Anthropic just shipped Claude Opus 5, a near-frontier model at half the price of its best one. Nvidia is weighing a $250 billion guarantee so OpenAI can lease a 10-gigawatt Ohio campus it cannot yet pay for. And 25 companies signed a letter asking Washington not to restrict open-weight models - with two big names missing. Here is today's Pulse, with the numbers, the sources, and what to do about each.

Anthropic released Claude Opus 5 on July 24 at $5 per million input tokens and $25 per million output - half the rate of Fable 5, its top model. It scored 30.16% on ARC-AGI-3, the highest result recorded on that benchmark, and 43.3% on FrontierBench against GPT-5.6 Sol's 37.5%. A fast mode runs at roughly 2.5 times default speed for double the base price, and Anthropic published a dedicated prompting guide alongside the launch.

Why it matters: The price of near-frontier intelligence just halved, and that changes the maths on any workflow you priced out earlier this year. Coding agents, research pipelines and document-heavy work that only penciled out on cheaper, weaker models now fit a frontier-class budget. It also puts pressure on every rival still charging flagship rates.

What to do:

Source: Anthropic

Nvidia, Microsoft, Meta, IBM and 21 other signatories sent a letter on July 24 warning Washington against "premature restrictions" on open-weight AI models, as the administration weighs a ban on advanced Chinese models. OpenAI added its name without announcing it; Anthropic and Google are absent. The letter argues open weights drive competition, customisation and cyber defence, and that curbs would push development offshore.

Why it matters: Open-weight Chinese models like Kimi K3 now compete at frontier level, and this letter is about which of them your business will be allowed to build on. The split in who signed is the tell: companies selling closed subscriptions have different incentives from companies selling chips and cloud capacity. A White House framework is expected before August 1, so the answer lands within days.

What to do:

Source: CNBC

Nvidia is in talks to backstop up to $250 billion so OpenAI can lease SoftBank's 10-gigawatt data center campus in Piketon, Ohio - a site expected to cost more than $500 billion, built on a former uranium enrichment plant. A separate deal under discussion would see Nvidia finance around $350 billion of chips for the same site. Reports landed July 27, two days before Nvidia's biggest customers report earnings.

Why it matters: This is the chip seller guaranteeing the debts of its own biggest buyer, so the same dollars keep circulating between supplier and customer. That structure holds while AI demand grows and gets fragile fast if it slows - and there is now enough of this circular financing stacked across the industry that a wobble at one company travels. If your business plans assume cheap, abundant AI compute forever, this is the financing that assumption rests on.

What to do:

Source: CNBC

Bill Oliver, a Progressive Conservative member of New Brunswick's legislature, recited his AI assistant's leftover editing note mid-speech: "Here's a more natural, flowing version of that section that reads like a legislative speech rather than a series of short points." The clip spread across Reddit in late July, with some commenters calling for his resignation.

Why it's interesting: The AI did the researching, drafting and polishing - and the one step it could not automate, actually reading the result before saying it out loud, is the one that failed.

Key takeaway: Read the whole draft before it leaves your hands - leftover prompts are the easiest tell of unedited AI text.

Source: 1News

"Claude of Duty," a browser FPS published by Matt Shumer, spans roughly 55,000 lines of code across 11 subsystems - rendering, physics, weapons, enemy AI, audio - with every texture and sound generated procedurally in code. Its only runtime dependency is Three.js, and the whole thing is open source under MIT. Peer reviewers scored it 5.05 out of 10 against actual Call of Duty standards.

Why it's interesting: A 5/10 sounds unimpressive until you remember the comparison point is a franchise built by hundreds of people over two decades, and this came from orchestrated AI agents working against a single architecture spec.

Key takeaway: Complex interactive software is now a prompt-and-orchestration problem - clone the repo and see for yourself.

Source: GitHub

Nvidia open-sourced the first GPU-accelerated medical physics simulation framework, letting surgical robots train across 8,192 parallel simulations at once. Training runs that took over 5 hours now finish in under 2 minutes, and the whole framework is free inside Isaac for Healthcare.

Why it's interesting: Simulation speed is the real bottleneck in robotics - collapse it by two orders of magnitude and skills that took months of physical trial-and-error become an overnight batch job.

Key takeaway: Watch surgical and medical robotics move faster than the hardware suggests - the learning loop just got 150 times shorter.

Source: Nvidia

Context: Claude Opus 5 just halved the price of near-frontier AI. Before you switch anything, run a controlled bake-off on your own work instead of trusting benchmark charts.

Prompt: I currently use [current model] for [specific task, e.g. drafting client proposals]. I'm evaluating [new model] as a replacement. Design a head-to-head test: (1) pick 3 real examples of this task from my workflow, [paste or describe them], (2) define 3-5 scoring criteria that matter for this task, with a 1-5 scale for each, (3) give me an identical prompt to run on both models, (4) build a simple scoring table I can fill in, and (5) tell me what score gap would justify switching once I factor in a price of [old price] versus [new price].

Do not: Do not let the test use toy examples or generic samples - it must use real inputs from your actual workflow, including your messiest one.

If/Then: If the cheaper model wins or ties on your top two criteria, then switch your default and keep the expensive model only for the specific tasks where it clearly won.

Example: A consultancy founder named Daniel ran this bake-off on proposal drafting before moving from his premium model to Opus 5. The scores tied on structure and tone, so he switched his default and cut his monthly AI bill roughly in half without a quality drop.

Microsoft released public previews of MAI-Image-2.5-Pro and MAI-Voice-2-Flash, its highest-fidelity image model and a voice model that is twice as fast and 32% cheaper than its predecessor. The company says the new models cut costs by up to 89% compared with OpenAI equivalents, and they roll out across Foundry, Bing, PowerPoint and Azure.

Why it's important: Microsoft is methodically replacing its most important AI supplier with its own models, one modality at a time, after the two loosened their exclusivity in April. When the vendor with the deepest OpenAI ties builds its exit ramp, every buyer should at least know where theirs is.

Business takeaway: Price your AI stack against Microsoft's in-house models - an 89% claimed gap is worth an afternoon of testing.

Source: Microsoft AI

ChatGPT Work can now sign into a user's sites and keep that login across sessions - the user enters credentials, the agent resumes, and passwords stay hidden from the model. Saved logins can be cleared at any time, and agents can now finish work on platforms behind login walls such as Jira and Notion.

Why it's important: Login walls were the practical ceiling on what AI agents could actually finish. Persistent authentication turns agent demos into end-to-end workflows - and widens how much of your stack an agent can reach without anyone re-approving it, which is a security question as much as a productivity one.

Business takeaway: Decide which systems an agent may hold logins for before your team decides it for you.

Source: OpenAI Devs

From August 2, Article 50 of the EU AI Act requires businesses to tell users when they are interacting with an AI system, and to label AI-generated content and deepfakes. The obligations apply to companies serving EU users regardless of where the company sits.

Why it's important: This is the first broad AI transparency mandate with real enforcement behind it, and it arrives in five days. If your chatbot, support flow or content pipeline touches EU customers, "we didn't label it" stops being a style choice and becomes a compliance gap.

Business takeaway: Audit every customer-facing AI touchpoint this week for disclosure and labeling before August 2.

Source: EU AI Act, Article 50

That's it for today's Daily Pulse. Forward this to one person who wants to stay ahead of AI. See you in the next one. - Nicolas