Transparency report

What it does

When a lab releases a new frontier model, it publishes a short report on what the model is for, which risks it was tested for, and what the tests found.

Applies to: All frontier developers; extra content for large ones.

• Before or concurrently with deploying a new frontier model, publish a report with intended uses, restrictions, modalities and, for large developers, summaries of catastrophic-risk assessments and the role of third-party evaluators (SB 53 § 22757.12(c)).
• Send the state a summary of internal-use catastrophic-risk assessments every three months.
• EU Code signatories file a Model Report at deployment.

Where things stand

OpenAI publishes a system card with every frontier release (GPT-5.5 on April 23, 2026; GPT-6 Astra on September 3, 2026). Anthropic publishes system cards at launch and Risk Reports every three to six months. The legal document is the system card with a checklist.

Why it costs almost no lead

Labs already publish these on launch day. Formatting them to legal requirements costs almost nothing in research time.

Biggest unknown: Whether a legally required document ever finishes later than the launch it accompanies.

Why it lowers p(doom) by ~0.03%

Tells the public what a new model can do and what the lab tested before it shipped.

Disclosure alone changes little, but it gives outsiders what they need to catch problems.

The strongest case that it costs more

The definition of deployment is broad. GovAI found that in the first five months of 2026, the frontier models withheld from general release were instead given to select entities under trusted-access programs. If a transparency report is required before any third party touches the model, those programs would have to wait for the write-up. That could cost a week or two on the access clock per model, and it is why the high end of the range is a week.

The debate

For

• California SB 53 requires a transparency report before or when a new frontier model is deployed, 2025.
• EU AI Act requires documentation and a public summary of training content for general-purpose models.
• Anthropic says it already publishes system cards and backs making disclosure mandatory, 2025.

Against

• Meta refused to sign the EU's code of practice, saying Europe is heading down the wrong path, 2025.
• Chamber of Progress warned SB 53's mandates expose intellectual property, 2025.

Sources

• California Bus. & Prof. Code § 22757.12(c) (SB 53): transparency reports
• GPT-5.5 System Card: published on release day
• Anthropic RSP v3.4: Risk Reports every 3 to 6 months
• GovAI, Delays to Frontier AI in the EU and UK: trusted-access programs in 2026