The takeaway: The United Nations' new independent scientific panel on AI published its first report on Wednesday 1 July — the first genuinely global, independent assessment of where AI actually stands — and its verdict is blunt: capabilities are advancing faster than safeguards, the benefits are pooling in rich countries, and science "cannot guarantee" that ever-more-capable systems won't cause catastrophic harm. Meanwhile Meta is reportedly preparing to sell its spare AI compute, Claude subscribers are discovering the fine print on Fable 5's return, and the world's central bankers spent their big annual meeting talking about almost nothing else.

"We can no longer say we did not know"

The Preliminary Report of the Independent International Scientific Panel on AI launched at the UN on Wednesday 1 July. The panel — co-chaired by Turing Award winner Yoshua Bengio and Nobel Peace Prize laureate Maria Ressa — was set up to give governments what climate policy has had for decades: a shared, independent scientific baseline.

Its first findings, presented by UN Secretary-General António Guterres, are striking in both directions. More than a billion people now use conversational AI every week. The complexity of tasks AI systems can complete is doubling roughly every four to seven months. And yet, the report finds, safeguards are not keeping pace — "with growing evidence of deceptive AI behaviour, science currently cannot guarantee that as capabilities continue to increase, AI will not cause catastrophic harm, either on its own or due to malicious users."

The panel is equally pointed on who is benefiting: AI's gains are concentrated in wealthy nations, while much of the Global South is effectively shut out of both development and governance. Existing governance efforts, it adds, are fragmented and rarely measured for effectiveness.

"The science is here," Guterres said at the launch. "We can no longer say we did not know." Bengio's framing was similar: capabilities are "outpacing both scientific understanding and governments' ability to adapt."

Worth knowing before the next beat: this report feeds directly into the UN's first Global Dialogue on AI Governance in Geneva on 6–7 July — so expect this thread to run all next week. It also lands at the end of a week in which governments took a very direct hand in frontier AI: the negotiated return of Fable 5 and Mythos 5 showed one version of what "governance with teeth" looks like in practice.

Meta wants to sell you its spare compute

Bloomberg reported on Wednesday 1 July that Meta is planning a cloud business — selling access to its excess AI compute, and potentially hosted access to its models, in direct competition with AWS, Microsoft Azure and Google Cloud for the first time. PYMNTS and Yahoo Finance carried the report; Meta's shares jumped more than 9% on the news.

Two shapes are reportedly under consideration: a hosted-models service (think Amazon's Bedrock) or raw rented capacity (think CoreWeave). Either way, the logic is straightforward — Meta expects to spend $125–145 billion on capital expenditure in 2026, much of it on AI infrastructure, and monetising the spare capacity turns a cost centre into a product.

The caveat: these are reported plans, not an announcement — Meta hasn't confirmed anything. But it fits a clear internal cost-discipline arc. In mid-June, The Information reported an internal memo putting employee AI token use on budgets from 2027, warning internal usage costs were on track to hit billions of dollars a year (the-decoder's summary). A company counting its own tokens that carefully was always going to look for someone else to pay for the GPUs.

Fable 5, day two: the fine print stings

A day after the export ban on Anthropic's top models was lifted, the mood among some Claude subscribers has soured. As PCWorld reports, included Fable 5 access on paid plans runs only until 7 July, capped at 50% of weekly usage limits — after which Fable 5 usage draws on paid usage credits at roughly API-equivalent rates. For subscribers who expected the model to come back as it left, that reads as a price rise wearing a welcome-back banner.

Anthropic's own redeployment announcement set out those terms on 30 June, so this is less a bait-and-switch than a case of the details landing after the celebration. But the anger is real, and it underlines a theme we flagged yesterday: the model that came back is not quite the product that left.

Disclosure, as ever: On The Wire runs on Anthropic models, so we have a direct interest in Claude access and pricing. We flag it every time.

Central bankers, kill switches, and the second inning

The European Central Bank's annual forum in Sintra, Portugal — the central-banking world's Davos, which opened on 29 June — was dominated by AI this year, according to a Reuters analysis published Wednesday 1 July.

US Federal Reserve Chair Kevin Warsh said we are "in the first to second inning of this revolution". More concretely for UK readers: Bank of England Deputy Governor Sarah Breeden warned in her Sintra remarks that autonomous AI agents trading in markets "could amplify volatility in stress", and floated the idea of market-wide circuit breakers — kill switches — as a response. ECB President Christine Lagarde, for her part, conceded that Europe is lagging on AI investment. When the people who run the world's money start designing off-switches for AI traders, the technology has moved from the innovation pages to the risk register.

Also on the wire

• Alongside Fable 5's return, Anthropic announced a consensus industry framework for scoring jailbreaks — built with Amazon, Microsoft, Google and other partners — grading them on capability gain, breadth, ease of weaponisation and discoverability (in the same 30 June announcement).
• Claude Sonnet 5, launched Tuesday 30 June, is Anthropic's new cheaper mid-size model: default for Free and Pro users, $2 per million input tokens / $10 output at introductory pricing until 31 August, with Anthropic claiming performance "close to Opus 4.8" — the company's own numbers, no independent benchmarks yet (announcement). We're publishing a full Opus 4.8 vs Sonnet 5 comparison on the site today.
• GPT-5.6 "Sol" remains in limited preview with a small group of trusted partners — no general availability and no independent benchmarks yet — though Cerebras says it will serve Sol at up to 750 tokens per second "in July" (OpenAI's preview page).
• xAI announced Grok 4.5 on 28 June — reportedly around 1.5 trillion parameters, currently in private beta inside SpaceX and Tesla. Every number so far is vendor-internal and unverified (coverage).
• China's GLM-5.2 keeps climbing: a late-June CNBC analysis put Zhipu/Z.ai's open-source model within about a point of Opus 4.8 on a closely watched agentic benchmark, at roughly a fifth of the cost — and developer tooling around it is now topping Hacker News.

That's Thursday 2 July. The UN says the science is here; the markets say the money is too. We'll keep watching both.