yellow3 Model Intelligence · Monthly Report

The Model Adoption Report

July 2026

Regional power shifts, model momentum, provider intelligence, the economics of the top tier, Europe Watch, and the signals to monitor.

106 models tracked · live routed traffic · published 12 Jul 2026
01

Executive summary

Five conclusions, three minutes. What changed across model routing, economics and the European picture in July 2026.

  1. Asia holds the lead at 60.6% of routed tokens (+6.3 pp over four weeks) - the concentration of demand by region of origin is not easing.
  2. Z.ai GLM 5 was the month's biggest climber, up 15 places to #29.
  3. No tracked price cuts landed inside the window this month.
  4. 3 new model(s) entered the ranking, led by Tencent Hy3 (free).
  5. Europe remains marginal at 0.51% of routed traffic; the regulatory picture is tracked in Europe Watch below.
02

Global AI market

Regional share of routed tokens

RegionShare4-week change
Asia60.60%▲ +6.26 pp
US32.67%▲ +0.23 pp
Other6.22%▼ -6.52 pp
Europe0.51%– +0.03 pp

Region reflects where each model's developer is headquartered. Four-week change is computed over ranked models.

Top gaining models

  • Z.ai GLM 5 #29 +15 places
  • Anthropic Claude 5 Fable #24 +8 places
  • OpenAI GPT 5.5 #14 +7 places
  • Google Gemma 4 31b It #27 +7 places
  • Anthropic Claude Sonnet 5 #13 +6 places

Top losing models

  • Tencent Hy3 Preview #21 -18 places
  • DeepSeek V3.2 #23 -13 places
  • Google Gemini 3.5 Flash #30 -10 places
  • OpenAI GPT 4o Mini #28 -5 places
  • Anthropic Claude 4.6 Sonnet #11 -4 places
Observed

Demand stayed concentrated in Asia-built models. The month's movement was driven by Z.ai GLM 5 on the way up and price competition at the low end.

03

Model rankings - top 25

#Model / originRouted share4-weekReasonMomentum
1Tencent Hy3 (free) Asia11.66%New entrantFlat
2Xiaomi Mimo V2.5 Asia11.30%▲ 2HoldingRising
3DeepSeek V4 Flash Asia9.93%▼ 2HoldingFalling
4MiniMax M3 Asia8.09%▼ 2HoldingFalling
5Z.ai GLM 5.2 Asia6.07%▲ 5New entrantRising
6DeepSeek V4 Pro Asia5.37%▲ 2HoldingRising
7Anthropic Claude 4.7 Opus US4.29%▼ 1HoldingFalling
8NVIDIA Nemotron 3 Ultra 550b A55b (free) US3.88%▲ 4Adoption climbingRising
9Anthropic Claude 4.8 Opus US3.86%HoldingFlat
10StepFun Step 3.7 Flash Asia2.31%▲ 3Adoption climbingRising
11Anthropic Claude 4.6 Sonnet US1.89%▼ 4Losing shareFalling
12Google Gemini 3 Flash Preview US1.78%▼ 1HoldingFalling
13Anthropic Claude Sonnet 5 US1.75%▲ 6New entrantRising
14OpenAI GPT 5.5 US1.66%▲ 7Adoption climbingRising
15Google Gemini 2.5 Flash US1.16%HoldingFlat
16Xiaomi Mimo V2.5 Pro Asia1.15%▲ 2HoldingRising
17Google Gemini 2.5 Flash Lite US1.07%▼ 3Losing shareFalling
18Poolside Laguna M.1 (free) US1.03%▼ 2HoldingFalling
19Google Gemini 3.1 Flash Lite US0.96%▲ 6Adoption climbingRising
20OpenAI GPT OSS 120b US0.96%▲ 2HoldingRising
21Tencent Hy3 Preview Asia0.78%▼ 18Losing shareFalling
22Moonshot Kimi K2.6 Asia0.74%▲ 2HoldingRising
23DeepSeek V3.2 Asia0.72%▼ 13Losing shareFalling
24Anthropic Claude 5 Fable US0.71%▲ 8Adoption climbingRising
25NVIDIA Nemotron 3 Super 120b A12b (free) US0.69%▲ 2HoldingRising

Reason is derived from the data (new entry, observed price change, or rank movement); it is not an editorial claim.

04

Provider intelligence

DeepSeekAsia3 ranked · 16.02% share

Best move: DeepSeek V4 Pro +2 to #6 · Cheapest: DeepSeek V4 Flash at $0.090/1M in

AnthropicUS5 ranked · 12.50% share

Best move: Anthropic Claude 5 Fable +8 to #24 · New: Anthropic Claude Sonnet 5 · Cheapest: Anthropic Claude Sonnet 5 at $2.00/1M in

XiaomiAsia2 ranked · 12.45% share

Best move: Xiaomi Mimo V2.5 +2 to #2 · Cheapest: Xiaomi Mimo V2.5 at $0.10/1M in

TencentAsia2 ranked · 12.44% share

New: Tencent Hy3 (free) · Cheapest: Tencent Hy3 (free) at $0.00/1M in

MiniMaxAsia1 ranked · 8.09% share

Cheapest: MiniMax M3 at $0.30/1M in

GoogleUS7 ranked · 6.81% share

Best move: Google Gemma 4 31b It +7 to #27 · Cheapest: Google Gemma 4 26b A4b It at $0.00/1M in

Z.aiAsia2 ranked · 6.64% share

Best move: Z.ai GLM 5 +15 to #29 · New: Z.ai GLM 5.2 · Cheapest: Z.ai GLM 5 at $0.60/1M in

NVIDIAUS2 ranked · 4.57% share

Best move: NVIDIA Nemotron 3 Ultra 550b A55b (free) +4 to #8 · Cheapest: NVIDIA Nemotron 3 Ultra 550b A55b (free) at $0.00/1M in

OpenAIUS3 ranked · 3.20% share

Best move: OpenAI GPT 5.5 +7 to #14 · Cheapest: OpenAI GPT OSS 120b at $0.030/1M in

StepFunAsia1 ranked · 2.31% share

Best move: StepFun Step 3.7 Flash +3 to #10 · Cheapest: StepFun Step 3.7 Flash at $0.20/1M in

PoolsideUS1 ranked · 1.03% share

Cheapest: Poolside Laguna M.1 (free) at $0.00/1M in

MoonshotAsia1 ranked · 0.74% share

Best move: Moonshot Kimi K2.6 +2 to #22 · Cheapest: Moonshot Kimi K2.6 at $0.66/1M in

05

Economics

Best value (cost + adoption)

  • Tencent Hy3 (free) · $0.00
  • NVIDIA Nemotron 3 Ultra 550b A55b (free) · $0.00
  • NVIDIA Nemotron 3 Super 120b A12b (free) · $0.00
  • Google Gemma 4 26b A4b It · $0.00
  • Google Gemma 4 31b It · $0.00

Monthly cost of a standard coding-agent workload; adoption flat-or-rising.

Most expensive, same workload

  • Anthropic Claude 5 Fable · $870.00
  • OpenAI GPT 5.5 · $485.00
  • Anthropic Claude 4.7 Opus · $435.00
  • Anthropic Claude 4.8 Opus · $435.00
  • Anthropic Claude 4.6 Sonnet · $261.00
Observed

The cheapest tracked model runs the standard coding-agent workload for $0.00/month; the most expensive costs $870 - a 87000x spread for comparable work. This month no tracked model cut list price.

Cost per coding task, per agent and per reasoning workload are on each model page's calculator.

06

Enterprise Watch

  • Availability - median provider uptime across tracked models: not available (last 30 days, real).
  • Open weights - 15 of the top 25 ship open weights; the rest are proprietary.
  • Longest context - OpenAI GPT 5.5 (1050K), Xiaomi Mimo V2.5 (1049K), DeepSeek V4 Flash (1049K).
  • Not yet tracked - SLAs, EU data-residency, latency and formal compliance are not in the routed-traffic feed; yellow3 does not estimate them. Flagged for the enterprise data set.
07

Europe Watch

  • Routed share - European-built models hold 0.51% of routed traffic.
  • EU AI Act - Edition 02 ยท 13 July 2026: The Digital Omnibus became final: the Parliament adopted it on 16 June, the Council on 29 June, locking in the slide of high-risk obligations to 2027 and 2028. Only Official Journal publication remains. Fines issued stayed at zero.
  • Digital Product Passport - public call record 0-1; First call graded: the EU missed its own service-provider deadline. The registry's 19 July test is next.
  • Sovereignty - Mistral remains the only European developer in the routed top tier.

Europe Watch draws on yellow3's live EU AI Act and Digital Product Passport instruments.

08

Emerging models

  • New this month - Tencent Hy3 (free), Z.ai GLM 5.2, Anthropic Claude Sonnet 5.
  • Open-source momentum - 15 of the top 25 are open-weight, a structural shift toward models teams can self-host.
  • Challengers to watch - the fastest climbers were Z.ai GLM 5, Anthropic Claude 5 Fable, OpenAI GPT 5.5.
09

Forecast

Where yellow3 believes the next month goes. This is analysis, not data - presented for your sign-off before it carries the yellow3 name.

yellow3 view · for sign-off

The observed signals point one way: Asia share +6.3 pp over four weeks; Z.ai GLM 5 climbing fast. On that basis, the low-cost, high-context challengers keep taking routed share while premium frontier models hold value in reasoning-heavy and enterprise use. The forward call - and the confidence level - is to be finalised by the yellow3 analyst before publication.

10

Appendix

Methodology

Adoption is measured from OpenRouter routed-token traffic over trailing windows; region reflects developer headquarters. Pricing and capabilities are pulled from the OpenRouter public model feed and provider uptime. This is developer routing behaviour, not the whole market.

Tracked providers

Anthropic, DeepSeek, Google, MiniMax, Moonshot, NVIDIA, OpenAI, Poolside, StepFun, Tencent, Xiaomi, Z.ai.

Definitions

Routed share = a model's tokens as a percentage of all routed tokens in the window. Standard workload = 1,000 coding-agent tasks/month (100K input, 10K output, 70% cached).

Data

Underlying figures: model-adoption-data.json. CSV exports are a Professional feature.

Media and press may use the data and graphics with clear attribution to yellow3.io. Full historical data, CSV exports and the archive are Professional features.