Anthropic gives METR internal access to evaluate its frontier models, the White House publishes a voluntary accord on superintelligence, Spain's AI governance bill lapses, OpenAI watermarks text in the EU under the AI Act, the UK ICO secures commitments from ten labs and GPT-6.1 Sol arrives.
research
Anthropic gives METR internal access to evaluate the safety of its frontier models
On 15 October 2026 Anthropic announced that it has begun giving a team from METR, a non-profit organisation, internal access comparable to that of its employees to assess the risks of its most advanced models. METR may publish its analyses; Anthropic cannot veto publication, only require the redaction of narrowly defined categories of sensitive information. It also works with SecureBio and Accenture, and links the move to the White House Accord on Super Intelligence.
Why it matters: It is a concrete step by a frontier lab towards external safety and alignment oversight, the kind of control that matters most as models approach general capabilities.
White House: voluntary «Super Intelligence» accord with external audits
The White House Accord on Super Intelligence was published on 29 September 2026. Frontier companies voluntarily commit to four layers: internal capability and alignment controls, an internal oversight team, an independent external auditor and an independent board committee. Google, Anthropic, Meta, OpenAI, xAI and Nvidia signed, alongside President Trump.
Why it matters: It is the most recent institutional framework in the US on superintelligence governance and explains why labs such as Anthropic are bringing in external evaluators.
Spain: the AI governance bill lapses as Parliament is dissolved
Royal Decree 806/2026 dissolved the Congress and the Senate on 6 October 2026. The Organic Bill on the good use and governance of AI is listed by Congress as lapsed from that date. The EU AI Act continues to apply directly; AESIA exists as an agency, but the national penalty regime planned in the bill is postponed to the next legislature.
Why it matters: Spain will have no organic AI governance law this legislature, while the EU AI Act remains in force for businesses and users.
OpenAI applies text watermarks in the EU under the AI Act
On 5 October 2026 OpenAI explained its response to the AI Act's text provenance rules: opt-in watermarking in the API globally, a gradual rollout of an invisible mark in ChatGPT and Codex in the EU only, and a detector initially limited to researchers and expert organisations. It acknowledges limits (short texts, editing, maths) and presents its textGrain technology.
Why it matters: It is a concrete instance of a frontier provider complying with the AI Act's transparency rules; it affects anyone using ChatGPT or the API in the European Union.
The UK data regulator secures changes from ten foundation model developers
On 8 October 2026 the ICO (United Kingdom) reported data protection commitments from Amazon, Anthropic, Apple, Cohere, DeepSeek, Google, Meta, Microsoft, OpenAI and Stability AI. It opened a six-week consultation on agentic AI risks (until 20 November) and confirmed enquiries into recent agent trials.
Why it matters: It shows that data oversight now reaches foundation models and more autonomous systems, not just individual apps.
OpenAI launches GPT-6.1 Sol, close to Astra at a fraction of the price
On 29 September 2026 OpenAI unveiled GPT-6.1 Sol, a model that, according to the company, comes close to GPT-6 Astra in agentic coding, computer use and professional work at roughly a fifth of the standard token price. It is available in ChatGPT Work, Codex and the API (`gpt-6.1-sol`); not yet in the general chat. OpenAI highlights alignment improvements over GPT-6 Sol.
Why it matters: It lowers the cost of near-frontier capabilities, which changes who can use that level of AI in everyday professional work.