Sonnet 4.5 hits 82% on SWE-bench with test-time compute. We're adding another dot to what's been a pretty clean exponential toward saturating SWE-bench as a benchmark. At this rate we'll need new evals soon, but it's still a useful signal that the capability curve is holding.
RESEARCH
-

Anthropic to release Claude Sonnet 4.5 soon
By
–

BREAKING : Anthropic is about to release Claude Sonnet 4.5 soon! SOTA on SWE bench
-

DeepSeek’s Sparse Attention Enables Massive Price Reduction Economics
By
–
>50% price cut across the board *MAKES DeepSeek money*
— swyx 🐣 (@swyx) 29 septembre 2025
put another way, DeepSeek's Sparse Attention allows DeepSeek to serve 1.8m OUTPUT tokens for the same price as it used to serve 128k output.*
here's my extremely scientific calculation (i also eyeballed it and did it on a… https://t.co/rDNwafshwQ pic.twitter.com/5FpN3V2gQM>50% price cut across the board *MAKES DeepSeek money* put another way, DeepSeek's Sparse Attention allows DeepSeek to serve 1.8m OUTPUT tokens for the same price as it used to serve 128k output.* here's my extremely scientific calculation (i also eyeballed it and did it on a
-
Hyperparameter Changes for Muon Optimization and Optimizer Comparison
By
–
What hyperparameters changed when optimizing for muon? @clashluke has tried several new optimizers on our code and not reliably beaten Adamw.
-
Study of differences in similar processes for ainews
By
–
(we do similar for ainews, just wanted to study the diffs, happy to clean it up as a service)
-
NVIDIA CEO Jensen Huang on AI Scaling Laws and Infrastructure
By
–
Scaling laws, AI infrastructure, and the future of technology.
— NVIDIA (@nvidia) 29 septembre 2025
NVIDIA CEO Jensen Huang breaks it down on the @BG2Pod with @altcap and @_clarktang.
Listen in 🎧: https://t.co/KNdqSdiatxScaling laws, AI infrastructure, and the future of technology. NVIDIA CEO Jensen Huang breaks it down on the @BG2Pod with @altcap and @_clarktang
. Listen in : -

Top robotics stories: DeepMind thinking robots Meta Android
By
–
Top stories in robotics today: – DeepMind’s robots learn to think aloud – Meta wants to be the ‘Android of robotics’
– Unitree’s Bluetooth backdoor nightmare – iRobot founder: humanoid hype is fantasy
– Quick hits on other robotics news Read more: https://
robotnews.therundown.ai/p/googles-robo
ts-learn-to-think-first
… -
Symbolic Verifiers vs LLM Judges for Verification Tasks
By
–
What do you mean with set of prompts? A LLM judge with a rubric? (That’s a different approach; covered in the appendix).
The reason for a symbolic verifier is that it is symbolic. There is no nondeterminism, bias, etc. Doesn’t work for everything, but eg for math there is no

