Good Morning!
Two new names climbed into the top ten while the board shrank by a fifth. MiMo v2.6 Flash went from fifteenth to ninth, Nemotron-3 Ultra from fourteenth to tenth.
MiMo is a code machine: almost all its traffic is code, and calls keep getting shorter, so more people doing smaller jobs. Nemotron is the opposite, a steady industrial load, same volume and mix daily, as if one company's batch job skipped the weekend.
GLM 5.3 and Claude Haiku 5.5 left the ten shedding far more than the board, so their users dislike weekend work even more than the market does.
Model | tokens · vs. previous day (Oct 30) | ||
1 | DeepSeek V4.1 Flash · DeepSeek | 8.76T | – -12% |
2 | GLM 5.3 Flash · Z.ai | 1.84T | – -31% |
3 | GPT-5.6 Luna · OpenAI | 1.11T | ▲2 -4% |
4 | Hy4 preview · Tencent | 0.96T | ▼1 -36% |
5 | MiniMax M3.1 Flash · MiniMax | 0.89T | ▲1 -15% |
6 | DeepSeek V4 Flash · DeepSeek | 0.84T | ▲2 -6% |
7 | GPT-6 Luna · OpenAI | 0.82T | ▼3 -32% |
8 | Claude Opus 5.5 · Anthropic | 0.62T | ▼1 -37% |
9 | MiMo-V2.6-Flash · Xiaomi | 0.45T | ▲6 +16% |
10 | Nemotron 3 Ultra · NVIDIA | 0.44T | ▲4 +5% |
Grok Watermarks Every Answer

TL;DR: Supported Grok API models will add invisible text watermarks by default from November 7, 2026, including Grok 4.7 and the Grok 4.20 variants. The policy covers ordinary text, structured output and tool-call arguments. Team admins can disable it per model where permitted, after confirming a legal exception; some models may remain locked on. Developers can inspect the x-watermark response header. Read more →
OpenAI Lists GPT-6 Astra for Law

TL;DR: OpenAI's official pricing page lists gpt-6-astra-law as a legal model configuration at $10 per million input tokens, $50 per million output tokens and $1 for cached input. Promotional pricing is available at least through December 31, 2026. The listing confirms published pricing, not general account access, a launch date or independently validated legal performance. Read more →
MI355X Improves Kimi Inference Value

TL;DR: SemiAnalysis reports that an optimized AMD MI355X stack achieved up to twice the performance per dollar of Nvidia B200 on its Kimi K3 agentic-inference benchmark. The cited configuration uses 2,304 GB of HBM per server, helping avoid the cross-server networking required by the compared B200 setup. Results depend on the workload, software configuration and modeled hardware costs. Read more →
GLM 5.5 Leak Hints at Post-Training

TL;DR: Lumina reports that GLM 5.5 is in post-training, citing a leaked development record. The account clarified that the 32B, 122B and 256B models mentioned are internal models being post-trained for merging into the next base model, rather than three announced releases. Read more →
Musk Maps Colossus 2 Expansion

TL;DR: Musk says Colossus 2's Macroharder and Minihard buildings each contain 220,000 GB300 GPUs. He targets a similarly sized Gun Range training cluster for late November and an approximately 200,000-GPU Spam inference cluster for December. These are self-reported counts and expansion targets; the cited post provides no independent verification of installed capacity or assurance that the planned sites will open on schedule. Read more →
GPT-4o Mini Ends January 14

TL;DR: OpenAI announced on October 14 that GPT-4o mini's API shutdown is scheduled for January 14, 2027, alongside GPT-4o, GPT-4.1 and GPT-4.1 mini. The published replacement is gpt-6-luna; fine-tuned versions follow their base model's schedule. Treat this as a migration reminder: inventory model IDs, test replacements on your own workloads and update production calls before the deadline. Read more →

Top AI models. Lower costs.
Access GPT, Claude, Gemini, and more through one OpenAI-compatible API—and save up to 70% on AI API costs. Create your DIT account, get your API key, and start building.