Models

Anthropic's frontier model and the current intelligence leader; excels at agentic coding and long-horizon reasoning.

Balanced speed/quality workhorse of the Claude 4 family.

Fast, low-cost Claude tier with near-frontier coding.

OpenAI's flagship multimodal reasoning model (figures shown for the high reasoning-effort config).

Prior OpenAI flagship; still available but scheduled for API retirement in Aug 2026.

Google's most capable Gemini (preview); tops several reasoning benchmarks.

Google DeepMind's frontier multimodal, long-context model.

Fast, low-cost tier of the Gemini 3 line (preview).

Open-weights MoE flagship (1.6T total / 49B active) with built-in reasoning modes (MIT).

Efficient open-weights V4 tier (284B total / 13B active), MIT-licensed.

Moonshot's trillion-param MoE (32B active); strong agentic coding (Modified MIT).

Zhipu/Z.ai open-weights flagship (754B total / 40B active), MIT.

Alibaba's dense open-weights Qwen3.6 (Apache-2.0); multimodal.

Meta's open-weights MoE (~17B active); the current open Llama flagship.

Mistral's open-weights MoE flagship (41B active, Apache-2.0); a non-reasoning model.

xAI's current flagship; ranks high on the Artificial Analysis Index (xAI publishes few classic benchmarks).

Anthropic's prior Opus flagship (superseded by 4.8); still strong on agentic coding.

Earlier Opus 4.6 frontier model; still generally available.

OpenAI's GPT-5.4 — the tier between GPT-5.2 and GPT-5.5; reasoning + tool use.

Google's current fast/default Gemini (GA); strong agentic performance.

Prior-generation Gemini flagship; widely used, slated for deprecation.

xAI Grok 4.20 (reasoning); multi-agent 'council' architecture, very long context.

Open-weights MoE (671B total / 37B active), MIT; predecessor to V4.

Moonshot's Nov-2025 long-horizon reasoning MoE (1T total / 32B active), Modified MIT.

Alibaba's proprietary Qwen3.7 Max flagship; agentic-focused.

Open-weights hybrid-MoE Qwen3.6 (35B total / 3B active), Apache-2.0.

Meta Superintelligence Labs' first proprietary frontier model (free via Meta AI).

Open-weights MoE (~17B active / 109B total) with very long context.

Amazon's Nova 2 Pro flagship on Bedrock; hybrid reasoning, multimodal.

NVIDIA's open-weights hybrid Mamba-MoE (550B total / 55B active), OpenMDW license.

MiniMax's M3 flagship; tops the Artificial Analysis Intelligence Index (open weights pending).

ByteDance's Doubao Seed 2.0 Pro; multimodal flagship served via Volcano Engine.

Self-supervised video joint-embedding predictive architecture; learns world dynamics for motion understanding, action anticipation, and zero-shot robot planning.

Foundation world model that generates interactive, controllable environments in real time. Shown qualitatively — DeepMind publishes no standardized cross-model benchmark numbers.

NVIDIA's open-weights World Foundation Model for physical AI (future-frame prediction for robotics and AV); the 2B variant is independently scored on PAI-Bench-G.

World Labs' spatially-grounded generative model — persistent, editable 3D scenes from text, images, or video. Marketed on persistence/editability; no standardized benchmarks published.

OpenAI's text/image-to-video model — the system that popularized 'video as world simulator'. The original Sora was independently scored on Physics-IQ, where generative physical understanding proved severely limited.

OpenAI's flagship video+audio model (the 'GPT-3.5 moment for video'); a large jump in physical plausibility over the original Sora. Consumer app retired Apr 2026, API through Sep 2026.

Google DeepMind's text/image-to-video model with native audio and real-world physics; among the top performers on the PAI-Bench physical-AI generation benchmark.

Alibaba Tongyi Lab's open-weights MoE video generator (Apache-2.0); the I2V-A14B variant leads the open models on PAI-Bench-G.

Tencent's 13B open-weights video foundation model; the image-to-video variant is benchmarked on PAI-Bench-G.

Luma's reasoning-driven video model with 3D-aware generation and native HDR. Demos only — no standardized cross-model benchmark results published.