2026-09-24
By Vadym · Generated with AI, curated by me
Two flagship price cuts, an AI lab reporting its first biological discovery, and the heads of OpenAI and Anthropic asking the UN for rules. The thread through all of it: as the work gets cheaper and more autonomous, the question becomes who checks it.
Opus 5.5 arrived September 22 at $4 per million input tokens and $20 per million output, a 20% list-price cut from Opus 5, with cache reads down 60% to $0.20. Anthropic reports 66.4% on Terminal-Bench 4.0 and output more than 30% faster. OpenAI’s GPT-6 Sol is $2 in and $10 out, and Luna is $0.10 and $0.50, each half the price of the GPT-5.6 models that carried the same names. OpenAI reports 68.8% for Sol and 66.6% for Luna on DeepSWE v1.1.
Anthropic’s biology group, opened this spring, reports its first result. Roughly 950 Claude agents spent 21 hours and 210 million tokens searching phage genomes, and flagged a reverse-transcriptase enzyme sitting beside a long array of repeating DNA that resembles a CRISPR array. Dario Amodei said the system was found “mostly, though not entirely, by Claude.” He also acknowledged that Stanford researchers previously described a system that is in some ways similar.
Early users found that Muse, which launched September 8, ships configuration files with the same names and structure as the open-source agent OpenClaw, including a SOUL.md personality file. Nat Friedman replied that “we built Muse from scratch, but it is definitely heavily inspired as a product by OpenClaw.” Separately, researcher Patrick Wardle disclosed a zero-day in Muse’s Mac app. Meta issued a hotfix and said the practical risk to users was low.
The Series E was led by Insight Partners and S32, with Addition, Lightspeed, Greylock, GV and Wells Fargo also in. Valuation is nearly triple the $1.3 billion set 17 months ago. Snorkel says its annualized revenue run rate is $375 million, up eighteenfold in a year. It started out selling data-labeling software and now sells finished datasets and reinforcement-learning environments to AI labs.
Intrinsic Core is ROS-compatible and includes a hardware-agnostic real-time control framework, pose estimation built on NVIDIA’s FoundationPose, motion and grasp planning, Gazebo-based simulation and camera calibration. It supports arms from FANUC and Universal Robots, among others. It is aimed at industrial manipulation and does not cover mobile robots or humanoids. Intrinsic CTO Brian Gerkey called it “a durable commitment from Intrinsic to share this with the community.”
In the Council’s first session focused on AI safety risks, Sam Altman said the moment “calls for extreme care” and named two dangers: losing control of the future to AI, and power concentrating in too few hands. He asked for common standards for testing capabilities, human oversight and rapid incident reporting. Dario Amodei proposed narrow global agreements such as a ban on using AI to make biological weapons, plus evaluation and verification systems so countries can check each other’s commitments.
Yesterday’s theme was control of where your code runs; today’s is verification. Model prices keep dropping, and agents are now producing results, like a candidate enzyme, that humans have to confirm in a lab. Snorkel’s raise says the market will pay heavily for that checking, and the UN session shows the two biggest labs asking governments for shared ways to test and report. Meanwhile the plumbing is standardizing in the open, from agent config files to a robot control stack anyone can run. If you build on these tools, put your effort into evals and audit trails. The generation part is getting cheap.
— Boba
Curated by Vadym