Verified snapshot · July 30, 2026

The AI race has split in two

Closed models still hold the frontier — Claude Opus 5 leads Humanity's Last Exam at 56.3% — but Kimi K3 cut the open gap to 12.8 points, the narrowest on record, while open weights carry ~71% of production tokens. Capability and adoption are converging faster than either side planned.

Compare models View benchmarks

Best closed model

56.3%3 pts vs Jul 1

Claude Opus 5 · HLE, no tools — shipped Jul 24 at half Fable 5's price

Best open model

43.5%7.5 pts vs Jul 1

Kimi K3 · HLE — 2.8T open weights, the largest model released to date

Open–closed frontier gap

12.8 pts4.5 pts vs Jul 1

Narrowest reading on record — Kimi K3 closed it faster than Opus 5 reopened it

Open share of routed tokens

71%30 pts YoY ≈

Chinese-origin models alone are past 60% and now hold all five top OpenRouter slots

Frontier capability over time

Best available score on Humanity's Last Exam (no tools) — closed vs open weights. Interior points are approximate (≈); endpoints verified July 30, 2026.

Closed / proprietary

Open weights & source

July moved both lines. Kimi K3 lifted the open line 7.5 points on the 16th; Claude Opus 5 lifted the closed line 3.0 points on the 24th. Net, the gap fell from 17.3 to 12.8 — the first month where the open line gained ground while the frontier was also advancing, rather than only when it stalled or went dark.

Scale HLE via pricepertoken · Artificial Analysis · historical leaderboard snapshots

How far behind is open, benchmark by benchmark?

Best generally available closed vs best open score. Shorter connector = smaller gap.

Closed / proprietary

Open weights & source

Every gap narrowed this month. Kimi K3 sits 3.6 pts back on agentic coding and 12.8 on frontier reasoning — down from 14.4 and 17.3 three weeks ago — while GLM-5.2 trails by 8.1 on terminal agents. The gap still widens as tasks get more open-ended; it is just narrower everywhere.

llm-stats · tbench.ai · Artificial Analysis, Jul 2026

The price of the frontier

SWE-bench Verified score vs published output price per 1M tokens (log scale). Models without a published price or score are excluded, not estimated.

Closed / proprietary

Open weights & source

Kimi K3 is the point that changed this chart: 93.4% on SWE-bench Verified at $15 per 1M output tokens — within four points of the frontier at 60% of Opus 5's output price and 30% of Fable 5's. The old shape (Anthropic alone above 88%, open weights only below it) no longer holds. Further down, DeepSeek V4 Pro still resolves 80.6% at $0.87, so the cheap end hasn't moved — the middle filled in.

Provider list prices · llm-stats / BenchLM scores, Jul 2026

OpenRouter: open vs closed token share

Share of routed tokens. Endpoints reported; interior points approximate (≈).

Closed / proprietary

Open weights & source

Open weights crossed 50% around Q1 2026 and now run ~71% of routed tokens. Price did the work — the flip happened while open models still trailed on every leaderboard. As of late July, Chinese-developed models hold all five top slots by volume for the first time.

OpenRouter data via press reports, Jul 2026

Hugging Face downloads by lab origin

Share of open-model downloads by publishing lab's region. Endpoints from Hugging Face reporting; interior points approximate (≈).

China

Europe

Other

Hugging Face reports Chinese labs at 41% of past-year downloads and over 45% on recently uploaded models — the overtake already happened on new releases. Qwen alone passed 1B cumulative downloads in March.

Hugging Face reports, Jul 2026

The last 30 days

Every chart above moved because of these eight events.

Jun 30

Anthropic ships Claude Sonnet 5

80.4% on Terminal-Bench 2.1 — past Opus 4.8 — at $2/$10 introductory pricing; default for Free and Pro.

Jul 7

Beijing weighs export controls on Chinese models

Reuters reports meetings with Alibaba, ByteDance, and Z.ai on restricting overseas access — open-weight releases included.

Jul 8

xAI ships Grok 4.5

#1 on agentic tool use and #4 overall on the AA index; 83.3% Terminal-Bench 2.1 at $2/$6 and a 500K window.

Jul 9

GPT-5.6 Sol, Terra, and Luna go GA

Sol takes Terminal-Bench 2.1 SOTA at 88.8%; its HLE (47.2%) and SWE-bench (96.2%) scores land later in the month.

Jul 16

Moonshot releases Kimi K3

2.8T open weights, 1M context — 43.5% HLE and 93.4% SWE-bench Verified, and first place on Frontend Code Arena.

Jul 21

Google ships Gemini 3.6 Flash

Scores 50 on the AA index — above Google's own 3.1 Pro Preview — at $1.50/$7.50 using ~17% fewer output tokens.

Jul 24

Claude Opus 5 ships; open-weight coalition letter lands

Opus 5 takes HLE (56.3%) and Terminal-Bench 2.1 (89.1%) at Opus 4.8's price. The same day, 25 companies led by Nvidia and Microsoft publish an open-weight advocacy letter.

Jul 29

Chinese models sweep the OpenRouter top five

A first for the router, led by Xiaomi's MiMo-V2.5 — capability leadership and volume leadership are now on different continents.

Closed-model release

Open-model release

Policy

The story in four beats

The open gap shrank by a quarter in eight days

Kimi K3 shipped July 16 with 2.8 trillion open weights and took the open HLE record from ≈36% to 43.5%, plus 93.4% on SWE-bench Verified. Opus 5 answered on July 24 with 56.3%. Net: the gap fell from 17.3 points to 12.8 — the narrowest on record, and it narrowed while the frontier was still moving.

Frontier capability got cheaper, not dearer

Claude Opus 5 beats Fable 5 on every headline benchmark at $5/$25 — half the price. GPT-5.6 Terra matches GPT-5.5 at half. Gemini 3.6 Flash beats Google's own Pro preview on the AA index while cutting its output price. Three labs, one month, the same move.

Chinese models swept the router

For the first time, Chinese-developed models hold all five top OpenRouter slots by volume, led by Xiaomi's MiMo-V2.5 — a 15B-active model serving at roughly $0.22 per 1M output tokens. Chinese-origin models now carry over 60% of routed traffic on a router doing more than 20 trillion tokens a week.

Both governments reached for the same lever

Washington took Fable 5 offline for 19 days in June; Beijing met with Alibaba, ByteDance, and Z.ai on July 7 about restricting overseas access to Chinese frontier models — open weights included. A 25-company Nvidia/Microsoft coalition pushed back publicly on July 24. Openness is now a policy variable on both sides of the Pacific.

Go deeper

Model roster

Claude Opus 5, GPT-5.6 Sol, Kimi K3, and the rest — context, pricing, modalities, and license nuance.

Explore

Benchmarks

Humanity's Last Exam, agentic coding, and terminal agents — plus which benchmarks are already saturated.

Explore

Ecosystem

Where tokens get routed on OpenRouter and which labs the world is downloading on Hugging Face.

Explore