AI Model Roundup: Aug 17-23, 2026
This week was about open weights and prices. Alibaba published the weights for its 2.4T-parameter Qwen3.8 flagship and a 27B model it says matches a model ten times its size. A small lab, Ornith AI, claimed Opus 4.8 parity with a 397B fine-tune. DeepSeek's peak-hour billing took effect, and OpenAI cut GPT-5.6 Sol's price by 20% to 33%.
Qwen3.8 weights go public
On August 17, Alibaba released the weights for Qwen3.8-2.4T-A95B: 2.4 trillion total parameters, 95 billion active. Alibaba's announcement describes the license as Apache 2.0; the Hugging Face repository metadata lists a custom license, so read the license file before you build on it.
The more practical release is Qwen3.8-27B, a dense, natively multimodal model with a 262K context window that extends to 1M. Alibaba says it matches Qwen3.7-Plus, an MoE model ten times its size, on agentic tasks, and that it runs on consumer hardware, including laptops with quantization. If that holds up, it is the most capable model most teams can run locally.
Ornith 1.5: parity claims from a fine-tune
On August 18, Ornith AI published three MIT-licensed models on Hugging Face, led by Ornith-1.5-397B, a 397B MoE with a 262K context window (about 1M with YaRN scaling). The model card says the Ornith line is built on Qwen3.5 and Gemma 4 with continued pretraining and post-training. That context matters: this is a strong fine-tune, not a new foundation model.
The self-reported comparison with Claude Opus 4.8:
| Benchmark | Ornith-1.5-397B | Claude Opus 4.8 |
|---|---|---|
| Terminal-Bench 2.1 (Terminus-2) | 86.1 | 85.0 |
| SWE-bench Verified | 86 | 85.8 |
| DeepSWE | 56.0 | 59.0 |
| GPQA Diamond | 92.8 | 93.6 |
Two wins, two losses, all within a few points. A new lab claiming parity with Opus 4.8 on its own harness needs independent reproduction before anyone routes production traffic to it. The smaller siblings, a 35B-A3B MoE and a 9B dense model, are the ones to try locally.
DeepSeek: official V4, peak pricing, and a vision model
- Peak pricing is live. DeepSeek V4 Pro reached general availability on August 13 (DeepSeek changelog), and the peak and off-peak billing announced in June took effect on August 17, Beijing time (TechNode). Rates double from 9:00 to 12:00 and 14:00 to 18:00 Beijing time. Background: DeepSeek Peak Pricing.
- V4 Flash Vision Exp. On August 21, DeepSeek V4 Flash Vision Exp appeared on OpenRouter at $0.22/$0.66 per million tokens: an experimental V4 Flash build with image input. It is labeled experimental; treat it that way.
OpenAI cuts GPT-5.6 Sol
On August 21, OpenAI cut GPT-5.6 Sol's API price from $5/$30 to $4/$20 per million tokens, a 20% input and 33% output reduction. The rate runs through at least November 21 and also applies to Codex credits and eligible ChatGPT Work plans. At $20 output, Sol undercuts Claude Opus 4.8 ($25) on output for the first time.
Takeaway
The open-weight frontier moved from API previews to downloadable weights this week, while the cheapest API is getting more expensive at peak hours and the most expensive one is getting cheaper. If you run agents on DeepSeek, move batch work to off-peak windows. If you're weighing Sol against Opus 4.8, recheck the numbers at the new price before November. And if you want a capable local model, Qwen3.8-27B is the one to test first.
Related: Best Open-Weight Models · Best Local Models · Best Cheap Models