Good Morning!
OpenRouter moved 29.95 trillion text tokens on Nov 13, up 2.1% week over week.
DeepSeek V4.1 Flash held the top spot at 11.36 trillion across 229 million requests, 72.7% of them machine traffic, and today it put its app's search, file and voice tools on the API.
Model | tokens | WoW (11/6) | |
1 | DeepSeek V4.1 Flash · DeepSeek | 11.36T | – +10% |
2 | Hy4 · Tencent | 2.62T | – -14% |
3 | GLM 5.3 Flash · Z.ai | 1.54T | – -10% |
4 | GPT-6 Luna · OpenAI | 1.26T | – +8% |
5 | Claude Opus 5.5 · Anthropic | 1.25T | – +19% |
6 | DeepSeek V4 Flash · DeepSeek | 0.74T | ▲1 +2% |
7 | MiMo-V2.6-Flash · Xiaomi | 0.60T | ▲9 +62% |
8 | GPT-6.1 Sol · OpenAI | 0.55T | ▲1 -19% |
9 | GLM 5.3 · Z.ai | 0.52T | ▲2 -16% |
10 | MiniMax M3.1 Flash · MiniMax | 0.52T | ▼2 -25% |
DeepSeek Opens App APIs

TL;DR: DeepSeek added web search, file reading and safety checks to the V4.1-Flash API, with separate endpoints for OCR, speech recognition and text-to-speech. The model has a 1M-token context and supports up to 384K output tokens. Off-peak rates are $0.15 per million uncached input tokens and $0.60 per million output tokens. Extended tools can incur separate fees. Read more →
Salesforce Opens SOV-Transcribe-1

TL;DR: Salesforce released SOV-Transcribe-1, a speech-to-text model for recorded audio, under Apache 2.0. It scored a 2.28% word error rate on Artificial Analysis’s non-streaming benchmark, placing first among open-weight models and seventh overall at launch. The model supports custom vocabulary for names and domain-specific terms. Read more →
Six Models Show Blackwell Traces

TL;DR: Researcher Will DePue found patterns in six open-weight models consistent with Blackwell GPUs generating their initial weights. The models come from DeepSeek, Moonshot AI, MiniMax and Ant Group. He reproduced preserved values from DeepSeek V4 Flash on a B300 and from Kimi K3 on a GB200 using the original random seeds. Read more →
GPT-6 Luna Adds Local Inference

TL;DR: OpenAI made in-country API inference available for GPT-6 Luna in India, Japan and Australia. Eligible organizations with active Zero Data Retention can use a regional project to process supported text and image requests locally through the Responses and Chat Completions APIs. Read more →
Ai2 Tests False Premises

TL;DR: Ai2 released XYBench, a benchmark of 8,115 questions built around misconceptions. It tests whether a model catches the mistaken premise and helps with the user’s underlying goal. The questions cover programming and everyday tasks such as cooking and home repair. Read more →
Edge0 Opens 145B and 120B

TL;DR: Edge0 released two models derived from Qwen3.8-Flash-Next, with Q3_K and NVFP4 weights for single-GPU inference on Linux and Windows. In Edge0’s test, the 145B Q3_K build decoded chat at 36.8 tokens per second on an RTX 5090 D capped at 16 GiB of VRAM and 20 GiB of system RAM. Read more →
Ant Opens Duo 1.2 Weights

TL;DR: Ant Group’s inclusionAI released weights and offline inference code for Duo, a research preview for two-way audio and video conversations that lets users interrupt while it responds. The release includes the Duo 1.2 base checkpoint and distilled and quantized versions of Duo 1.0. The reference offline setup uses two CUDA GPUs. 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.