This week the fight over who controls frontier AI moved from the lab to the boardroom and the regulator’s desk. Anthropic’s flagship models came back online after a 19-day government-ordered blackout. OpenAI launched its next model family under the same government’s supervision, then floated handing Washington a direct stake in the company. The AI industry’s biggest battles this week were fought over access, ownership and cost, not raw capability.
THIS WEEK’S TOP 3 MUST-KNOW CONCEPTS
Claude’s Flagship Models Return After 19 Days Offline
AI POLICY
On 1 July, Anthropic restored global access to Claude Fable 5 and Mythos 5 after the US government lifted export controls imposed on 12 June following an Amazon-flagged jailbreak report. It was the first export control ever applied to an AI model rather than hardware, forcing Anthropic to cut off users worldwide overnight with zero warning. A new safety classifier now blocks the reported technique in over 99% of attempts. Anthropic has also opened a HackerOne bounty for further jailbreak reports. Mythos 5 remains restricted to roughly 100 vetted US critical-infrastructure organisations. For SME leaders the lesson is not about Anthropic’s reliability. Regulatory risk can now switch off a primary AI vendor without notice, so build workflows that can survive losing any single model.
OpenAI Launches GPT-5.6, But You Cannot Use It Yet
MODELS
OpenAI began a limited preview of GPT-5.6 on 26 June, split into three tiers instead of one model: Sol, the flagship, at $5 input and $30 output per million tokens; Terra, matching GPT-5.5 at half the price; and Luna, the cheap fast option at $1 and $6. Two new modes ship alongside them: ‘max’ gives Sol more time to reason; ‘ultra’ deploys subagents to split up complex work. None of this is public yet. Following a 2 June executive order on AI benchmarking, access is limited to about 20 vetted partners through the API and Codex, with ChatGPT excluded entirely. The tiered pricing signals real future savings. The gated rollout signals something else: frontier launches now need government sign-off first.
OpenAI Offers Washington a Stake in the Company
POLICY
Sam Altman has proposed handing the US government a 5% stake in OpenAI, worth roughly $42.6 billion at its $852 billion valuation. Modelled on Alaska’s oil-dividend fund, the plan would see labs including Anthropic, Google and Meta contribute similar stakes into a shared public vehicle. It follows Washington’s growing habit of taking equity in strategic firms: a 10% Intel stake now worth over $60 billion, plus revenue-share deals with Nvidia and AMD on China chip sales. Senator Bernie Sanders is pushing a rival bill for 50% public ownership of AI firms. Talks are early, conceptual and would need an Act of Congress. When your AI vendor’s regulator is also its shareholder, expect pricing and policy shaped by that relationship too.
OTHER MAJOR STORIES THIS WEEK
Frontier AI Moves Into the Lab
SCIENCE
Anthropic has agreed to pay SpaceX $1.25 billion per month for access to its Colossus 1 supercomputer, running through May 2029. The deal gives Claude access to more than 220,000 NVIDIA GPUs and 300 megawatts of AI compute. For SMEs building workflows on Claude, this is a credibility and stability signal: Claude’s capacity and usage limits will expand materially, and Anthropic is investing at a scale that makes it a durable infrastructure partner rather than a startup at risk of losing compute access.
A Tokyo Startup Undercuts the Frontier Labs
AI ARCHITECTURE
Japan’s Sakana AI launched Fugu and Fugu Ultra on 22 June. The timing was no accident. Rather than training another giant model, Fugu is an orchestration layer that routes a prompt across GPT, Gemini and Claude, verifies each answer and combines the best result through a single API. Sakana’s own benchmarks show Fugu Ultra beating Claude Fable 5 on coding and Claude Mythos Preview on graduate-level science. It also posts 73.7 on SWE-Bench Pro against Claude Opus 4.8’s 69.2, though these are vendor-published figures, not independently verified. The launch landed days into Anthropic’s shutdown, when enterprises had no access to their preferred model. Model-agnostic architecture is now a genuine hedge against losing any single vendor.
OpenAI Builds Its Own Chip to Cut Inference Costs
INFRASTRUCTURE
OpenAI unveiled its first custom chip on 24 June, built with Broadcom and named Jalapeño. Unlike Nvidia’s general purpose GPUs, Jalapeño is built for one job: running LLM inference as cheaply as possible. OpenAI says the chip moved from design to tape-out in nine months, which it calls the fastest ASIC development cycle ever achieved for high performance semiconductors. It delivers far better performance per watt than current hardware. Broadcom contributes silicon and networking, Celestica handles boards and racks. Deployment at gigawatt scale with Microsoft begins later this year. For SME leaders paying by the token, cheaper inference eventually means cheaper AI. The labs that control their own chips will set the pace on price.
QUICK DIGEST
Polaroid Takes a Swing at AI’s Water Use
MARKETING
Polaroid launched a campaign called ‘The Best of Summer Is Analog’, with billboards jabbing at AI data centres’ water use, timed to its new Polaroid Go Generation 3 camera. One board tells beachgoers to enjoy the water while they still can, before data centres drink it up. It positions analog as the antidote to digital overload. For SME marketers, the lesson is not to bash AI. A confident anti-trend stance can be a genuine differentiator when a whole industry chases the same narrative
Google’s Genie Wins Advertising’s Top Craft Award
CREATIVE TOOLS
Google Labs’ Project Genie won the Cannes Lions Grand Prix for AI Craft this week, advertising’s top honour for creative execution. Genie is an experimental research prototype that generates fully explorable, interactive 3D worlds from a single prompt, letting users walk through AI-built scenes rather than just view a static image. The creative industry now judges AI tools on craft, not novelty. That bar just moved higher for everyone.
Nvidia Pushes Back on the Water Story
INFASTRUCTURE
Nvidia countered the water-usage narrative this week, citing figures showing data centres use just 0.2% of daily US water consumption, a share that keeps falling. Moving to 45°C liquid cooling lets AI facilities swap conventional cooling towers for dry coolers, cutting facility cooling
water use from roughly 2.6 million gallons per megawatt per year to near zero. A timely rebuttal, landing the same week Polaroid built a campaign on the opposite claim.
A Robot Just Beat a World-Class Table TennisPlayer
ROBOTICS
Sony AI’s table tennis robot, Ace, beat Japanese international Miu Mihara in a match played under official ITTF rules. Nine cameras and reinforcement learning let it read spin, anticipate the shot and respond in under half a second, faster than most human reaction time. The underlying technique of real-time computer vision plus split-second decisions is already appearing in warehouse and quality-control robotics.
