Frontier AI Laboratories
Frontier research laboratories driving machine intelligence forward. Each organization maintains a primary foundation model alongside focused experimental checkpoints.
Creator of the Claude family; pioneers of adaptive thinking, Enterprise Frontier Safeguards (EFS), and agentic knowledge work.
Released Sept 22, 2026; Anthropic's flagship with Fable-level agentic coding at 40% lower cost than Opus 5 ($4/$20), beating GPT-6 Astra on FrontierCode at a fifth of the cost per task.
Released Sept 28, 2026; 70.6% on Terminal-Bench 4.0 versus 10.3% for Sonnet 5, running 30%+ faster at the same $2/$10 pricing.
Pioneered the AGI era with GPT-6 Astra, the GPT-6 Sol/Luna efficiency tiers and near-flagship GPT-6.1 Sol, and the Astra critical-capability cybersecurity model.
Released Sept 29, 2026; upgrade to GPT-6 Sol with near-Astra coding and computer use at one-fifth of Astra's price ($2/$10).
Released Sept 22, 2026; high-volume tier at $0.10/$0.50 - half the 5.6 Luna price - matching GPT-5.6 Sol on factuality at ~1% of the cost.
Creators of the Gemini 4 family, led by Gemini 4 Argon for long-horizon software engineering, enterprise knowledge work, cyber defense, and 1M-token output at $2/$10 per 1M tokens.
Announced Sept 30, 2026; Gemini 4 Argon is Google DeepMind's frontier model for long-horizon software engineering, enterprise knowledge work, and cyber defense, with 1M output tokens and an introductory price of $2/$10 per 1M tokens.
Released Sept 15, 2026; Google's native speech-to-speech model that reasons and talks at once, with background tools, 97-language switching, and #1 82.6 on the Speech-to-Speech Quality Index.
Elon Musk's lab powered by the Memphis Colossus supercluster, creators of Grok 4.7 flagship for coding and knowledge work with 500K context.
Released Sept 21, 2026; xAI's flagship for coding and knowledge work on a larger base with longer RL, matching Grok 4.6 price/speed ($2/$6) with 46.3% CursorBench 4.0 and 71.0% DeepSWE v1.1.
Released Sept 18, 2026; xAI's speech-to-text 2.0 ranked first for accuracy among 32 streaming models on Artificial Analysis, at $0.10/hr batch and $0.20/hr streaming.
Open-source frontier champion behind the DeepSeek-V4.1 family (552B Causal Encoder-Decoder MoE, native multimodal, MIT-licensed) with DSpark speculative decoding and 1M context.
Released Sept 10, 2026; first of DeepSeek's Causal Encoder-Decoder family with native image+text, 1M context and 384K output. Replaces V4-Flash and V4-Pro from Sept 14.
DeepSeek's primary 1.6T parameter powerhouse with 49B activated per token, configurable thinking budget, and 1M context.
Global AI foundation leader with the newly launched Muse Wonder frontier agentic models (Oct 12, 2026) with Active Memory and Active Focus, plus Muse Spark 1.3 and the Llama 4 family.
Released Oct 12, 2026; Meta's first frontier model with Active Memory and Active Focus for personal and professional agents, frontier coding and computer use at $4/$20 per 1M tokens.
Announced Sept 1, 2026; Meta's streaming speech-to-text model with diarization and endpointing, trained on 70+ languages (25 validated), billed at $3.00 per 1,000 audio minutes ($0.18/hr).
Massive open and proprietary foundation ecosystem, leading global charts with the Qwen3.8 series and 2.4T open MoE models.
The largest open-weight MoE model in existence. 2.4 Trillion parameters with 95B activated per token and 1M context, available on HuggingFace and ModelScope.
Released August 26, 2026; open-weight multimodal MoE (125B + 51B N-gram, 6B active) with 1M context and 131K output at $0.16/$0.47.
European frontier lab behind Mistral Large 4 (1.05T MoE, 49B active, native multimodal, 1M context) - new open-weight flagship in preview API with weights 27 Oct 2026 - alongside Medium 3.5, Small 4, and Devstral 2.
Released Oct 6, 2026 in public preview on Mistral Studio ($1.36/$4.18); Mistral's largest model - 1.05T total / 49B active MoE with 1.6B vision encoder, native multimodal, 1M context, 160+ languages; SOTA open-weight outside China for cyber, finance and legal with 61.7% DeepSWE v1.1; weights drop Oct 27.
Dense 128B multimodal instruction-following model with native text and image understanding, tuned for European enterprise compliance.
Creator of the Hunyuan Hy3 and Hy4 series; current OpenRouter volume leader with efficient open MoE models for reasoning and agentic workflows.
Tencent's stable production flagship; 295B MoE with configurable no-think/low/high reasoning, grounded anti-hallucination behavior, and sustained 30T+ monthly OpenRouter volume.
Current #1 on OpenRouter by weekly tokens with +639% growth; preview of Tencent's next-generation reasoning model for agentic workflows.
Builders of the GLM family; dominating high-volume inference with GLM 5.3 Flash and GLM 5.2 at aggressive price-performance.
Z.ai's primary flagship; high-throughput foundation model with strong coding share and large-document handling at scale.
Top-5 OpenRouter volume driver; ultra-cheap large-context model for high-volume extraction, long documents, and agent fleets.
Creators of MiniMax M3, a favorite for autonomous agents and batch processing with a strong balance of reasoning depth and throughput.
Top-5 OpenRouter volume; the default for OpenClaw-style autonomous agents, balancing reasoning depth with batch throughput.
MiniMax's reigning foundation flagship is also its newest deployed model.
Builders of the Nemotron family of open, high-throughput models including the 550B Nemotron 3 Ultra for large-scale inference.
550B open model with a generous free tier; top-10 OpenRouter volume for high-throughput open inference and experimentation.
NVIDIA's reigning foundation flagship is also its newest deployed model.
Creators of MiMo-V2.6 Pro, the strongest open model (AA Index 46) with scaled RL self-improvement, omni-modal reasoning and 1M context.
Released Sept 22, 2026; Xiaomi's flagship reasoning model scaled with 6-day Live RL (750K trajectories), scoring 46 on AA Index as strongest open model with 72.6% DeepSWE v1.1 at unchanged V2.5 pricing.
Released Sept 22, 2026; full-modality low-cost reasoning model outperforming MiMo-V2.5 Pro, with 65.7% DeepSWE v1.1 at $0.14/$0.28 and open weights.
Builders of the Kimi family, widely used for agent-swarm orchestration and long-context production workloads.
Top-12 OpenRouter volume; agent-swarm orchestration pick with ~90% cache hits making effective cost far lower than sticker price.
Moonshot AI's reigning foundation flagship is also its newest deployed model.
Kuaishou's video generation lab behind Kling 3.0 - the value leader in AI video with native audio and multi-shot generation.
Kuaishou's Feb 2026 video flagship: native multilingual audio, motion brush and best-in-class image-to-video at ~$0.84 per 10s clip.
Kuaishou Kling's reigning foundation flagship is also its newest deployed model.
Creative-tools company behind Gen-4.5 - the director-grade video model topping text-to-video arenas with motion brush and camera control.
Launched Dec 2025 with 1,247 Elo - No. 1 in text-to-video: director camera moves and reference characters at ~$1.49 per 10s clip.
Runway's reigning foundation flagship is also its newest deployed model.
Builders of Sarvam 105B sovereign MoE, SOTA across 22 Indian languages, trained from scratch in India under IndiaAI mission.
Released Mar 6, 2026; India's sovereign 105B MoE trained from scratch on 12T tokens, SOTA across 22 Indian languages with 98.6 Math500 and 49.5 BrowseComp.
Sarvam AI's reigning foundation flagship is also its newest deployed model.
Builders of System One decision models; first model Jev returns typed calibrated decisions in 70-500ms with free output tokens.
Released Sept 15, 2026; TypeSafe's first System One model returns typed decisions with calibrated confidence in 70-500ms, at $0.042 per 1M input tokens with free output.
TypeSafe AI's reigning foundation flagship is also its newest deployed model.