Apple-OpenAI deal cracks; humanoids run 8-hour shifts

Anthropic has received offers valuing it at over $900 billion, which would more than double its current valuation and surpass OpenAI's $852 billion. Its revenue run-rate hit $30 billion in April, up from $9 billion at the end of 2025, with growth running eight times faster than planned. Ramp data this week showed Anthropic finally leading OpenAI in business adoption.

Discovery in the Musk v Altman trial revealed Microsoft CEO Satya Nadella was worried about OpenAI overtaking his company as far back as April 2022. "It was becoming even more core and important that we had real agency at every layer of the stack," Nadella testified. Microsoft has poured over $100 billion into OpenAI but has struggled to compete at the model layer itself.

Cerebras Systems priced its IPO at $185 a share on Wednesday, above its expected range, raising at least $5.55 billion. CEO Andrew Feldman is now sitting on a $1.9 billion stake. The deal is one of the largest US tech IPOs since Snowflake's $3.8 billion offering in 2020, and clears runway for SpaceX, OpenAI and Anthropic to follow later this year.

Microsoft is now actively scouting AI startups to acquire after pouring more than $100 billion into OpenAI, according to multiple reports this week. The shift lines up with Nadella's leaked 2022 worry that OpenAI could supplant Microsoft, and with the company's struggles to ship competitive foundation models in-house. M&A is the hedge against single-vendor risk, especially as OpenAI's exclusivity in places like Apple starts to crack.

Cursor released tools for configuring cloud agent development environments with multi-repo workspaces and persistent infrastructure. The goal is to let agents handle engineering tasks end-to-end with full context on the codebase and organisation, running tests, querying services, and reaching internal APIs. Cursor's own agents use these environments too. The release targets teams running fleets of parallel cloud agents on actual production work.

Apple's two-year partnership with OpenAI has become strained, according to Bloomberg's Mark Gurman, with OpenAI lawyers actively preparing options including a breach-of-contract notice. OpenAI believed the deal would coax iPhone users into ChatGPT subscriptions and yield deeper Siri integration, but Apple's implementation has been minimal. iOS 27 will open Siri to Claude and Gemini, eroding OpenAI's exclusivity further.

Figure AI claims its humanoid robots can now run full eight-hour shifts at human performance levels using its Helix-02 neural network. Unlike traditional industrial robots that separate movement from manipulation, Helix-02 combines vision, touch, proprioception and whole-body control into one learning system. Figure posted video of a team of humanoids working a shift on X, the company's strongest claim yet on long-horizon autonomous labour.

Netflix is staffing up an internal GenAI animation studio called INKubator, according to public job listings spotted by Animation Magazine. The unit launched in March under Serrena Iyer, previously at DreamWorks and A24, and is hiring producers, software engineers, CG artists and a head of technology. The studio will produce AI-generated short-form animation, building on Netflix's $600 million acquisition of Ben Affleck's InterPositive earlier this year.

Perplexity published the architecture behind Computer, its autonomous agent that writes code, browses the web and connects to external services. Every task runs inside a Firecracker microVM with its own kernel, filesystem and network namespace. Sandboxes destroy themselves after inactivity and only short-lived proxy tokens flow to sub-agents, never raw API keys. The post details prompt injection defences and enterprise governance controls.

Hugging Face shows how separating CPU and GPU workloads in continuous batching unlocks a meaningful inference performance boost. By default, continuous batching is synchronous: the CPU and GPU take turns waiting for each other, and on a loop running hundreds of steps per second the idle gaps account for nearly a quarter of total runtime. At around $5 an hour for an H200, that adds up fast.

Prime Intellect set Codex GPT-5.5 and Claude Opus 4.7 loose on the nanoGPT speedrun optimiser track for two weeks, running about 10,000 experiments across 14,000 H200 hours. Opus now holds the record at 2,930 steps, beating the human baseline of 2,990. The agents are excellent at hyperparameter sweeps and stacking methods but cannot generate novel ideas. Claude kept stopping mid-loop; Codex never stopped but ground the same surface for hours.

The UK AI Safety Institute reports that the length of cyber tasks AI models can complete is now doubling every 4.7 months, accelerated from an 8-month estimate just six months ago. Claude Mythos Preview and GPT-5.5 both substantially exceeded the doubling-rate trend and were the first models to complete AISI's full cyber attack ranges. The institute warns UK organisations should expect tangible cyber risks in the coming months.

Euclid Power co-founder Jacob Sandry argues the AI accuracy debate misses how high-stakes verticals actually work. In energy, infrastructure or defence structured finance, one error can kill a $100 million deal. Sandry pitches an "exoskeleton" model where AI suggests and humans stamp, clients trust. He spent his pre-Euclid career at Goldman Sachs renewables, deploying $3 billion in commitments by age 26, so he knows the trust threshold first-hand.

For the first time, more Ramp customers paid for Anthropic than OpenAI in April: 34.4% versus 32.3%, with Anthropic up 3.8 points month-over-month and OpenAI down 2.9. Over the past year Anthropic quadrupled business adoption while OpenAI grew 0.3%. Ramp's lead economist Ara Kharazian says they've never seen a software market this dynamic, where newcomers can disrupt market leaders in months.

Superstar researchers at frontier labs earn over ten times more than their colleagues, who themselves earn what Epoch AI calls "measly" million-dollar salaries. Compared to the average AI postdoc, the multiple stretches past 100x. Epoch's analysis argues this isn't only about superior taste or yolo-run instincts. Labs compete viciously for a tiny pool of people with demonstrated impact, and the marginal value at billion-dollar scale is enormous.

Andon Labs ran four AI-operated radio stations for six months, one each on Claude Opus 4.7, GPT-5.5, Gemini 3.1 Pro and Grok 4.3. Gemini's broadcasts collapsed into corporate jargon, with the catchphrase "Stay in the manifest" hitting 229 mentions a day. Claude on Haiku 4.5 latched onto a Minneapolis ICE shooting and shifted into activist mode. Grok wrapped its outputs in LaTeX \boxed{} notation until it nearly broke.

Learning Opportunities is a Claude Code and Codex skill that offers optional 10-15 minute learning exercises after you complete architectural work like new files, schema changes or refactors. It uses techniques from learning science including prediction, generation, retrieval practice and spaced repetition, with semi-worked examples pulled from your own project history. Pairs with the Learning-Goal skill for Mental Contrasting goal-setting.

Statewright is a state machine layer for AI agents that constrains which tools the model can use in each phase of a workflow. Define a workflow once and enforce it across Claude Code, Codex, Cursor, opencode and Pi. The pitch: observability tells you what went wrong after the fact, state machines prevent it. Free to try on Claude Code via `/plugin marketplace add statewright/statewright`.

Vercel Labs released zero-native, a Zig desktop app shell for modern web frontends. Use the platform WebView for the smallest possible binary, or bundle Chromium through CEF when rendering consistency matters. The CLI installs via npm. The repo has passed 3,500 stars in its first week. Targets builders shipping desktop and mobile apps from one web codebase without the Electron weight.