AI Industry
News Digest

GPT-Realtime-2 Arrives with 128K Context, Live Translation, and Streaming Transcription

If you build anything that talks back, this is the most consequential update of the day. OpenAI has shipped three new audio models inside the Realtime API: GPT-Realtime-2 with GPT-5-class reasoning; GPT-Realtime-Translate covering 70-plus input languages live; and GPT-Realtime-Whisper for low-latency streaming transcription. The context window has jumped from 32K to 128K, parallel tool calls are now supported, and you can pick reasoning effort from minimal through xhigh.

Voice agents can now hold longer flows without losing track. Pricing is set at $32 per million audio input tokens for Realtime-2 and $0.034 per minute for live translation, so this is firmly priced for production rather than demo work. Anyone shipping a phone agent, support bot, or in-car assistant should plan an evaluation this week.

GPT-5.5-Cyber Enters Limited Preview as Trusted Access for Cyber Widens

OpenAI is opening up a defender-only tier of GPT-5.5 with looser refusals on legitimate security work, and adding a new GPT-5.5-Cyber preview for verified red and blue teams who need exploit proof-of-concepts and live target validation. For most users, the practical effect is access tier reshuffling: regular GPT-5.5 stays as is, GPT-5.5 with Trusted Access for Cyber becomes the default for vetted defenders, and GPT-5.5-Cyber sits behind stronger account checks.

Anyone running internal vulnerability tooling on the API should review whether they need to apply for Trusted Access, because Advanced Account Security becomes mandatory for cyber-permissive models from 1 June 2026.

OpenAI, AMD, Broadcom, Intel, Microsoft, and Nvidia Ship the Multipath Reliable Connection Protocol

This is plumbing rather than a product, but it matters because it is the protocol that will sit underneath the next round of frontier-model training runs. MRC spreads each transfer across hundreds of network paths, reroutes around failures in microseconds, and is now baked into 800 Gb/s NICs from the same six companies. OpenAI says it is already deployed across its largest GB200 clusters at Oracle and at Microsoft Fairwater.

The specification has been released to the Open Compute Project, making it available for anyone building large GPU fabrics, including European sovereign cloud projects and the next wave of neocloud providers. Expect the practical knock-on to be faster training cycles and fewer mid-run failures across most major labs.

Trusted Contact Arrives in ChatGPT for Adult Users at Risk of Self-Harm

ChatGPT now lets adults nominate one trusted person who can be alerted by email, SMS, or in-app notification if reviewers and automated systems judge a conversation to indicate a serious self-harm risk. The feature is opt-in on both sides and will not share transcripts, only a brief notice. For most builders this is not directly actionable, but it is worth knowing about because it materially changes ChatGPT's safety surface. It is also a useful reference point if you are designing safety flows in any consumer product on top of the API, where similar duty-of-care expectations are starting to land.

ChatGPT Self-Serve Ads Manager Opens to US Small Businesses

OpenAI has opened up its ChatGPT advertising test so any small business in the United States can sign up directly, set budgets, upload creative, and launch campaigns inside a new self-serve Ads Manager. Ads only appear for logged-in adult users on the Free and Go tiers, are clearly labelled, and are blocked in sensitive categories. Bid mechanics use cost per click. If you are an end user on Free and would rather not see ads, there is now an opt-out which trades sponsored placements for a lower daily message cap.

AlphaEvolve Graduates from Research Project to Operational Tool across Science and Infrastructure

DeepMind has published a one-year retrospective on AlphaEvolve, the Gemini-powered coding agent that designs and optimises algorithms, and the headline numbers are strong enough to make this more than a marketing post. AlphaEvolve has cut variant detection errors in PacBio sequencing pipelines by 30%, lifted feasible-solution rates on the AC Optimal Power Flow problem from 14% to over 88%, and reduced quantum-circuit error rates by 10x on Google's Willow chip.

For practitioners, the interesting signal is that DeepMind is happy to position AlphaEvolve as a partner for working scientists rather than a benchmark-chaser, which suggests the underlying loop is robust enough for serious problems. If you work on optimisation-heavy domains like genomics, energy, or simulation, this is a good prompt to revisit whether evolutionary code search belongs in your stack.

Personal Computer Agent Rolls Out to All Pro and Enterprise Users inside a New Native Mac App

Perplexity has replaced its old Mac client with a fully rebuilt native macOS app that bundles its Personal Computer agent for every Pro and Enterprise subscriber, not just Max. The agent can drive native Mac apps, touch the file system, and run web tasks inside a sandbox where every action is auditable and reversible.

If your team has been waiting for a credible alternative to Anthropic's computer use or OpenAI's Codex desktop app, this is a real one to evaluate this week. The older Perplexity Mac app is being deprecated, so existing users will need to migrate. For Windows users, this also signals where Perplexity intends to take its agent stack next.