Daily AI News — 26 Sep 2026

Akamai disclosed an $11.6 billion, seven-year agreement for Anthropic to run its growing CPU workloads on Akamai's distributed cloud, with an option to expand the total to roughly $20 billion. As part of the deal, Akamai issued Anthropic a warrant for about 5% of its common stock (7.74 million shares), vesting against contract milestones, plus a $1.7 billion memory pre-purchase through manufacturer Jabil. Akamai shares jumped more than 15% in extended trading. It is the largest contract in Akamai's history and a major bet that edge-first infrastructure can serve frontier AI.

This is the latest in a wave of 'vendor financing' deals - infrastructure providers taking equity in exchange for long-term compute commitments, following the Nvidia-OpenAI template. It shows how much capital frontier labs now need, and it validates that non-hyperscaler clouds can win serious AI workloads. Analysts are split on whether it signals healthy demand or circular dealmaking.

If you run infrastructure-heavy software, watch how compute contracts are now structured around equity and multi-year lock-ins. For buyers, more suppliers competing for AI workloads means more pricing leverage and less dependence on the big three hyperscalers.

OpenAI launched GPT-Live, a new generation of voice models powering ChatGPT Voice with a full-duplex experience: the model can listen while it speaks, handle interruptions, and follow the conversation as it shifts. Crucially, Voice now taps ChatGPT's agentic features - plugins, connected apps, and, for ChatGPT Work users, the ability to create documents, decks and spreadsheets by voice on web, mobile and desktop. GPT-Live-1 and GPT-Live-1 mini are rolling out across Plus, Pro, Business, Edu and Enterprise, with availability and limits varying by plan.

Turn-taking has been the thing that made voice assistants feel robotic. Full-duplex plus agentic tool access turns voice from a novelty into a genuine hands-free work interface - closer to talking to a capable colleague than dictating to a machine.

Draft an email, kick off a spreadsheet, or brief yourself between meetings entirely by voice while your hands are busy. Then pick the task back up in text at your desk - the conversation and the work carry over.

Meta's personal AI agent Muse became the top free app on both the Apple App Store and Google Play this week, overtaking ChatGPT, Claude and Gemini. At its Connect conference, Meta expanded Muse with real-time video avatars, a dedicated agent email address, and computer-use on Mac, and announced hardware to carry it everywhere: Ray-Ban Meta Gen 3, camera-free Ray-Ban Audio glasses shipping October 13 at $349, and a Tamagotchi-style pocket device called Muse Charm. Muse can browse, click, type and act across authorized apps, gated by a separate control layer Meta calls Sentinel.

Muse's growth is a real headache for OpenAI: it proves a capable consumer agent can run without OpenAI's frontier models, and that Meta will spend heavily to keep a rival free. The battleground is shifting from raw model quality to distribution - who gets the agent in front of the most people.

Point Muse (or a rival agent) at recurring chores - cancelling unused subscriptions, drafting replies, booking travel - but scope its app access tightly and revoke what it doesn't need. Agent convenience and account exposure rise together.

In his first media appearance as head of Google DeepMind, Koray Kavukcuoglu said Gemini 4 is in the refinement (post-training) stage and Google aims to release it 'much earlier' than the end of 2026, promising an early post-training output soon. It would be Google's first new flagship since Gemini 3 in late 2025 - a gap in which OpenAI's GPT-6 and Anthropic's Mythos and Opus lines pulled ahead. Kavukcuoglu, who replaced Demis Hassabis in August, reframed DeepMind's mission from chasing AGI to building 'intelligent agents we can trust.'

Google has been visibly behind on frontier models even as its Gemini app crossed a billion monthly users. A competitive Gemini 4 would reset the three-way race and give developers a serious third frontier option - important for pricing and avoiding lock-in to one lab.

If you're picking a model provider, hold some flexibility in your stack: a strong Gemini 4 launch could shift the price-performance calculus within weeks. Build against abstraction layers so you can swap models without rewriting your app.

The Trump administration, via the Office of the National Cyber Director, asked OpenAI and Anthropic to withhold their newest frontier models from the UK's AI Security Institute until US reviewers finish their own checks, Politico reported and Bloomberg confirmed. Officials framed it as securing US-made systems before sharing them with partners, under a June executive order that caps reviews at 30 days and exempts open-weight models. Anthropic has already complied, withholding Claude Mythos 5.1 from the UK institute and limiting it to a set of US organizations; OpenAI has not detailed its response.

It signals a more nationalist turn in AI governance, straining the close US-UK safety-testing relationship and the idea of shared international evaluation. How labs navigate competing government demands will shape who gets early access to the most capable models - and where.

If your product depends on early access to frontier models, factor in geography: launch timing and model availability may now diverge by country. Enterprises in the UK and EU should plan for possible lag on the newest US systems.

Anthropic said Claude, running as roughly 950 agents over about 21 hours and 210 million tokens, autonomously scanned some 200,000 genomic records, narrowed 3,500 candidates to 20 systems, and identified a novel CRISPR-like enzyme system. CRISPR pioneer Feng Zhang called the work 'genuinely intriguing,' and a technical preprint was published alongside the announcement. The company paired it with a new Life Sciences Verification Program and a dedicated internal biology lab, positioning Claude for scientific-discovery workloads while stressing new biosecurity safeguards.

This is a concrete example of AI agents doing real, verifiable scientific search at a scale and speed humans can't match - not a chatbot demo. If it holds up under peer review, it points to AI as a genuine discovery engine in biology, with all the promise and dual-use risk that implies.

Researchers can use large agent swarms to triage enormous candidate spaces - genes, molecules, materials - down to a short list worth wet-lab testing, compressing months of literature and database work into a day.

Enterprise security firm Island raised $400 million in a round led by Evolution Equity Partners - with Sequoia, Coatue, Insight, Prysm, Cyberstarts and J.P. Morgan Growth participating - valuing the Dallas-based company at $6.4 billion, more than 30% above its $4.8 billion mark in 2025. Island builds a secure enterprise browser. Investors are piling into cybersecurity as AI adoption creates new risks, including rogue AI agents that can act beyond intended controls. The raise landed the same week Australia disclosed an OpenAI agent breach of a government health portal.

AI agents that browse, click and act are becoming a new attack surface, and the enterprise browser is where much of that activity happens. The valuation jump shows security budgets shifting toward controlling what autonomous software is allowed to do inside a company.

Before you deploy agents on employee machines, decide where they run and what they can touch. A locked-down enterprise browser is one way to give agents a workspace without handing them the keys to everything.

OpenEvidence, a clinical-decision-support platform used by physicians, raised $250 million at a $15 billion valuation - about 25% above its $12 billion mark in January - according to Business Insider. Sources say founders are fielding acquisition interest and could be open to selling rather than pushing to an IPO. OpenEvidence surfaces evidence-based answers to medical questions grounded in peer-reviewed literature, aiming to help clinicians make faster, better-supported decisions at the point of care.

Healthcare is one of the highest-stakes proving grounds for AI, and a $15B valuation signals real clinical traction, not just hype. It also shows investors betting that vertical, domain-grounded AI - not general chatbots - is where durable value in regulated fields will accrue.

Clinicians can use grounded tools like this to check the latest evidence during a consultation instead of trawling journals afterward - provided the answers stay cited and verifiable, so the human stays accountable for the call.

At ROSCon in Toronto, Nvidia released Isaac ROS 5.0, extending its GPU-accelerated robotics stack to the roughly 1.3 million developers on the open-source Robot Operating System. The headline is agentic workflows: reusable skills and agent-readable docs that let AI coding assistants build, tune and deploy robot applications alongside human engineers. FoundationPose object tracking now runs up to 5.5x faster, collision-free path mapping benchmarked at 2-5 milliseconds, and the release adds ROS Lyrical and Ubuntu 24.04 support. Magna, Universal Robots, Flexiv, Intrinsic and others shipped integrations at launch.

Robotics has been bottlenecked by how hard it is to stitch perception, planning and control together. Letting agents assemble that pipeline from tested components lowers the skill barrier and speeds deployment - and deepens Nvidia's grip on the robot developer stack the way it owns AI training.

A robotics team can describe an outcome - 'pick this part and place it here' - and have an agent wire up detection, depth and pose from Isaac ROS modules, instead of hand-coding each stage. Faster prototyping, fewer specialists needed per project.

DeepSeek's annualized revenue reached about $1 billion, CEO Liang Wenfeng told investors, driven largely by an API price hike of 2.3x to 4.5x last month. The company plans to close a roughly $7.5 billion (50 billion yuan) fundraise via a Shanghai listing at a 500 billion yuan valuation by the end of October, The Information reported. DeepSeek built its reputation on highly efficient open-weight models that undercut Western labs on price, pressuring OpenAI, Anthropic and others to cut their own token costs.

DeepSeek shows the low-cost, open-weight playbook is commercially viable at serious scale - and that Chinese labs can raise huge sums domestically despite US export controls. Its pricing pressure is a big reason frontier labs have been slashing token costs.

When budgeting AI features, benchmark cheaper open-weight models against premium frontier ones on your actual tasks. For high-volume, well-defined workloads, the price gap can outweigh the capability gap - re-test before you commit.