THE 60-SECOND VERSION
● K2 Horizon: six models, 0.9B to 375B, with training data and logs published.
● Daybreak first partners: state cyber defenders and water systems, with training attached.
● 9 days to Claude Code weekly limits settling 17 percent lower.
A model family released with its training data
The Institute of Foundation Models published K2 Horizon — six models from 0.9 billion to 375 billion parameters, with weights, training code, training data, checkpoints and training logs.
Every other open-weight release this summer shipped the weights and kept the rest, which is exactly what the Open Source Initiative has been objecting to: you cannot inspect a weights file, reproduce it without the data, or verify what went into it. This one you can.
It is not a capability claim — nothing here says frontier performance. It is a verifiability claim, and for regulated work or research that is the more valuable axis.
Why publishing training data is the part nobody does →
Daybreak names its first participants
Two days after gating GPT-6 Astra's cyber capability behind an application programme, OpenAI announced a pilot with MS-ISAC covering state, local, tribal and territorial cyber defenders, plus water systems — pairing Daybreak access with guided training and hands-on assistance.
Municipal water utilities run critical control systems and are typically defended by teams of one. They are precisely the defenders who could not otherwise buy this capability, which makes it a better answer to the gating objection than a policy statement would have been.
What is still unpublished is how the gate is enforced technically — the detail that decides whether it is a control or a policy.
What the pilot covers, and what remains unanswered →
Also this week
- Anthropic published an agent-first software playbook, with the Claude Code team reporting that 70 to 80 percent of its work now runs through AI-driven remote workflows.
- Astra benchmarks are starting to appear from third parties. Reported figures include 84 percent on Epoch's Mystery Game Puzzles, and a MirrorCode result landing between Opus 4.7 and Fable 5.
- Google Research and HHMI Janelia mapped the complete male fruit-fly nervous system — over 166,000 neurons and 125 million synapses.
The month ahead
| Date |
What happens |
| Sept 14 |
Claude Code weekly limits settle 17 percent below current levels |
| Sept, unconfirmed |
OpenAI listing window. No public S-1 on EDGAR |
| Early-mid Sept |
Grok 4.7 window per Musk. No model ID published |
| Oct 1 |
OpenAI vs Apple hearing |
| Nov 12 |
OpenAI models leave Cursor |
Still open
FAQ
What is K2 Horizon?
Six models from the Institute of Foundation Models, 0.9 billion to 375 billion parameters, released with weights, training code, training data, checkpoints and training logs.
Why does publishing training data matter?
It is the part that allows reproduction and audit, and almost no lab publishes it. Without it, open weights and open source are different things.
Who gets Daybreak access first?
A pilot with MS-ISAC covering state, local, tribal and territorial cyber defenders plus water systems, with guided training alongside the access.
What is the next confirmed deadline?
14 September, when Claude Code weekly limits settle at 25 percent above the pre-May baseline — 17 percent below current levels.