AI model leaderboard · LinkWord

OpenAI 新旗舰模型，三款中的最强版本。在 HealthBench Professional 得分 60.5%，网络安全和生物风险评估均为 High 级别。

GPT-5.6 系列性价比版本，能力和旗舰接近，价格减半。

Fugu Ultra is Sakana AI's flagship multi-agent orchestration model. Rather than a single monolithic model, it dynamically orchestrates a pool of expert models to tackle complex multi-step tasks. Benchmarks competitive with Fable 5 and Mythos Preview. 1M context window, text+image input.

Anthropic 1M 上下文推理模型，标准版 $5/$25，快速版 $30/$150

Google lightweight flagship model with built-in Computer Use, function calling, Search/Maps Grounding — ideal for agent scenarios

3B dense reasoning model. AIME26: 94.3, based on Qwen2.5. Uses Spectrum-to-Signal post-training. No tool calling support, focused on math and code reasoning.

1T MoE 编程专用模型，256K 上下文，Modified MIT 开源，推理 token 消耗降低 30%

xAI multi-agent reasoning model built on Grok 4.20, supports multi-agent collaborative orchestration with 2M context, ideal for complex task decomposition and parallel execution

xAI推理模型，2M上下文，最低幻觉率，支持Agent工具调用

First open-weights model combining frontier coding, 1M context, and native multimodality. MSA sparse attention architecture. SWE-Bench Pro 59.0%, TerminalBench 66.0%. Aggressively priced.

1.6 trillion parameter MoE model from Meituan (LongCat), ~48B activated per token, trained on domestic AI ASIC superpods, 1M context window, MIT license

GPT-5.4高效变体，400K上下文，优化高吞吐场景

NVIDIA Nemotron 3 Ultra is an open frontier-reasoning and orchestration model with 550B total params (55B active, MoE). Built on hybrid Mamba-Transformer architecture, it achieves 95% Long Context Ruler @1M and 65-70% SWE-bench Verified. Fully open weights, data, and training recipes.

Google 超低价 1M 上下文模型，适合大批量处理

StepFun's latest multimodal MoE model with 196B parameter language backbone and vision encoder for native image/video understanding

Poolside Laguna XS 2.1 is a 33B-A3B MoE coding agent model optimized for local deployment. Builds on XS.2 with improved SWE-bench Multilingual (63.1%) and stronger terminal-style task performance. Supports 256K context, runs locally in vLLM/SGLang.

Mistral Large 2025年12月更新版，262K上下文，定价大幅下降

Cohere North Mini Code is Cohere's first agentic coding model and the debut of its North family. A sparse MoE with 30B total / 3B active parameters, optimized for end-to-end coding tasks. 256K context window, text input only.

xAI's coding-focused model trained for agentic software engineering workflows, supports text+image input

IBM's open-source 8B dense enterprise model, matching 32B MoE performance, 131K context, designed for enterprise tasks