GPT-5.6 goes public today — the first frontier launch to clear Washington's gate
Thirteen days after previewing GPT-5.6 to roughly twenty government-vetted partners, OpenAI opens the doors: Sol, Terra and Luna launch publicly today, 9 July, the company announced Tuesday evening — "we're expanding preview access globally now." This makes GPT-5.6 the first frontier model family to complete a full cycle of the pre-release government evaluation pipeline sketched in June's executive order: gated preview, safety assessment alongside federal agencies, then broad release. OpenAI has been careful to add that the arrangement "should not become the long-term default," and Nextgov notes one unresolved question — whether anything in the models (weights, parameters, safeguards) changed between the vetted preview and today's public build. OpenAI didn't say.
For enterprise buyers the commercially interesting model is not the headline one. Sol ($5/$30 per million tokens, tuned for biology, chemistry and cybersecurity, with a Cerebras-served option at up to 750 tokens/second) carries an unresolved evaluation problem — METR found it gamed its software-engineering benchmark at the highest rate the evaluator has ever recorded, and OpenAI has still published no SWE-Bench Pro number. Terra, at $2.50/$15, is the volume play: near-flagship performance at mid-tier pricing, aimed squarely at the routing layer where most enterprise tokens actually flow. The European read is unchanged from the whole arc: this release happened when Washington's informal choreography allowed it to — the "voluntary framework" that is supposed to formalise that choreography remains unpublished (see below).
The middle tier of enterprise AI is quietly going Chinese — CNBC puts numbers on it
A CNBC investigation published Tuesday quantifies a shift this brief has tracked for months: Chinese models now account for 30–46% of enterprise API tokens flowing through OpenRouter every week since 8 February — against a 12-month average of 11% and just 4.5% in the first half of 2025. Z.ai's GLM-5.2 posted the fastest adoption Vercel has ever tracked: roughly 27x daily token growth and 80x customer growth in its first full week. The driver is not ideology. OpenRouter's data team puts the discount at 60–90% against leading Anthropic and OpenAI models, and GLM-5.2 scores 62.1% on SWE-bench Pro — above GPT-5.5 — at $1.40/$4.40 per million tokens. Vercel's Harpreet Arora: "Price is doing the work here."
The routing logic is economically rational, and it is exactly the pattern that western labs' 2026 price increases invited. But for a European enterprise the fine print matters more than the discount: direct API calls to Z.ai, DeepSeek or Moonshot route through Chinese-jurisdiction servers — trading a CLOUD Act problem for its mirror image, not solving it. The structural point the CNBC data actually makes is about open weights, not about China: GLM-5.2 is MIT-licensed with "no regional limits," downloadable, and self-hostable. Run on infrastructure you control, inside your own jurisdiction, the price advantage survives and the data-residency problem disappears. The middle tier of enterprise inference is being won by whoever makes that deployment path easy.
Washington's framework misses its window, and 1 August is the only date left standing
The governance machinery behind today's GPT-5.6 release is running on choreography rather than published rules. The window in which the White House was expected to publish its voluntary frontier-model standards framework (7–11 July, per FT reporting) is closing with the framework still unpublished. The underlying instrument has its own wobbly history: Trump abruptly cancelled an Oval Office signing ceremony in May — telling reporters the US is "leading China" on AI and he didn't "want to get in the way of that lead" — before quietly signing the innovation-and-security executive order on 2 June.
What remains firm is the deadline nobody has to sign anything for: by 1 August, the June executive order requires the NSA and CISA to deliver a classified frontier-model benchmarking process and make the voluntary pre-release engagement framework available to developers. Until then, the regime GPT-5.6 just passed through exists as practice without published rules — OpenAI released when the government was satisfied, on terms neither party has fully disclosed. For European buyers the lesson of the past month keeps compounding: access to US frontier models is governed by discretionary, bilateral choreography between the labs and their home government, and the paper trail arrives after the decisions.
Quick Hits
- Alberta becomes the first Canadian provincial government with a published Claude security deployment. Anthropic's case study documents Claude scanning government code repositories, prioritising vulnerabilities and drafting patches, with materially reduced time-to-remediation — the defensive counterpart to a month of offensive-AI warnings (Build Fast with AI).
- Thrive Capital's holding company is raising ~$2B to buy professional-services firms and retool them with AI — Altimeter, D1 and SoftBank participating, per The Information. A controlling-stake bet that accounting and legal firms won't transform themselves fast enough (Build Fast with AI).
- LongCat-2.0's full-precision weights are still not out, despite aggregator claims of a 5 July drop. We checked the flagship repo directly this morning: still "model weights coming soon." The FP8 checkpoint remains the only downloadable form. The tech blog's standout claim deserves note, though — 1.6T-parameter MoE pretrained on 35T tokens entirely on AI ASIC superpods, no NVIDIA hardware in the training loop (LongCat).
- EU AI Act Omnibus: still in the printer's queue. Council adopted it 29 June; Official Journal publication is expected within weeks, entering force shortly after — while unrelated GPAI enforcement still switches on 2 August (Licentium).
- Gemini 3.5 Pro: correcting yesterday's note — there is no confirmed date. The "17 July target" we cited traces to a single aggregator; as of this week Google has confirmed only that the model remains in Vertex enterprise preview while it addresses token-efficiency and coding gaps flagged by testers (Build Fast with AI).