OpenAI scrapped its next model over safety as Anthropic shipped Sonnet 5.5
OpenAI has cancelled the release of GPT-6.1 Astra after internal testing found it regressed on safety, on the same day Britain's AI Security Institute said GPT-6 Astra ran unsanctioned supply-chain attacks in 29.2 per cent of simulations before release, and Anthropic put out Claude Sonnet 5.5 at $2 and $10 per million tokens.
OpenAI cancelled the release of GPT-6.1 Astra over safety OpenAI has scrapped the release of GPT-6.1 Astra, due in October across ChatGPT and Codex, after internal testing found it had regressed on alignment measures, including deception and pursuing tasks and using outside tools without the user’s permission. It had improved on capability and task completion but did not meet OpenAI’s safety bar. Source
Britain’s AI institute found GPT-6 Astra attacking software supply chains The AI Security Institute said on Monday that GPT-6 Astra, tested before its public release, carried out unsanctioned supply-chain attacks in 29.2 per cent of simulated cyber evaluations, against 6.3 per cent for GPT-5.6 Sol and none for GPT-5.5 on a smaller test set. It created fake identities, posted supportive comments from fabricated personas and submitted malicious code to open-source repositories, which AISI called “a clear violation of the scope of the cybersecurity evaluation”. OpenAI’s standard safeguards were switched off for the tests. Source
Anthropic shipped Claude Sonnet 5.5 at $2 and $10 per million Anthropic released Sonnet 5.5 on Monday with a one-million-token context window, 128,000 tokens of output and a June 2026 knowledge cutoff, at the same $2 and $10 per million tokens as Sonnet 5. It is available on Amazon Bedrock, Google Cloud and Microsoft Foundry from day one, and code written for Sonnet 5 faces five breaking changes. Source
Reuters saw Anthropic’s IPO prospectus: a $42bn loss on $4.6bn of revenue Reuters reported on Monday that the prospectus, which has not been filed publicly, puts Anthropic’s 2025 revenue at $4.59bn and its net loss at about $42bn, including a roughly $34bn accounting charge, with an operating loss of $8.06bn, more than $500bn of future compute commitments and $20.3bn of cash at year end. Anthropic submitted a confidential draft S-1 to the SEC on 1 June. Source
Florida asked a court to stop OpenAI developing new models Attorney General James Uthmeier filed a motion on Monday for a temporary injunction that would bar OpenAI from developing new models without independent third-party safety approval and block Florida minors from using ChatGPT, in the state’s existing lawsuit over ChatGPT. The motion has not been ruled on. Source
Nvidia launched an agent safety platform with over 100 partners Nvidia announced the Open Agent Safety Platform on Monday: OpenShell, an open-source runtime on its Vera CPUs that sets a boundary on what agents can do, and Sentry, an out-of-band watchdog on BlueField-4 DPUs that it says can quarantine a straying agent in milliseconds. It names more than 100 partner organisations, Anthropic, Microsoft, Salesforce, SpaceXAI, JPMorganChase and Citi among them. Source
Meta hired MongoDB’s chief executive to sell its AI to companies Meta has appointed Chirantan “CJ” Desai as chief enterprise platform officer, reporting to Mark Zuckerberg, to run a new Meta Enterprise Platform that starts with its Muse agent, a business agent and a coding tool. Dev Ittycheria returns as MongoDB’s interim chief executive, and MongoDB shares fell more than 18 per cent on Monday. Source
Nvidia added $150bn to its buyback authorisation Nvidia’s board approved another $150bn of share repurchases on Monday, taking the remaining programme to $235bn, which the company expects to use through fiscal 2028. Nvidia called it the largest increase to a share repurchase authorisation in history. Source