10xAI Wiki

12分钟阅读

AI Models Landscape 2025–2026

Overview

The 2025–2026 period marks a fundamental shift in AI model economics and capability. Chinese open-source models (Kimi K2, Qwen3, GLM-4.5/5, DeepSeek V4) have reached near-parity with top Western proprietary models on coding benchmarks while costing 10–100x less. The industry has moved from a US-dominated landscape to genuine global competition.

Key defining characteristics of this era:

  • Cost collapse: API prices dropped 10–100x compared to 2024
  • Context explosion: 1M token context windows became standard
  • Agentic shift: Models now act (write/run code, call APIs, browse) not just respond
  • Open-source competitiveness: Open models routinely challenge closed models on key benchmarks
  • Multimodal breakthrough: Image generation models (GPT-Image-2) solved Chinese text rendering, a long-standing bottleneck

Key Players

Western / Closed-Source

CompanyModel FamilyFlagship (2025-2026)Pricing (Input/Output per M tokens)
AnthropicClaudeClaude Opus 4.615 / 75
OpenAIGPTGPT-5.42.50 / 10
GoogleGeminiGemini 3.1 Pro / Gemini 32.00–4.00 / 12–18

Chinese / Open-Source Leaders

CompanyModelSWE-bench VerifiedContextPrice (Input/Output per M tokens)License
Alibaba (Qwen)Qwen3-Coder-480B67%256K–1M~$0.10–0.80Apache 2.0
Moonshot AI (Kimi)Kimi K265.8%128K0.15 / 2.50Modified MIT
Zhipu AI (GLM)GLM-5.1~70%+128–200K0.60 / 2.20MIT
DeepSeekDeepSeek V480.6% (SWE-Verified)1M (standard)~1/7 of Claude Opus 4.6Open
MiniMaxM2.756.22% (SWE-Pro)Long context~$0.03/MOpen
ByteDanceDoubao-Seed-2.069.16 SuperCLUELong context￥9.60/MProprietary
XiaomiMiMo-V2-ProTop tier1M1–2 / 3–6Proprietary

For context: Claude Sonnet 4.5 leads at 72.7% on SWE-bench Verified. Chinese models are closing the gap with GLM-5.1 achieving ~70%+, and DeepSeek V4 at 80.6%.

Benchmark Comparison (2025-2026 snapshot)

SWE-bench / SWE-Verified (Real-world software engineering)

ModelScoreOriginNotes
DeepSeek V480.6%ChinaMatches Claude Opus 4.6
Claude Sonnet 4.572.7%USAHistorical leader
Claude Opus 4.672.5%USA
Gemini 376.2%USALatest release
OpenAI o371.7%USA
GLM-5.1~70%China#3 global, #1 open-source
Qwen3 Coder67%ChinaOpen-source leader
Kimi K265.8%China
Doubao-Seed-2.069.16ChinaSuperCLUE score
MiniMax M2.756.22%ChinaSWE-Pro
Gemini 2.5 Pro63.8%USA
GPT-4.154.6%USA

Key insight: DeepSeek V4 now matches Claude Opus 4.6 on SWE-Verified (80.6% vs 80.8%), while Chinese models dominate the open-source rankings.

Architecture Comparison

All leading Chinese models use Mixture-of-Experts (MoE) — activating only a fraction of total parameters per inference:

ModelTotal ParamsActive ParamsContext
Kimi K21 trillion32B128K
Qwen3-Coder480B35B256K–1M
GLM-4.5355B32B128K
GLM-5744B40B200K
DeepSeek V4~671B~37B1M

1. Cost Collapse

API pricing has dropped dramatically:

  • 2023: GPT-4 at $30/M input tokens
  • 2025: Comparable quality available at $0.10–0.60/M input tokens
  • Prediction: Sub-$0.01/M tokens by 2026

Drivers: MoE efficiency, hardware optimization, competition from Chinese labs.

2. Long Context as Standard

  • 2024: 128K tokens was “long context” (Gemini had exclusive 1M)
  • 2025: DeepSeek V4 made 1M context standard across all tiers
  • Use case: Entire codebases, legal documents, book-length analysis in single prompt

3. Agentic Capabilities

Models shifted from text generators to autonomous agents:

  • Tool-calling: GLM-4.5 achieves 90.6% tool-calling success (highest among open models)
  • Coding agents: Models like Claude Code, Kimi K2, GLM-5 autonomously write/debug/deploy code
  • MCP standard: Model Context Protocol enables standardized tool integration

4. Reasoning Models

All major labs introduced “thinking” modes where models spend more compute reasoning before answering:

  • OpenAI: o-series (o3, o3-mini)
  • Anthropic: Extended thinking in Claude
  • DeepSeek: R1 series
  • Qwen/GLM: Hybrid thinking/non-thinking modes

5. Chinese Models Catching Up

In July 2025, three Chinese companies simultaneously released models that redefined price-performance:

  • Focused optimization on practical coding over general knowledge
  • Massive domestic developer base for training data
  • Open-source strategy to build ecosystem
  • Hardware constraints (no H100 access) drove efficiency innovations

6. Hardware Decoupling

DeepSeek V4’s migration from NVIDIA CUDA to Huawei Ascend chips — achieving 35x inference speed improvement — broke the assumption that frontier AI requires NVIDIA hardware. This is strategically significant for China’s AI independence.

Model Deep Dives

Qwen3 / Qwen3-Coder (Alibaba)

  • Architecture: 480B total / 35B active MoE, Apache 2.0
  • Strengths: Best-in-class context (256K–1M), 119 human languages, 358 programming languages, top SWE-bench score among open models (67%)
  • Use case: Large codebase refactoring, multilingual tasks, research
  • Tools: Qwen Code CLI (forked from Gemini Code), compatible with Claude Code, Cline
  • See: China-AI-Model-Landscape-2025-2026

Kimi K2 (Moonshot AI)

  • Architecture: 1T total / 32B active MoE, Modified MIT
  • Strengths: Native MCP support, agentic capabilities, lowest cost ($0.15/M input)
  • Performance: 65.8% SWE-bench, 97.4% MATH-500
  • Use case: Green-field development, autonomous coding agents
  • See: China-AI-Model-Landscape-2025-2026

GLM-4.5 / GLM-5.1 (Zhipu AI)

  • Architecture: 355B/32B active (4.5); 355B/32B active (5.1), MIT license
  • Strengths: Highest published tool-calling success (90.6%), runs on 8 H20 chips
  • GLM-5.1 milestone: First open-weight model to achieve 8-hour continuous autonomous work
  • SWE-bench: #3 globally, #1 open-source
  • Use case: Debugging, agentic workflows, long-horizon task automation
  • See: GLM-5

DeepSeek V4 (DeepSeek)

  • Context: 1M tokens standard
  • Coding: SWE-Verified 80.6% (near Claude Opus 4.6 at 80.8%); Codeforces 3206 Elo
  • Price: ~1/7 of Claude Opus 4.6
  • Notable: First dual NVIDIA CUDA + Huawei Ascend validation; Hybrid attention (CSA + HCA)
  • Key Innovation: Engram memory module (O(1) lookup), FP4 quantization
  • See: DeepSeek

Claude (Anthropic)

  • Current frontier: Opus 4.6 and Sonnet 4.5 lead SWE-bench at 72–73%
  • Coding dominance: Historically “断崖式领先” (cliff-edge leading) — Google/OpenAI engineers use it internally
  • Strategy: Deep enterprise focus, MCP protocol inventor, safety-first
  • See: Anthropic-Claude-Company

Gemini 3 (Google)

  • Context: 1M input / 64k output
  • Architecture: TPU-based, MoE
  • Key Innovation: Deep Think mode, multi-modal reasoning
  • Performance: 1501 Elo LMArena, 91.9% GPQA Diamond
  • Coding: 76.2% SWE-bench Verified (surpasses 2.5 Pro)
  • See: 2025-11-19-Gemini-3-Released

MiniMax M2.7

  • Architecture: MoE with self-evolution capability
  • Key Innovation: Model builds and optimizes its own Agent Harness
  • SWE-Pro: 56.22% (near Opus levels)
  • MLE Bench: 66.6% medal rate (tied with Gemini-3.1)
  • See: MiniMax

GPT-Image-2 (OpenAI)

  • Type: Text-to-image generation model
  • Release: April 20, 2026 (free to all ChatGPT users)
  • Key Innovation: First model to reliably generate high-quality Chinese text in images — solves the “Chinese typography problem” that plagued previous image gen models
  • Capabilities: Short-prompt driven (3-5 sentences), magazine covers, infographics, UI mockups, vintage print simulation, multi-platform social media screenshots
  • Significance: Represents a watershed moment for Chinese-language AI content creation
  • See: GPT-Image-2, ../sources/2026-05-07-GPT-Image-2 全量上线，中文顶到爆，50+ Case 生图实测

Doubao Seed 2.0 (ByteDance)

  • SuperCLUE: 69.16 score (#2 globally behind Claude Opus 4.6)
  • Math Reasoning: 87.83 (global #1, +10 over Claude)
  • Limitation: Code generation trails Claude (48.77 vs 59.91)
  • See: 2026-02-25-Doubao-Seed-2

Xiaomi MiMo V2-Pro

  • Architecture: >1T params, 42B active, Hybrid Attention (7:1 ratio)
  • Context: 1M tokens
  • Ranking: #8 global, #2 domestic (Artificial Analysis)
  • Key: Agent-native design, surpasses Claude Sonnet 4.6 in practice
  • See: Xiaomi

AI Cloud Economics

China Enterprise MaaS Market (IDC 2025)

ProviderMarket ShareParent
Volcano Engine49.5%ByteDance
Alibaba Cloud28%Alibaba
Baidu AI Cloud10%Baidu
  • 2025 total: 1,944 trillion Tokens on China public cloud MaaS (16x YoY growth)
  • Volcano Engine’s share grew from 46.4% (2024) to 49.5% (2025) despite intensifying competition
  • Key insight: MaaS is becoming the dominant cloud paradigm — selling Tokens, not servers

Token Economy Scaling

Token consumption has experienced three step-changes:

  1. ChatGPT era (2022.11): Text-only, thousands of tokens per request
  2. Reasoning models (O1/R1, 2025): Compute-heavy inference, 10x higher per-task consumption
  3. Long-horizon Agents (Claude Code, ArkClaw, 2025-2026): Continuous multi-hour execution, thousands of times more tokens

Each step-change directly multiplied Token consumption, with multimodal (image/video generation) adding further acceleration (single tasks consuming tens of millions of tokens).

China Token Supplier Rankings (Deben Consulting TOP100, 2026)

  1. Huawei Ascend (chips + infra)
  2. Alibaba (Qwen)
  3. ByteDance (Doubao + Volcano Engine)
  4. Tencent
  5. DeepSeek
  6. Zhipu AI (GLM)
  7. Baidu

Note: Rankings differ from IDC MaaS consumption data. Deben likely weights strategic/infrastructure value, while IDC measures actual Token consumption (where ByteDance/Volcano dominates at 49.5%).

AI Model API Economics

Cost comparison at enterprise scale (1B tokens/month)

ModelMonthly CostAnnual
Kimi K2$150$1,800
GLM-4.5$110$1,320
GPT-5.4$2,500$30,000
Claude Opus 4.6$15,000$180,000

The economic case for Chinese models is compelling for cost-sensitive applications.

Open-Source vs. Proprietary

Open-source advantages

  • Data sovereignty: Self-host, data never leaves premises
  • Cost: 3–30x cheaper than proprietary APIs
  • Customization: Fine-tune with LoRA or full training
  • Compliance: Meets GDPR, HIPAA, regulated industry requirements
  • Licences: MIT/Apache 2.0 allow commercial use

Proprietary advantages

  • Top performance: Still 5–8% ahead on key benchmarks (as of mid-2025)
  • Safety/alignment: More rigorous evaluation and constitutional AI
  • Multimodal: More mature vision/audio capabilities
  • Support: Enterprise SLAs

Predictions for 2026

  1. Sub-$0.01/M tokens for open models
  2. Specialized models for specific languages (Rust, Go) and frameworks
  3. Western response: OpenAI/Anthropic will either match pricing or pivot to premium features
  4. Hybrid strategies: Use Chinese models for bulk coding, Western models for sensitive/creative tasks

Sources:

[2026-07-17] State of GPT：2023-05 助手榜历史快照

  • Karpathy 在 Microsoft BUILD 2023（2023-05-23）公布的助手模型排行榜：GPT-4 (1274) > Claude-v1 (1224) > ChatGPT (1155) > Vicuna-13B (1083) > Koala-13B > RWKV-4-Raven-14B > ChatGLM-6B > Alpaca-13B (904) > LLaMA (826)。
  • “开源最好的 Vicuna 比 GPT-4 差 190 分”——与本页 2025-2026 数据对比（DeepSeek V4 80.6% 追平 Claude Opus 4.6 80.8%），可见两年内开源与闭源的差距闭合速度之快。
  • 四阶段训练流水线（预训练→SFT→RM→RLHF）与缩放定律（“数据比参数更重要""训练永不收敛”）是理解当前 2025-2026 格局的技术基础框架。
  • 实体见 Andrej-Karpathy。

来源：../sources/2026-07-17-个人成长与学习方法

关系图谱

反向链接