AUG 24, 2026 · MAPLEHILL LABS · WEEKLY · MARKET
Model market week of August 24, 2026
The market this week
New in the index (2):
- Devstral 2 2512 (Mistral) — from $0.440/MTok input
- Hy-MT2-7B (Tencent) — from $0.074/MTok input
Price moves — largest 8 of 21 observed moves:
- DeepSeek V4 Flash 0731 input: $0.064 → $0.040/MTok (-37.4%)
- Qwen3 235B A22B Instruct 2507 output: $0.550 → $0.350/MTok (-36.4%)
- GLM 5.2 output: $2.00 → $1.54/MTok (-23.0%)
- DeepSeek V4 Flash 0731 output: $0.100 → $0.080/MTok (-20.0%)
- Gemma 4 26B A4B input: $0.050 → $0.042/MTok (-16.0%)
- Qwen3.8 27B input: $0.400 → $0.350/MTok (-12.5%)
- Qwen3.6 35B A3B input: $0.080 → $0.070/MTok (-12.5%)
- Gemma 4 26B A4B output: $0.250 → $0.220/MTok (-12.0%)
Provider availability:
- Qwen2.5 7B Instruct: left Together
- Nemotron 3 Nano 30B A3B: now on Nebius
- Mistral Small 3.2 24B: now on Mistral
- Mistral Nemo: now on Mistral
Most-used models (OpenRouter tokens, trailing 7 days):
- DeepSeek V4 Flash 0731 — 11.6T tokens
- Ox Alpha — 11.6T tokens
- MiMo-V2.5 — 9.9T tokens
- Hy3 — 8.2T tokens
- DeepSeek V4 Flash 0423 — 5.5T tokens
Around the industry
- Reuters reported that DeepSeek raised API prices for its V4 models effective August 17 — increases ranging from 50% to over 10× depending on model, token type, and time of day. [1]
- OpenAI previewed an Ultrafast service tier for GPT-5.6 Sol, running up to 14× faster than standard processing. [2]
- OpenAI made GPT-5.6 Luna the default model for Free and Go ChatGPT users. [3]
- Mistral opened its platform to third-party open-weight models — starting with Z.ai's GLM-5.2 — alongside expanded in-region European inference. [4]
- Google DeepMind introduced Gemini 3.7 Flash, its coding-and-agents workhorse, with intro pricing through December. [5]
- Z.ai released GLM-5.2 Turbo on August 17. [6]
- Alibaba open-weighted its Qwen 3.8 generation: the 2.4T-parameter Qwen3.8-2.4T-A95B on August 12 under a revenue-gated license, and the dense Apache-2.0 Qwen3.8-27B on August 14. [7]
- Tencent released Hy-MT2-30B-A3B, a translation flagship covering 33 language pairs plus five Chinese dialect pairs, on August 20. [8]
Assembled automatically from ModelIndex's observed data and cited external sources, with light human curation of the industry section. Sourced claims link to their citations; our own figures link to the live pages they come from.