Editor's note: we lead with regulation today only because the news cycle is genuinely dominated by a single dated event — the EU AI Act's general-purpose-AI enforcement powers and its content-transparency rules both take effect today, 2 August. On an ordinary day this bucket earns a quick hit, not the top slot.
The EU AI Act Grows Teeth Today — and Starts Labelling the Machines
From today, 2 August 2026, the European Commission's AI Office and national authorities can actually enforce the AI Act's rules for general-purpose AI models, one year to the day after those obligations first applied. The grace period was the point: providers had until now to get compliant, and from this morning the AI Office can compel technical documentation, demand access to a model to run its own evaluations, order risk-mitigation, market-restriction or withdrawal, and levy fines of up to €15 million or 3% of global turnover under Article 101 — not the €35M/7% headline cap, which sits in Article 99 for prohibited-practice breaches. Models placed on the market before 2 August 2025 get until 2 August 2027 to comply; anything newer is already exposed.
The quieter half of today's deadline may matter more to how AI feels to ordinary users: the Article 50 transparency rules also start to apply. Chatbots and other interactive systems must now tell people they are talking to a machine, deepfakes must be labelled, and AI-generated or altered content has to carry machine-readable marks so it can be detected downstream. To operationalise this the Commission has published a first list of more than 180 organisations that have signed a voluntary Code of Practice on transparency of AI-generated content. For a platform that has to answer the question “who can be compelled, in which jurisdiction, over which data,” today is the moment that stops being a slide in a compliance deck and becomes an operational fact.
OpenAI Cuts Luna 80% — the Token Price War Is Now the Headline
Three weeks after the GPT-5.6 family launched, OpenAI has already recut its price list. On 30 July it dropped its cheapest tier, GPT-5.6 Luna, by 80% — from $1/$6 to $0.20 per million input and $1.20 per million output tokens — and trimmed the mid-tier Terra 20% to $2/$12. The top Sol tier stays at $5/$30, now paired with a new “Fast Mode” that replaces Priority Processing. VentureBeat read the move plainly: competition among frontier labs has shifted from capability to cost, and pricing power is eroding fast enough that a three-week-old price sheet is already stale.
The backdrop is a market consolidating even as it discounts. Forbes' AI 50, published in April, put the cohort's cumulative venture funding at $305.6 billion, of which OpenAI and Anthropic account for roughly $242.6 billion — about 80 cents of every dollar raised. Capability is converging and per-token prices are collapsing, but the capital and the leverage are pooling into two US-jurisdiction labs. For a European buyer, that combination sharpens rather than softens the real question: when the cheap, capable model you built on can change its price, its terms, or its availability on three weeks' notice, the durable variables are portability, data location, and who holds the contract.
Compliance Day Makes the Open-Weight Case, Not the API One
The two threads above meet on the same ground. If the frontier is converging and getting cheaper, and the binding regulatory question is jurisdiction, then the model you can hold — weights on your own hardware, on EU soil — is worth more than a marginally higher benchmark score you rent. That is the through-line of this summer's open-weight run: DeepSeek's V4-Flash-0731 shipped self-hostable under MIT last week, and Moonshot's Kimi K3 remains the strongest open model on the Artificial Analysis Intelligence Index at around 57, close enough to the frontier that the gap is a serving problem, not a capability one.
The next test of that thesis is imminent. Zhipu's GLM-5.5 is widely expected in August as an open-weight, trillion-parameter successor to GLM-5.2, though the company has not confirmed a date and its channels still promote the current model — the announcement-versus-artifact caveat that has caught several of this year's “imminent” open releases. The honest caveat cuts the other way too: these are serious models that need real multi-GPU nodes to serve, so “self-hostable” is a Q4-2026 engineering task for most teams, not a weekend one. But the direction of travel is clear. On the day the EU AI Act starts labelling and compelling, the models that answer to your own jurisdiction are the ones you don't have to renegotiate.
Quick Hits
- The Transparency Code has broad backing — The Commission's first list of signatories to the AI-content transparency code runs past 180 organisations (about 190 by end-July), spanning the big labs (OpenAI, Google, Meta, Anthropic, Mistral, Aleph Alpha, Black Forest Labs, Cohere, Synthesia) and deployers (Getty Images, Lufthansa, Lenovo, Iberdrola); roughly half are SMEs.
- Forbes AI 50 shows a two-horse race at the top — The April AI 50 counted 20 newcomers, but OpenAI ($182.6B raised, >$25B run-rate revenue) and Anthropic ($60B raised, >$30B run rate) still dominate the capital table.
- Mistral's “fat but sparse” MoE is still partner-only — The open-weight Mixture-of-Experts CEO Arthur Mensch teased in July remains in early access with no params, benchmarks, or licence published; a broader release is still expected “later this summer.”
- Gemini 3.5 Pro slips again — Google's next Pro model reportedly continues to stumble on coding evals with no fresh timeline, even as Gemini 4 pre-training is said to have begun.