← All updates

Qwen4-Max Enters Preview

Plus: Google Ships Decision 1, 800M BotPsy Beats Qwen 2B

Good Morning!

OpenRouter routed 31.33T text-model tokens on Nov 11, up only 1.7% week over week.

Z.ai holds two slots. Claude Opus 5.5 fell 10% to 1.04T. Volume is now a contest among low-cost Chinese models: Qwen4-Max opened as a preview with no per-token price to compete on.

DIT
Model Ranking
WEDNESDAY
Top 10 text models by real-world token usage · Nov 11, 2026 (UTC)
Real-world use:
Agent
Code
General
Data
Model
tokens
WoW (11/4)
1
DeepSeek V4.1 Flash
· DeepSeek
12.02T
–
-9%
2
Hy4
· Tencent
2.71T
▲1
+110%
3
GLM 5.3 Flash
· Z.ai
2.61T
▼1
+36%
4
GPT-6 Luna
· OpenAI
1.25T
▲1
+10%
5
Claude Opus 5.5
· Anthropic
1.04T
▼1
-10%
6
DeepSeek V4 Flash
· DeepSeek
0.97T
–
+10%
7
GLM 5.3
· Z.ai
0.80T
▲5
+47%
8
MiMo-V2.6-Flash
· Xiaomi
0.67T
▲11
+140%
9
GPT-6.1 Sol
· OpenAI
0.65T
▲1
+10%
10
Nemotron 3.5 Super VL
· NVIDIA
0.53T
▼1
-15%
How we rank: tokens processed over the full 24 hours of Nov 11, 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.

Qwen4-Max Enters Preview

Image: xenospectrum

TL;DR: Alibaba Cloud opened Qwen4-Max Preview to Token Plan subscribers through Nov. 25 (UTC+8). The model takes text, images and video, with a 1 million token context window and up to 393,216 output tokens shared between reasoning and the final answer. The preview is subscription-only. Read more →


Gemini Decision 1 Costs $0.10/M

Google DeepMind signage in the reception area of its London office
Photo: Jose Sarmento Matos / Bloomberg via Getty Images

TL;DR: Google’s Gemini Decisions API scores up to 1,000 yes/no, multiple-choice or rubric questions against one multimodal input of up to 256K tokens. It charges $0.10 per million input tokens and bills no output tokens. Google reports 150–250 ms end-to-end latency in tests with inputs up to 16K tokens. Read more →


800M BotPsy Beats Qwen 2B

Image: tether.io

TL;DR: Tether released BotPsy, an open-weight 800 million parameter model tuned for tool calls. In Tether’s tests, it beat Qwen3.5-2B-Think on four agent benchmarks. Compared with its own Qwen3.5-0.8B-Think base, BotPsy rose from 23.4 to 44.0 on BFCL v4 and from 30.5 to 64.6 on ToolSandbox. Read more →


Carbon Ties for Second in Accounting

Photo: Luke Chesser, Unsplash

TL;DR: Gemini 4 Carbon scored 88.1% on DualEntry’s 101-task accounting benchmark, tying GLM-5.3 Flash and Grok 4.7 behind Gemini 4 Argon at 91.1%. Carbon led the hardest group at 89.3% and scored 66.7% on bank reconciliation, its weakest category. Read more →


AgiBot Releases RoboBeats Weights

Image: BigGo Finance

TL;DR: AgiBot released RoboBeats weights in 0.8B, 2B and 4B sizes on Nov. 3, alongside inference code and SSS-Bench. In its LIFE framework, RoboBeats watches a video stream and flags moments to act on or save. The memory system stores those frames and structured records for later use. Read more →


ERNIE Light Costs More Than Flash

Baidu co-founder and CEO Robin Li speaking at VivaTech in Paris in 2024
File photo: Chesnot / Getty Images

TL;DR: OpenRouter lists Baidu’s ERNIE Light DT 1.0 at $1.13 per million input tokens and $3.94 per million output tokens, compared with $0.28 and $1.13 for ERNIE 5.5 Flash Preview. Light DT has a 1,024,000-token context window and a 131,072-token output limit. Read more →


Reflection Tests Beam’s Safety

Image: Futurum

TL;DR: Reflection’s Nov. 5 report says Beam refused or redirected 88.9% of harmful requests and complied with 81.5% of harmless requests in its conversational safety tests. The evaluation covered 1,500 human-written conversations across 15 categories. Beam has 501.3 billion parameters, with 23.4 billion active. Read more →


Meta Details Its Triton Extension

Image: Engineering at Meta

TL;DR: Meta detailed TLX, its extension to Triton for controlling how GPU kernels schedule work and move data. Meta uses it in recommendation systems and language model training. It reports a 4x speedup for one recommendation kernel on an Nvidia H100 after several optimizations, including a new TLX implementation. 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.