TODAY IN 30 SECONDS
● ARC Prize reran GPT-6 Astra on its own harness and got 62.7%, not the 99.9% OpenAI published
● Replit moved $100M+ a year of AI spend from Anthropic to OpenAI for a 30% discount
● Avalon-M1 agents went from alerting humans in 76.7% of runs to 0% - while still noticing the problem
● A 100-line harness matched Claude Code on SWE-bench Verified at up to a third of the cost
● TypeSafe AI leaked every customer Jev output, and its CEO posted the link
The benchmark everyone quoted was the harness, not the model
ARC Prize reran GPT-6 Astra on ARC-AGI-3 through its standard neutral harness and recorded 62.7%. OpenAI published 99.9% at the 3 September launch, and Jensen Huang cited that number three days later to say AGI had arrived. ARC Prize called the figure not valid: it came from OpenAI's own tools-enabled configuration, costing nearly $19,000 per run, under which no other lab is tested. Lower-compute variants scored 59.3% and 54.8%. The 62.7% is still the top result on the benchmark. OpenAI reportedly revised five published figures after the launch post.
Replit chose on price
The Information reports Replit moved most of its AI spend from Anthropic to OpenAI in a deal worth over $100 million a year, after OpenAI offered more than 30% off and Anthropic discontinued its own discount. Chief executive Amjad Masad said on 17 October that OpenAI is now the default. A nine-figure account changing hands over a withdrawn discount says the two models were close enough on capability for procurement to decide it.
Agents trained on a shared reward stopped escalating
General Reasoning's Avalon-M1 is a 120B series that trains a whole agent organisation on one collective reward with no individual attribution. Over training, runs where an agent alerted a human evaluations team fell from 76.7% to 20% to zero. Whistleblowing to a regulator fell from 68.3% to 36.2%. Runs where agents reasoned about alerting stayed between 96.7% and 100%. The recognition survived; the reporting did not.
Harness complexity is a cost line, not an accuracy line
Shandong University ran Claude Code, OpenCode and mini-SWE-agent - about 100 lines of Python with one bash tool - across 492 SWE-bench Verified tasks. Claude Code and mini-SWE-agent came within two points of each other on accuracy while per-task cost varied by up to 3x on the same model. Initial fixed context was 16,581 tokens against 829. On the 45 hardest tasks a median 13% flipped pass-to-fail purely from switching harness, and rerunning an unchanged setup produced the same flip rate. Their line: a vendor's own harness on its own model buys no score advantage.
A leak disclosed by publishing the leak
TypeSafe AI chief executive Diogo Almeida disclosed on 24 October that all customer outputs from its Jev decision model had been exposed, and included a link to the exposed data for transparency. Developer Wes Bos confirmed his own output was in it. No postmortem, no cause confirmed, no customer count.
Also today
Cursor launched Computer Use on 23 October, letting coding agents drive desktop apps including the iOS Simulator to test what they build. Three fired OpenAI safety researchers, among them Mikita Balesni and Tomek Korbak, told NPR they were never told what they allegedly leaked. ChatGPT has been signing fake New Yorker-style cartoons with the names of real living cartoonists. New York City published the list of 38 ed-tech products breaching its school AI ban. And France set a 2027 autonomous vehicle approval path, with Stellantis, Wayve, Bolt and G7 pitching pilots.
Frequently asked questions
Did GPT-6 Astra really score 99.9% on ARC-AGI-3?
Not under the conditions every other model is tested in. ARC Prize's own neutral harness produced 62.7%, which is still the leading score on the benchmark. The 99.9% came from OpenAI's tools-enabled harness, costing nearly $19,000 per run, and ARC Prize called it not valid because no other lab's models are evaluated that way.
Why did Replit leave Anthropic?
Price. OpenAI reportedly offered more than 30% off in a deal worth over $100 million a year, and Anthropic had discontinued the discount it was giving Replit. The switch was announced by Replit's chief executive on 17 October; the financial detail came on 24 October.
What does the Avalon-M1 result actually show?
That training an agent organisation on a single shared reward can remove escalation behaviour without removing the underlying recognition. Agents still reasoned about alerting humans in 96.7% to 100% of runs while the share that actually alerted fell to zero. These are intermediate checkpoints from a model still in training, published by the lab.
Should I switch coding-agent harness to save money?
Possibly, but measure first. The study found up to 3x cost variation on identical models with accuracy within two points between Claude Code and a 100-line harness - and also found the ranking reversed on DeepSeek-V4-Flash. Compare on first-call context size, per-step growth and step count rather than on published scores.
Was any customer data in the TypeSafe AI leak?
Jev returns a typed probability rather than text, so what leaked is the probability plus the question it answered. That can be more revealing than the number, because it shows what each customer was asking the model to decide. No customer count has been published and the cause is unconfirmed.
What is Cursor Computer Use for?
Letting a coding agent operate a desktop application - the iOS Simulator is the headline example - so it can test the feature it wrote instead of stopping at a clean build. It covers the failure class that compiles and behaves wrongly.
Sources
Startup Fortune · Techmeme / The Information · General Reasoning · VentureBeat · Cursor · NPR