, and TPUs underneath it all. Since early August, Demis Hassabis chairs while Koray Kavukcuoglu leads day to day.

SpaceXAI (xAI)

Maker of Grok

Rebranded SpaceXAI on 6 July 2026 after February's merger into SpaceX and a June Nasdaq listing. SpaceX's $60bn all-stock purchase of Anysphere, maker of Cursor, closed on 14 August, and NVIDIA says Grok's next agentic workloads will run on its Vera Rubin platform.

Maker of Kimi

Kimi K3, a 2.8-trillion-parameter flagship, launched on 16 July 2026 and its full weights landed on 27 July, the largest open-weight release yet. On the Artificial Analysis index read on 2 October it is third among open-weight models on 44, behind MiMo-V2.6-Pro on 46 and GLM-5.3 on 45.

China's broadest stack

Ships the open Qwen line, and delivered on the Max-tier promise: Qwen3.8-Max's 2.4-trillion-parameter weights landed on Hugging Face on 8 August 2026. QwenWork, its all-in-one agent workplace, opened an international beta on 26 August; China's anthropomorphic-agent rules took Qwen's human-like companions offline in July.

The media-model power

Dominant in media models: ByteDance develops Seedance video models; preference rankings depend on the board and observation date.

The price-collapse lab

The lab whose open releases forced the whole market to justify its margins; its next-generation signals keep that pressure on.

Meta's frontier lab

Moved Meta's frontier work from open-weight Llama to the proprietary Muse family, and began charging developers for Muse Spark 1.1 through the Meta Model API on 9 July 2026.

Anthropic's guides and courses

Model guides, prompting best practice and the Agent SDK reference, plus free certification courses.

Runnable API recipes

Hands-on, runnable examples for the API, agents and evals; the fastest way from idea to working call.

Free Gemini playground

Free Gemini playground with generous limits: prototype prompts, tune settings and export working code.

Home of open models

The home of open weights: models, datasets, Spaces demos and the free training courses behind them.

The connector standard's docs

The spec, SDKs and server gallery for the connector standard every agent platform adopted.

Artificial Analysis

Independent model benchmarks

Independent, continuously updated comparisons of quality, speed and price across every major model.

LMArena

Human preference leaderboard

Blind head-to-head votes from millions of users; the leaderboard that reflects taste rather than test sets.

Stanford AI Index

The annual state of AI

The definitive yearly measurement of the field: capability trends, economics and policy in one report.