Skip to content
IIAIT

AI Today · 4 November 2026

Apple opens Siri AI to developers, and Anthropic tests proofs of which model produced an answer

Apple lets any developer connect apps to Siri AI, though not on iPhone and iPad in the EU, Anthropic publishes early research on verifying which model produced a response, and Revolut describes how self-hosted models changed its AI coding costs.

  1. 01

    Apple opens Siri AI to all app developers, but not on iPhone, iPad or Apple Watch in the EU

    On 3 November Apple launched a developer preview that lets any developer connect an app to Siri AI through the App Intents framework, starting with the iOS, iPadOS, macOS and visionOS 27.3 betas. Apps can expose their content and actions, and a new Skills feature lets developers combine several actions into one request. Apple expects the release for customers in a few months, with App Store submissions opening in spring. Siri AI is not available in the EU on iOS, iPadOS and watchOS, and Apple says it is working with the EU on a way forward. In its annual filing with the US Securities and Exchange Commission on the same day, Apple listed AI agents acting in unintended or unpredictable ways as a business risk, according to CNBC.

    What it means for you

    If your company has an iOS or Mac app, you can start testing how Siri AI would use its actions, but colleagues in the EU cannot use it on iPhone, iPad or Apple Watch for now. Before you let any assistant act across apps, decide which actions it may take without your confirmation, such as sending messages or placing orders.

  2. 02

    Anthropic publishes a research prototype for proving which model produced a response

    Anthropic's Frontier Red Team, with researchers from EPFL, MIT and other organisations, released FlashProve, open-source code that creates a cryptographic proof that a previously committed model generated a given output, without revealing the model's weights. In tests on Meta's Llama 4 Maverick, producing proofs took roughly 70 to 95 times the GPU time of normal serving, and a proof could be checked on an ordinary processor in about 17 to 21 seconds. Anthropic calls it a research prototype with clear limits, including a 1,024-token context and support only for greedy decoding, where the model always picks its most likely next token. TechCrunch reports that there are no plans yet to deploy it at Anthropic or any other lab.

    What it means for you

    Nothing changes in your tools today. The work addresses a practical question for anyone buying AI services: whether a provider really runs the model you pay for. Until such checks exist, ask suppliers to state in the contract which model and version handle your requests.

  3. 03

    Revolut says self-hosted open-weight models cut its AI coding costs by about 90%

    According to The Register, Revolut's operations and infrastructure lead told the AI Engineer conference in New York that the bank reduced its AI coding inference costs by roughly 90% in under a year by making self-hosted models its default instead of commercial services such as Claude Code. Revolut runs open-weight GLM models from Z.ai on 56 Nvidia B200 GPUs for about 2,000 engineers, at around 0.30 dollars per million tokens, and sends tasks involving images to models such as Gemini or GPT. He said that tuning the system mattered more than adding hardware. The figures are the company's own and have not been independently verified.

    What it means for you

    Self-hosting can pay off at large scale, but it needs dedicated GPUs and an infrastructure team, which most smaller organisations do not have. When comparing options, count hardware, staff and security review, not only the price per token, and check the licence and data location of any open-weight model you consider.