Watchtower

Watchtower tracks when AI companies change their safety policies — and when they break them.

OpenAI

OpenAI updated the GPT-6 Astra system card to include information about GPT-6 Sol and GPT-6 Luna. There were also updated alignment evaluations and new results for Astra.

Major
Change

OpenAI

Sep 22, 2026

OpenAI updated its GPT-6 Astra System Card on Sept. 22, 2026. The update includes an appendix providing system card information for GPT-6 Sol and GPT-6 Luna, released the same day.

Both GPT-6 Sol and GPT-6 Luna are treated as “High” in the Cybersecurity and Biological/Chemical domains, and as below High in AI Self-Improvement. These are the same ratings as GPT-5.6 Sol and GPT-5.6 Luna, so they receive the same safeguards as the GPT-5.6 models rather than Astra’s Critical-level protections. Both use the same types of data and training as Astra, and the appendix mainly reruns the evaluations done in the main body of the system card.

OpenAI also updated five of its alignment evaluations. Updated results for GPT-6 Astra and GPT-5.6 Sol were placed alongside the results for GPT-6 Sol and GPT-6 Luna in the appendix. 

Astra’s numbers improved noticeably on coding deception and broken search. This results in the ratios given in the main body of the card becoming stale. The body says GPT-5.6 Sol’s deception rate is “4x” Astra’s and its search failure rate “10x,” while those ratios are about 20x and 50x for the updated runs.

The changelog also says a note was added in the main body of the card to point to GPT-6 Astra’s updated results, but no such note exists within the Alignment section of the card.

The above is not an exhaustive list.

A diff1 of the changes can be found below:

  1. This diff is between Sept. 9 and Sept. 29. We were unable to find a pdf version of the system card between Sept. 22 and Sept. 29. There is a separate Watchtower entry for the changes on Sept. 29.

OpenAI

Oct 29, 2026

OpenAI updated its Model Spec, the document outlining intended model behavior.

View details

Meta

Oct 23, 2026

Meta quietly updated Wonder’s model card nearly two weeks after the model’s release, with still no sign of a safety and preparedness report for the model.

View details

OpenAI

Oct 21, 2026

OpenAI updated its Usage policies. The new policies will go into effect on Nov. 20.

View details

Anthropic

Oct 19, 2026

Anthropic updated its Frontier Safety Roadmap which publicly charts some of its highest-priority goals.

View details

Google

Oct 13, 2026

Google launched Gemini 4 Argon along with the first “full” safety and preparedness report made under its Frontier Safety Framework (v3.1), published in April 2026.

View details

Meta

Oct 12, 2026

Meta publicly launched Muse Wonder with a 13-page model card and without concurrently publishing a safety and preparedness report.

View details

Frequently asked questions

Have more questions? Our team is happy to help, contact us.

What does The Midas Project do?
What does your name mean?
Who is behind The Midas Project?
Are you anti-AI?
How can I contribute?
How can I get in touch?