← All updates

DeepSeek Owns The Router

Plus: DeepSeek Takes 24.66%; Zhipu Legal Model; Darwin-180B Leads Open; TaoMate-HD Streams Talking; GLM-5.5 Leaks, St...

Good Morning!

DeepSeek V4.1 Flash ended the week with 24.66% of OpenRouter's routed tokens and the top share of a 29.3T-token day, down 4.8% from Thursday. Tencent's Hy4 more than doubled to 3.05T and moved into second.

The rest of today's issue is what that share buys: a legal-specialist GLM reaching public APIs, the open-weight reasoning leaderboard quietly changing hands, and a streaming avatar model cutting its own generation time.

DIT
Model Ranking
DAILY
Top 10 text models by real-world token usage · Nov 6, 2026 (UTC)
Real-world use:
Agent
Code
General
Data
Model
tokens · vs. previous day (Nov 4)
1
DeepSeek V4.1 Flash
· DeepSeek
10.37T
–
-21%
2
Hy4
· Tencent
3.05T
▲1
+135%
3
GLM 5.3 Flash
· Z.ai
1.71T
▼1
-11%
4
GPT-6 Luna
· OpenAI
1.17T
▲1
+2%
5
Claude Opus 5.5
· Anthropic
1.05T
▼1
-9%
6
crouton-alpha
· stealth
1.05T
▲5
+84%
7
DeepSeek V4 Flash
· DeepSeek
0.73T
▼1
-17%
8
MiniMax M3.1 Flash
· MiniMax
0.69T
▼1
-8%
9
GPT-6.1 Sol
· OpenAI
0.67T
▲1
+13%
10
Nemotron 3.5 Super VL
· NVIDIA
0.66T
▼1
+6%
How we rank: tokens processed over the full 24 hours of Nov 6, 2026 (00:00–24:00 UTC), counted and classified by DIT Radar's internal model.
Excluded: StepFun Step 5 Preview, whose traffic showed an anomalous, non-organic pattern.
Source: OpenRouter. The ranking reflects DIT Radar's interpretation of the data.

DeepSeek Takes 24.66% Of Routing

Photo: Domaintechnik, Unsplash

TL;DR: DeepSeek V4.1 Flash now holds 24.66% of routed tokens, up from 0.41% three weeks earlier, and ranks first globally after climbing from 34th. Listed price is $0.15 per million input and $0.60 per million output, with cached input at $0.003. The model carries 1M tokens of context and open weights. Read more →


Photo: Nasser Eledroos, Unsplash

TL;DR: Zhipu's GLM-5.3-FlashX-Legal entered OpenRouter and Vercel AI Gateway on Nov 2. On Harvey's law-firm benchmark it completes 19.2% of tasks at about $0.15 each, against 19.6% for Gemini 4 Argon at $10.55. Access is still gated per Z.ai's documentation and the headline figure is vendor-cited with no independent reproduction. Read more →


Darwin-180B Leads Open Reasoning

Image: DEV Community

TL;DR: Darwin-180B-RSI-R3 scored 73.68 in mathematical and scientific reasoning on SuperCLUE's November board, second overall behind GPT-6.1-Sol at 76.32 and first among open-weight models. It keeps the expert layers of Qwen3.8-Flash-Next but rebuilds attention, lifting the same test from the parent's 57.89. Weights ship under a commercial-use licence. Read more →


TaoMate-HD Streams Talking Avatars

Image: llms.blog

TL;DR: Alibaba's TaoLive team released a preview LoRA that generates speech and avatar video together on MiniMax H3, outputting 768x1376 at 24fps with 32kHz stereo. A coarse-to-fine plan cut generation from 491 to 307 seconds, and pipelined stages brought it to 160. Character error rate fell from 30.75% to 19.57%. Read more →


GLM-5.5 Leaks, Stays Unreleased

Photo: Taylor Vick, Unsplash

TL;DR: Details of Z.ai's next flagship, reportedly GLM-5.5, leaked through post-training trackers and a Tsinghua University lecture by co-founder Jie Tang describing an in-house pretraining data pipeline spanning web pages, documents, books, code, video and audio. Internal checkpoint sizes of 32B, 122B and 256B parameters and a rumoured system above a trillion parameters are circulating unconfirmed. The model is not yet released: as of early November 2026 Z.ai has published no model card, weights, API or price, and its shipping flagship remains GLM-5.3. Read more →


Self-Reported Rows Top SWE-bench

Image: DEV Community

TL;DR: On November 5, 2026 the Hugging Face SWE-bench Verified board carried 78 rows from 68 repos, and nine of them sat above 79.2, the highest entry in the official swebench.com Verified table, which was last updated February 26, 2026. All nine are self-reports: 51 rows come from still-open pull requests, 27 sit on a repo's main branch and none is marked verified. Six repos fill 16 of the 78 rows, and one model's own card spans 6.74 points across three harnesses. 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.