Good Morning!
Two names entered the top ten on a softer day. Kimi K3.1 Pro came from seventeenth to ninth, GPT-6.1 Sol from thirteenth to tenth.
They arrived by different roads. Kimi's calls grew fewer and much larger, so existing buyers are sending bigger code jobs rather than new ones arriving. Sol went the other way, with many more, smaller calls.
What fell was the human-facing traffic: GPT-5.6 Luna halved, and the general-chat crowd went home for the weekend.
Model | tokens · vs. previous day (Oct 31) | ||
1 | DeepSeek V4.1 Flash · DeepSeek | 8.14T | – -7% |
2 | GLM 5.3 Flash · Z.ai | 1.41T | – -23% |
3 | MiniMax M3.1 Flash · MiniMax | 0.90T | ▲2 +1% |
4 | DeepSeek V4 Flash · DeepSeek | 0.82T | ▲2 -3% |
5 | GPT-6 Luna · OpenAI | 0.81T | ▲2 -1% |
6 | Hy4 preview · Tencent | 0.73T | ▼2 -23% |
7 | Claude Opus 5.5 · Anthropic | 0.66T | ▲1 +5% |
8 | GPT-5.6 Luna · OpenAI | 0.57T | ▼5 -48% |
9 | Kimi K3.1 Pro · MoonshotAI | 0.57T | ▲8 +144% |
10 | GPT-6.1 Sol · OpenAI | 0.43T | ▲3 +14% |
GPT-6.2 Astra Appears in Codex Logs

TL;DR: Codex users are reporting gpt-6.2-astra in usage analytics and response metadata, even while the model picker still reads GPT-6 Astra. The sightings suggest OpenAI may be testing an Astra update behind the existing selection, with some accounts recording usage against the new model ID. Read more →
Mythos 5.5 Name Surfaces in Site Code

TL;DR: Anthropic's Mythos page briefly named Claude Mythos 5.5 in a hidden heading and image alt text before returning to 5.1. Chubby shared the source-code screenshot, and an archive comparison showed the 5.5 wording was absent on October 23. The visible page continued to describe Mythos 5.1, so the change points to a possible model refresh. Read more →
Claude Ultra Drops Five-Hour Cap

TL;DR: Verified Claude Ultra accounts no longer have a rolling five-hour usage limit, while the weekly allowance remains. The $200-a-month plan replaces Max 20x and includes 10 times Pro's weekly usage. Verification happens automatically based on account and payment history; until it completes, a temporary session limit applies. Chat and Claude Code draw from the same weekly budget. Read more →
OpenAI Limits Paid Accounts Per Device

TL;DR: OpenAI limits paid ChatGPT accounts per device and can block purchases when too many subscriptions share a card or device. The rules cover Go, Plus and Pro 100, 200 and 500; Free, Business, Enterprise and Edu are excluded. Signing out does not guarantee another slot, and buying credits or resetting usage does not remove the restriction. Read more →
MiniMax M3.1 Flash Goes Local

TL;DR: MiniMax highlighted community work bringing its 428B M3.1 Flash model to more local hardware. Unsloth's 2-bit GGUF is 143.9 GB, while separate recipes cover four RTX 3090s with system-RAM offloading, a DGX Spark, Apple Silicon and AMD Instinct. Quantization reduces weight memory and decoding traffic, with efficient kernels and expert placement making the savings practical. Read more →
xAI Targets 1.2M Blackwells by Year-End

TL;DR: Elon Musk says xAI is targeting around 1.2 million Nvidia Blackwell GPUs at Colossus 2 by year-end. He describes Macroharder and Minihard as 220,000-GB300 training clusters each, with Gun Range planned at the same scale and the roughly 200,000-GPU Spam building focused on inference. Training and serving workloads will use different network designs. Read more →
Retired A100 Cluster Listed for $59.7M

TL;DR: Core 4 Solutions offered a retired cluster on eBay for $59.7 million, comprising 300 racks, 9,600 Nvidia A100 GPUs and 4,200 AMD EPYC processors. Individual 32-GPU racks were listed at $199,000. The full-system listing has since ended. For buyers, the relevant cost includes the cluster's power, networking and operations, beyond its asking price. Read more →
Chinese Labs Reportedly Target GPT-6 Astra

TL;DR: The Information reports that Moonshot and DeepSeek used GPT-6 Astra outputs while training Kimi K3.1 Pro and V4.1-Flash 1017, respectively. The reported shift follows tighter Claude defenses: Anthropic says its measures reduced data harvested by distillers by around 90–99%. Extracting hidden reasoning remains a key target, as labs try to stop rivals using their outputs as training material. 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.