AI News: May 13, 2026

1. Thinking Machines Lab ships TML-Interaction-Small, a full-duplex voice model

Thinking Machines. Mira Murati’s lab released its first model, TML-Interaction-Small, a 276-billion-parameter mixture-of-experts model with 12 billion active parameters that processes audio, video, and text in 200-millisecond chunks rather than rigid turn-taking. The lab reports a 0.40-second response latency versus 1.18s for GPT-Realtime-2 minimum and 0.57s for Gemini Live, and pairs the fast interaction model with a background reasoning model. The release goes to a limited research preview group, with broader access planned for later this year. Source

2. Isomorphic Labs raises $2.1B Series B led by Thrive to push AI drug discovery into clinical trials

Isomorphic Labs. Demis Hassabis’s Alphabet drug discovery company closed a $2.1 billion Series B led by Thrive Capital, with Alphabet, GV, MGX, Temasek, CapitalG, and the UK Sovereign AI Fund participating. The round funds expansion of its IsoDDE platform, a pipeline of AI models that combine to design drug candidates, and advancement of internal candidates toward human clinical trials. Existing partnerships with Novartis, Lilly, and Johnson & Johnson remain. Source

3. Voice AI startup Vapi hits $500M valuation after Amazon Ring picks it over 40 rivals

Vapi. San Francisco-based voice agent platform Vapi raised $50 million in a Series B led by Peak XV with backing from Kleiner Perkins, bringing total funding to $72 million at a $500 million valuation. The company reports that its voice agents have handled more than one billion calls across finance, healthcare, insurance, and automotive, and that Amazon Ring selected the platform over 40 competing vendors for customer support automation. Source

4. Medicare’s new ACCESS payment model is built around AI agents

CMS. Medicare’s ACCESS model creates the first US governmental reimbursement mechanism for AI agents that monitor patients between visits and coordinate care, paying for autonomous workflows rather than only clinician time. The TechCrunch piece flags that most of the AI ecosystem has missed the policy shift even though it materially changes the addressable revenue for clinical AI vendors. Source

5. AI-native sales platform Monaco raises $50M Series B led by Benchmark

Monaco. Monaco closed a $50 million Series B led by Benchmark, bringing total funding above $85 million, after onboarding hundreds of customers during a public beta following a February stealth exit. The platform pitches an “AI-native” sales workflow that replaces stitched-together SaaS with agent-driven account research, outreach, and forecasting. Source

6. Cyber startup Frame, founded by ex-Wiz and Team8 leaders, raises $50M

Frame. Israeli cybersecurity startup Frame raised $50 million from investors led by Bessemer Venture Partners, with founders from Wiz and Team8 building an AI-driven security platform aimed at autonomous threat triage. The round arrives as security teams increasingly look at agent-based tooling for SOC work, and as governments warn that AI is already being used by attackers to find and weaponize vulnerabilities. Source

7. Tokenmaxxing spreads at Amazon as employees game internal AI leaderboards

Amazon. Amazon employees are reportedly automating unnecessary tasks to climb internal AI usage leaderboards, surfacing a familiar Goodhart’s-law problem now that token consumption is being used as a performance signal. The trend illustrates the limits of using raw AI usage as an incentive proxy and previews how other large enterprises tracking similar metrics may see distorted behavior. Source

8. Addy Osmani publishes a senior-engineer agent skills framework

Addy Osmani. A new post details an agent-skills framework that enforces senior-engineer production workflows rather than suggesting them through prompts, using anti-rationalization tables and a parallel fan-out command architecture. The piece targets teams running coding agents in production who keep running into prompt-only guardrails failing under load. Source

9. Self-engineering agents like DGM-H show independent emergence of memory systems

Alpha Signal. A roundup on self-improving agents highlights Darwin-Gödel Machine and Hyperagents research where systems modify their own code to build more robust scaffolding, with DGM-H showing independent emergence of memory systems and evaluation pipelines across diverse domains. The trend points to scaffolding, not raw model capability, as the next axis of agent performance gains. Source