TL;DR

Data snapshot: July 10 - Aug 6, 2026, cross-checked across boards. The overall #1 is Anthropic's Claude Fable 5 (released Jun 9, 2026; AA Intelligence Index 64.9, GDPval-AA Elo 1932), and the LMArena text top 5 is all Claude. Coding #1 is Kimi K3 (LMArena Coding Arena 1529). By real usage, DeepSeek V4 Flash is now the world's most-called model (~84% of one day's OpenRouter tokens). Chinese models have narrowed the gap to the frontier tier to within ~5 points (SuperCLUE May 2026).

LMArena Text Arena (2026-07-10)

RankModelScore (Elo)Notes
1Claude Fable 51509±9Released Jun 9, 2026; API $10/$50 per M tokens
2–5Claude family (Opus 4.8 / Opus 4.7 etc.)All top 5 are Claude models; strong human-preference lead
China #1 (Jan 2026 snapshot)ERNIE-5.0-0110 (Baidu)1460Global #8, China #1; math #2 globally (2026-01-15)

Coding & General Intelligence Boards (2026-07/08)

Board (data date)LeaderScore / Value
LMArena Coding Arena (2026-07-16)Kimi K3 (Moonshot)Elo 1529
Artificial Analysis Intelligence Index (2026-07)Claude Fable 564.9, ~5 pts ahead of GPT-5.5; Opus 4.8 #2
AA GDPval-AA (real-world tasks, 2026-07)Claude Fable 5Elo 1932 (previous record: Opus 4.8)
SWE-bench Verified (official, Jun 2026)Claude Fable 595.0% (Opus 4.8: 88.6%)
OpenRouter call volume (2026-08-06)DeepSeek V4 Flash#1 globally; ~6.3T of ~7.5T daily tokens (~84%)

Change vs. Earlier Snapshot (Jan 2026 → Aug 2026)

ChangeJan 2026 snapshotAug 2026 snapshot
LMArena text #1GLM-4.7; ERNIE-5.0 in top 10Claude Fable 5; top 5 all Claude
Chinese models on LMArena textERNIE-5.0-0110 global #8No Chinese model in top 10
Coding boardClaude/GPT ledKimi K3 (1529) #1
Call volume #1DeepSeek V4 Flash (price-war payoff)
Chinese general (SuperCLUE May 2026)Gemini 3.1 Pro #1; DeepSeek-V4-Pro / Qwen3.7-Max / Doubao Seed-2.0-pro within <5 pts of frontier

How to Choose (Aug 2026)

Use caseFirst pickAlternativeWhy
General intelligence / hard reasoningClaude Fable 5GPT-5.5 / Opus 4.8#1 on AA index; dominant on long-horizon agent tasks (GDPval-AA 1932)
Coding / software engineeringKimi K3 / Claude familyDeepSeek V4Coding arena #1; Fable 5 hits 95.0% on SWE-bench Verified
Chinese-language tasksGemini 3.1 ProDeepSeek V4 / Qwen3.7-MaxSuperCLUE Chinese #1; domestic gap <5 pts
Cost / high-volume callsDeepSeek V4 FlashV4 ProOutput ¥2/M tokens (Aug 6, 2026); #1 in real call volume
Self-hosted / open weightsDeepSeek V4 / Qwen3.7-MaxKimi K3Open weights, deployable on-prem, data stays in-country

Takeaways

FAQ

Which AI model is the strongest overall in 2026?

Per the July 2026 Artificial Analysis Intelligence Index, the overall #1 is Anthropic's Claude Fable 5 (64.9, ~5 pts ahead of GPT-5.5); it is also #1 on the LMArena text board (Elo 1509±9). For Chinese-language tasks, SuperCLUE's #1 is Gemini 3.1 Pro.

Which model is best at coding?

LMArena Coding Arena (2026-07-16) #1 is Kimi K3 (Elo 1529); on official benchmarks, Claude Fable 5 reaches 95.0% on SWE-bench Verified and 80.0% on SWE-bench Pro, best for long-horizon code engineering. Pick by task: algorithm problems → LMArena Coding; real-repo bug fixing → SWE-bench.

How good are Chinese models right now?

They are at the frontier's edge: Kimi K3 tops the coding board, DeepSeek V4 Flash tops real call volume, and SuperCLUE (May 2026) puts DeepSeek-V4-Pro, Qwen3.7-Max and Doubao Seed-2.0-pro within ~5 points of the global frontier. But no Chinese model currently ranks in the LMArena text top 10.

Why is DeepSeek V4 Flash the world's most-called model?

Cost-performance: as of Aug 6, 2026, output is only ¥2/M tokens (¥4 peak) — tens of times cheaper than Claude/GPT flagships — with near-frontier quality. On Aug 5, V4 Flash accounted for ~6.3T of ~7.5T daily OpenRouter tokens (~84%). Note DeepSeek announced a price increase on Aug 6; the new rates are pending.

Where should I check AI model rankings?

Five commonly used sources: LMArena (human-preference Elo), Artificial Analysis (benchmark index), SuperCLUE (Chinese-specific), OpenRouter (real call volume), SWE-bench (code engineering). The same model can rank differently across boards, so always cite "board + data date", e.g. "LMArena text, 2026-07-10".

Sources

最后更新:2026-08-09