The Brief: Faster Models, Firmer Rules
OpenAI's fastest model yet ships into a release pipeline that Washington and Brussels are moving, from opposite directions, to gate.
GPT-5.6 Sol goes to Cerebras at 750 tokens per second1
The Sol model we previewed on 7 July is now shipping on Cerebras wafer-scale hardware at up to 750 tokens per second — roughly an order of magnitude faster than the 40–120 tokens per second a GPU cluster streams from a frontier-class model. It was previewed on 26 June under the White House's voluntary pre-deployment review, and initial access is limited while Cerebras scales capacity.
Why it matters → Speed is quietly becoming its own capability axis: at 750 tokens per second an agent can take ten times as many reasoning steps inside the same wall-clock budget, which changes what interactive, tool-using workflows can attempt rather than just how quickly they answer.
Together AI raises $800M to bet on open-weight infrastructure2
Together AI, which builds serving infrastructure for open-source models, announced an $800 million Series C on 1 July, framing the round around a claim that "the economics of closed models don't scale."
Why it matters → It is the largest single wager yet that the durable margin in AI sits in running open-weight models cheaply for enterprises, not in owning the frontier weights — the opposite side of the trade from the labs racing Sol to market.
TSMC revenue holds up 30% as the AI build refuses to cool3
TSMC's most recent monthly disclosure put May 2026 net revenue at NT$416,975 million, up 30.1% year over year and 1.5% on April; year-to-date revenue through May reached NT$1,961,804 million, also up 30.0%.
Why it matters → The foundry that fabricates most of the world's AI accelerators is the cleanest demand signal there is, and a 30% run-rate with no sequential slowdown says the capex boom the hyperscalers keep guiding to is still landing as real silicon orders.
The EU AI Act's enforcement teeth arrive on 2 August4
From 2 August 2026 the European Commission gains its supervisory powers over general-purpose AI model providers — the authority to demand documentation, run evaluations, order mitigations, and levy fines — alongside the Act's Article 50 transparency obligations.
Why it matters → Until now the GPAI rules have been duties on paper; from next month they are duties a regulator can enforce, with extraterritorial reach over any model placed on the EU market — the first hard, binding regime any frontier lab has to clear.
Washington's voluntary release standards take shape5
The pre-release framework created by the June executive order we covered on 8 July is being finalised with the leading US labs. Its core is a classified benchmarking process, run by the Center for AI Standards and Innovation and the NSA, to designate "covered frontier models" and offer government up to 30 days of pre-release access.
Why it matters → It is the American answer to the EU Act, and a study in contrasts: voluntary rather than statutory, national-security framed rather than rights framed. GPT-5.6 Sol is the first model visibly moving through it.
"Every Eval Ever" tries to make benchmark results comparable again6
A paper posted 12 June proposes a single source-agnostic JSON schema for evaluation results and ships a community database on Hugging Face spanning 22,235 models, 2,273 unique benchmarks, and 31 evaluation formats.
Why it matters → As frontier claims fragment across HLE, ARC-AGI-2, FrontierMath and a dozen bespoke leaderboards, results scattered in incompatible formats can't be compared or reproduced; a shared schema is unglamorous plumbing that eval science needs before any headline score can be trusted.
Sources
- OpenAI — "Previewing GPT-5.6 Sol: a next-generation model"
- Together AI — "Announcing our Series C" ($800M)
- TSMC Investor Relations — 2026 monthly revenue disclosures
- European Commission — Regulatory framework for AI (the AI Act)
- The White House — Executive Order, "Promoting Advanced Artificial Intelligence Innovation and Security" (June 2026)
- arXiv:2606.14516 — "Every Eval Ever: A Unifying Schema and Community Repository for AI Evaluation Results"