A.I.

Rycode
Tue 27 Oct Due 31 Oct · DeepSeek starts billing Search and Fetch calls from server-side tools in Harness, Claude Code and Codex Due 31 Oct · Mistral Large 4's open weights are due by the end of October Due 9 Nov · Anthropic is expected to start pitching its IPO to investors the week of 9 November Due 4 Nov · Kimi K3.1 Pro weights promised by Moonshot Due 27 Oct · LiteLLM plans to disclose a separate, already-fixed security flaw Due 31 Oct · Mistral Large 4 open weights are due by the end of October Anthropic files its public S-1 and plans to list on Nasdaq as ANT Anthropic S-1: $16.3B first-half revenue; $80.7B net loss, mostly non-cash Anthropic S-1: up to $120.5B with SpaceXAI through 2029 for Nvidia GPU compute OpenAI ships GPT-6.1 Terra at $1/$5; Terminal-Bench 4.0 lab claim 50.3%

Rycode · /ai · leaderboard

Flagships, every score

Higher is better · darker = higher in its column · click a header to sort · Lab company's own result · AA measured by Artificial Analysis · ARC Prize verified by the ARC Prize organizers · Dash = not reported at this version: no score from the model's own lab at this column's test version

↗ ↗ ↗ ↗ ↗ ↗ ↗ ↗
Claude Opus 5.5 58 AA · max 57.8% Lab - 66.4% Lab 81.8% partial 1 Lab 40.0% Lab 67.7% Lab -
Claude Sonnet 5.5 56 AA · max 55.5% Lab - 70.6% Lab ----
Kimi K3.1 Pro 55 AA - 75.5% Lab 65.7% Lab -- 70.0% Lab -
Claude Fable 5.1 53 AA · max 73.4% 2 Lab - 55.8% Lab 77.9% partial / 41.7% strict 1 Lab 31.4% Lab 65.0% Lab -
GPT-6 Astra 53 AA · max - 74.1% Lab 57.9% Lab 72.6% 1 Lab 41.4% Lab 57.2% Lab 62.7% 34 ARC Prize
Gemini 4 Argon 53 AA · high - 77.9% Lab -- 51.3% Lab --
GPT-6.1 Sol 52 AA · max - 75.2% Lab -----
Muse Wonder 52 AA · max - 78.8% Lab -----
GPT-6.1 Terra 48 AA · max -- 50.3% Lab ----
Grok 4.7 46 AA · xhigh 46.3% Lab 71.0% Lab 38.0% Lab ----
Claude Haiku 5.5 43 AA · max -- 39.2% Lab -- 57.4% Lab -

Source: Artificial Analysis · as of 26 Oct 2026

Latest additions: Kimi K3.1 Pro (added 24 Oct 2026); GPT-6.1 Terra (added 26 Oct 2026). Board as of 26 Oct 2026.

  1. These OSWorld scores use different 2.0 versions, do not compare directly: Claude Opus 5.5 on 2.0, Claude Fable 5.1 on 2.0 Aug 2026 task release, GPT-6 Astra on 2.0 latency sim.
  2. Claude Fable 5.1 is scored on CursorBench 3.2.0, a different version from the others, do not compare directly.
  3. ARC-AGI-3 Standard harness: a harness is the setup a model runs inside during a test. The Standard harness lets the model carry forward only the notes it chooses to keep. GPT-6 Astra ran at max effort.
  4. Special setup for GPT-6 Astra: 99.9% with the Provider Adapter harness (high effort), which keeps the model's hidden reasoning between steps. OpenAI's own post also reports 99.9%, run with its Responses API harness.
AA Intelligence Index
one overall score from a fixed set of tests that Artificial Analysis runs itself.
CursorBench
coding tasks from real use of the Cursor editor.
DeepSWE
long coding tasks in real software projects, written from scratch for the test.
Terminal-Bench
completing tasks in a command-line terminal.
OSWorld
using a desktop computer through its screen, mouse and keyboard.
AutomationBench
Zapier's test of multi-step business automation tasks.
Humanity's Last Exam
very hard expert-level questions across many subjects.
ARC-AGI-3
the ARC Prize test of learning and beating new games without instructions.

Table reads public/ai-benchmarks.json. AA column: score and effort setting.

Score updates

New model pages on Artificial Analysis, an independent group that runs the same tests on every model.

Labs

Anthropic Claude frontier lineup. OpenAI GPT frontier lineup. xAI Grok frontier lineup. Google Gemini / Gemma lineup. Meta Muse and Llama stacks. DeepSeek V4.1 Flash and Pro API line.

Lab pages show current models only. Older models are in the model archive.

Labs & papers

Lab homes

New releases

Model cards on this site and downloadable model weights.

Lab posts

Announcements and recaps on each lab's own blog. These are the lab's claims.

Agent tools

Personal agents that work in the background on a cloud computer and through your apps.

  • Claude Duet Anthropic Always-on agent on a cloud VM for Pro and Max.
  • Muse Meta (Meta Superintelligence Labs) Personal agent in a secure VM; free tier plus paid plans.
  • dots OpenAI Always-on ChatGPT agent with its own cloud computer.
  • Grok Bot SpaceXAI (xAI), with Cursor Named Bots on a shared cloud computer and desktop apps.

All agent tools

Products and features

Apps, agents and features built on top of models. They are not models themselves.

Open weight

Open-weight frontier Downloadable weights near the closed frontier. Llama stays under Meta. 5 models

More labs

Mistral · Cohere · Thinking Machines · Apple · Amazon Current flagship from each, from the lab's own docs. 5 models

System One

TypeSafe · Jev Typed decisions for software, not a chat flagship.

Deep dive

  1. A next-token engine with an RLVR post-train and a tool loop. How staged training and tools shape a next-token engine.