Good Morning!
Two names climbed into the top ten while the board grew a tenth. GLM 5.3 took eighth from eleventh, Nemotron 3.5 Super VL ninth from twelfth.
They arrived by opposite mechanics. GLM 5.3 Flash gained tokens while its calls nearly halved, so the same users sent much heavier jobs. DeepSeek V4.1 Flash, still first, added a tenth more calls at unchanged job size.
The seats that emptied were call-heavy: the names that fell out lost requests faster than tokens, and the board now leans on fewer, heavier jobs.
Model | tokens · vs. previous day (Nov 2) | ||
1 | DeepSeek V4.1 Flash · DeepSeek | 13.24T | – +30% |
2 | GLM 5.3 Flash · Z.ai | 2.97T | – +15% |
3 | Hy4 preview · Tencent | 1.58T | – +0% |
4 | GPT-6 Luna · OpenAI | 1.13T | ▲1 +0% |
5 | Claude Opus 5.5 · Anthropic | 1.10T | ▲1 +12% |
6 | DeepSeek V4 Flash · DeepSeek | 0.90T | ▲1 -3% |
7 | MiniMax M3.1 Flash · MiniMax | 0.89T | ▲1 -2% |
8 | GLM 5.3 · Z.ai | 0.52T | ▲3 -2% |
9 | Nemotron 3.5 Super VL · NVIDIA | 0.52T | ▲3 -0% |
10 | GPT-6.1 Sol · OpenAI | 0.51T | – -8% |
Anthropic Threads Cut Time to First Token

TL;DR: Anthropic opened the Message Threads beta in its Messages API. After the first turn, agents send only new messages and the previous response ID. For Claude Code requests with 500,000 to 1 million tokens, Anthropic reports a roughly 35% drop in median time to first token. The full conversation still counts toward token usage. Read more →
Mistral Posts Large 4 Weights

TL;DR: Mistral has published the weights for Large 4, its 1-trillion-parameter mixture-of-experts model with 52 billion active parameters. The custom license permits deployment and fine-tuning. Companies that operate a model-as-a-service business need a separate commercial license before using the model or its derivatives commercially. Read more →
FlashProve Tests Provable Inference

TL;DR: Anthropic’s FlashProve prototype tests whether cryptographic proofs can verify that output came from a specified model. On a 4-bit Llama 4 Maverick, it generated proofs at about 110 tokens per second in batches on one H100 with CPU support. The benchmark used 70–95 times as much GPU time as ordinary vLLM inference. Read more →
Pruna Releases P-H3-Image Weights

TL;DR: Pruna adapted MiniMax’s H3 video model for image generation and editing and released the resulting weights. In Pruna’s tests, P-H3-Image generated a four-megapixel image in 2.6 seconds versus 12.4 seconds for the base model, using eight steps instead of 50. It supports Diffusers and ComfyUI. Read more →
Gemini 4 Carbon Edges Opus on One Index

TL;DR: Gemini 4 Carbon (High) scored 60 on the Artificial Analysis Intelligence Index, one point ahead of Claude Opus 5.5 in that comparison. Google announced Carbon on October 28 through Project Fairwind, an early-access program for cyber defenders. Read more →
Crouton Alpha Debuts With 1M Context

TL;DR: OpenRouter listed Crouton Alpha on November 3 as a free preview with a 1-million-token context window. The model comes from an unnamed third-party provider and is positioned for repeated agent subtasks such as tool use and document analysis. Read more →
Smaug Releases Two Audio Decision Models

TL;DR: Abacus.AI released Apache 2.0 weights for two Smaug models that score fixed-choice decisions such as whether a speaker has finished. The 4B multimodal Smaug-Decisions scored 60.1 on the company’s Earshot benchmark. Its smaller, audio-only sibling scored 53.0. Read more →
OpenAI Tests AI on Everyday Tasks

TL;DR: OpenAI published results from LifeVal, an evaluation of everyday tasks such as finding childcare, accessing government services and making a budget. It drew on time-use studies and a survey of 3,995 people. OpenAI says it used LifeVal internally for a year to track and improve model performance. Read more →
Fireworks Secures 5 MW of UK Inference Capacity

TL;DR: Fireworks AI signed a multiyear deal for 5 MW of liquid-cooled capacity at Era4’s Derbyshire data center. The site uses Nvidia B300 GPUs in an HPE modular data center powered by on-site landfill gas. The deal gives Fireworks UK-based infrastructure for serving open models. Read more →
OpenRouter Rewind: Chinese Models Topped 60%

TL;DR: A May 2026 report put Chinese models above 60% of weekly token volume on OpenRouter. Xiaomi, Alibaba, MiniMax, Zhipu, DeepSeek and StepFun accounted for roughly 60% combined in the report’s spring snapshot. The figure measures tokens routed through OpenRouter that week. 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.