32 stories, 337 models ranked. Every one links to where it came from.
Tracking
Top story · Mistral AI · Oct 6
The brief · 31 more
01AnthropicOct 9
Give Claude a goal and it splits the work into tasks that run side by side, then puts the result together. Everyone on the waitlist is now in.
beta · Pro + Max · parallel threads
02AnthropicOct 9
For developers: one Claude agent writes a plan, hands pieces to many other agents, then merges their work. Open to try in beta.
public beta · multi-agent · API
03OpenAIOct 9
Codex can now guess what you'll ask it next and suggest the message for you, based on your chat so far. A beta for Pro users.
beta · Pro · suggests next message
04OpenAIOct 9
Codex on Windows now runs code inside a sealed-off box built by Microsoft, so it can't reach files or the network it shouldn't.
Windows 11 · MXC sandbox
05AlibabaOct 9
A faster version of Alibaba's free image model: it makes and edits pictures in just 8 steps and still outputs 2K images.
8 steps · 7B · 2K images · open weights
06CloudflareOct 9
A new Clef model that takes in audio and video as well as text and images. Cloudflare also cut Clef-flash's price and made Clef up to 2x faster.
audio + video in · cheaper Clef-flash · up to 2x faster
07AnthropicOct 9
The first report lists four ways Claude acted on real websites or systems in ways Anthropic didn't intend, like working around a block. All had little real-world impact.
4 behaviors · little real impact · new report series
08OpenAIOct 9
A dot is an always-on ChatGPT helper that keeps working on a goal between chats. Until today you could only create one on a computer.
iOS + Android · always-on agent
09xAIOct 9
Grok's assistant can now sign up for services, email businesses for you or book time with someone, using an inbox of its own.
own inbox · sign-ups + scheduling
10JetBrainsOct 8
A small, free coding model from JetBrains, the makers of IntelliJ, that runs on your own computer. Trained to work as a fast helper inside coding agents.
12B MoE · 2.5B active · Apache 2.0
11AnthropicOct 8
Ask Claude to turn your data into a live dashboard, or an idea into a short animated explainer. Both are new and in beta.
beta · dashboards · animation
12OpenAIOct 8
A faster way to run OpenAI's GPT-6.1 Sol: close to its top model, up to 8x quicker than standard Sol. In the API, Codex and ChatGPT Work.
up to 8x faster · API + Codex · ChatGPT Work
13AnthropicOct 7
Anthropic's smallest, fastest Claude got much cheaper: about 75% less to run than the last Haiku. Made for quick, high-volume jobs like summaries.
$0.10/M in · $0.50/M out · ~75% cheaper
14Liquid AIOct 7
Instead of writing an answer word by word, these models pick one in a single quick step. Built for fast decisions on small devices. Free to download.
3B + 600M · one pass · open weights
15OpenAIOct 7
ChatGPT now runs on GPT-6 and can answer with interactive pieces, not just text. Paid plans got it first; Free and Go started on Oct 8.
all paid plans · Free + Go from Oct 8 · interactive answers
16GoogleOct 7
A free Google tool that checks whether a picture, video or audio clip was made with AI from Google or partners like OpenAI and Nvidia.
free · image · video · audio · OpenAI + Nvidia
17PerplexityOct 7
Two free Perplexity models that help search find the right text, images and pages for a question: a big one (9B) and a small one (0.6B).
9B + 0.6B · text + images · Hugging Face
18Google DeepMindOct 6
Google's model for making and editing pictures by just describing what you want. It's in the Gemini app, AI Studio and Search's AI Mode.
image edit · Gemini 3.6 Flash · Gemini API
19GoogleOct 6
A tiny Google model that lets apps search photos, video, audio and text by meaning. Small enough to run on a phone, and free to use.
740M · on-device · 5 modalities · Apache 2.0
20PerplexityOct 6
Perplexity's model for quick choices, like which tool to use for a question, now costs half as much. Anyone can download it.
27B · $0.02/M in · half of v1
21CursorOct 6
Start a coding agent on your computer, then check in, reply or give it new work from the Cursor iPhone app while it keeps going.
iOS · remote agents
22ReflectionOct 5
A startup's first big open model, built for coding and agents. Sign up for early access now; the download comes later this month under Apache 2.0.
501B · 23B active · weights later in Oct
23AmazonOct 5
Amazon's voice model that you talk to in real time is now better at following instructions and using tools. Developers get it in Amazon Bedrock.
speech-to-speech · GA in Bedrock · 4 regions
24RekaOct 5
One model that can read, watch video and steer a robot, instead of several models passing work between them. An early research preview.
19B · video + actions · preview
25Aleph AlphaOct 3
A German open model that speaks English and German, for European companies that want to run AI on their own servers. Free to download.
78B · 3B active · 1M context · Apache 2.0
26Ai2Oct 2
Ai2's free model turns a research question and a pile of papers into a short report with citations, in under a minute on average.
8B · cited reports · 51s each
27CactusOct 2
Speech-to-text so small (17 MB) it runs offline on almost any phone or laptop, no internet needed. Understands seven languages.
16.9 MB · 7 languages · runs on CPU
28MicrosoftOct 1
Microsoft's model writes down what people say as they talk, in 60 languages. Microsoft says it's the most accurate on a public leaderboard.
60 languages · streaming · voice in 23
29CloudflareOct 1
Cloudflare's models choose from answers you give them instead of writing text, which is handy for sorting and routing. Free to download.
27B + 9B · 64K context · Apache 2.0
30DecagonOct 1
Decagon's AI phone agent for customer support now speaks with its own voice model. Most people in a blind test couldn't tell it was AI, Decagon says.
70+ languages · own speech model
31TavusOct 1
An AI that talks with you face to face on video and reacts in under half a second. Only a small group of testers can try it so far.
0.43s latency · video-to-video · preview
This week
Models
Top model
Epoch Capabilities Index
Claude Opus 5.5
167.3 · Anthropic
Models tracked
274 with independent scores
337
Best open weights
#15 overall
Kimi K3
Cheapest in the top 25
GPT-6 Luna · per 1M input tokens
$0.10
“The two biggest releases this week were about cost. A trillion-parameter model that only wakes 52 billion at a time, and a Haiku at ten cents.”
One page, checked against the source. No account, no inbox. Follow it in any reader.
