CHINESE AI.

DeepSeek's flagship LLM released in April 2026. MoE architecture with 1.6T total parameters and 49B active, native 1M-token context, fully open-sourced under the MIT License. Scores 80.6% on SWE-bench Verified and 3206 on Codeforces, matching the world's top closed-source models.

Anthropic's Mythos-class flagship model released on June 9, 2026 — the first Mythos-class model available to the public. Leads substantially in coding, long-horizon agent tasks and complex reasoning. Supports Adaptive Thinking and can run complex task chains continuously for hours.

OpenAI's 2026 flagship model. 1.05M-token context, 130K max output, native multimodal (text/image/audio), leading multiple benchmarks.

ByteDance's flagship LLM released on February 14, 2026. Gold-medal level in the IMO/CMO math olympiads and ICPC programming contest, surpassing Gemini 3 Pro on the Putnam benchmark. Omni-modal, strong at scientific reasoning and agents, with over 100 million monthly active users.

Moonshot AI's 2.8T-parameter frontier flagship released in July 2026. 896-expert MoE with 16 experts active per token, and a flat-rate 1M context with no tiered surcharge. KDA hybrid attention delivers a 6.3x decode speedup; frontend coding ranks #1 globally, matching Claude Opus on coding & reasoning at half the cost. Open weights released July 27, 2026.

Z.ai (Zhipu)'s 2026 open flagship, MIT-licensed open weights on Hugging Face. Matches Claude Opus 4.8 at ~80% lower cost — overall spend ≈ 20% of Claude Opus. First-week API calls jumped 27x, praised by prominent engineers as matching US systems at a fraction of the cost.

MiniMax's efficient workhorse, consistently in the global top three by API call volume. Excellent price-to-performance for consumer-scale apps.

Released July 24, 2026. Near-Fable-5 performance at half the cost. 2x+ Frontier-Bench vs Opus 4.8. Default model on Claude Max. Supports effort ladder and Fast mode (2.5x speed).

Z.ai's Aug 14, 2026 flagship: 743B params on the same base as GLM-5.2, uplifted purely by post-training scaling (IndexShare, SAO and the open-sourced Slime RL framework). Coding +50% over GLM-5.2; open-source #1 on Terminal Bench 3.0 and Agents' Last Exam (CLI); agent performance on par with Kimi K3, Claude Opus 4.8 and Claude Fable 5. Live in ZCode, AutoClaw and GLM Coding Plan; full weights in two weeks, API coming soon.

OpenAI's new frontier model family released July 9, 2026, in three tiers: flagship Sol, workhorse Terra and lightweight Luna. Luna targets low-cost high-throughput with ~62% fewer factual errors than GPT-5.5 Instant. Permanent price cuts from July 30: Luna -80%, Terra -20%; Sol gains a Fast mode (2.5x speed at 2x price).

Tencent's new open-source flagship LLM released on August 28, 2026. MoE architecture with 770B total / 49B active parameters and a 1M-token context (960K max input, 64K max output), built for 'planning, executing and reliably delivering' agent capabilities. An open-source Preview build (not GA), live on Tencent Cloud TokenHub and OpenRouter.

Anthropic's point release on top of Fable 5, released in early September 2026. Retains the Mythos-class flagship capabilities — leading in coding, long-horizon agent tasks and complex reasoning, with Adaptive Thinking for hours-long task chains; 5.1 focuses on long-context stability, tool-call reliability and instruction following.

OpenAI's next-generation flagship model (codename Astra), released September 2026. First to reach the 'critical' cybersecurity capability tier in internal evals (ExploitBench perfect on known CVEs, autonomously found real zero-days). Native multimodal (text/image/audio); can solve hard math problems like high-dimensional sphere packing and group theory with Lean formal verification. Advanced safety-gated capabilities are review-user only.

Anthropic's September 2026 update to Opus 5, with 2x+ Frontier-Bench performance vs Opus 5. Supports new 'Adaptive Reasoning' mode for complex multi-step tasks. Default model on Claude Max.

DeepSeek's new-generation Flash model, officially released on September 10, 2026. It uses an entirely new model structure with native multimodal vision understanding and a 1M-token context, and drastically shrinks the KV cache (HBM demand down to 1/4 and SSD down to 1/8 versus the previous generation), cutting the cost of cached Agent workloads. Community tests measure 300–500 tokens/s. DeepSeek says it beats V4 Pro across performance, speed, cost and total latency, and it has taken over V4 Pro API traffic. The weights are open-sourced under the MIT License, enabling local deployment and fine-tuning.