The AI Brief

OpenAI launched GPT-5.6 Sol, Terra, and Luna to the global public on July 9, following White House clearance. The Department of Commerce's Center for AI Standards and Innovation (CASI) conducted additional tests after the June 26 preview, with OpenAI sending technical staff to Washington to address questions in real time. Sam Altman confirmed the launch on X the evening of July 8: "GPT-5.6 Sol launches Thursday! Happy building."

The three-tier structure: Sol is OpenAI's strongest model to date, targeting complex reasoning, coding, biology, and cybersecurity at $5/$30 per million tokens (standard) and $12.50/$75 per million in fast mode via Cerebras at up to 750 tokens per second. Terra is priced at $2.50/$15 and delivers performance comparable to GPT-5.5 at half the cost. Luna is the cheapest at $1/$6. GPT-5.6 also introduces a new "ultra" mode for Sol that deploys subagents to parallelize complex tasks. On Terminal-Bench 2.1, Sol scores 88.8% and Sol Ultra reaches 91.9%, both above Claude Fable 5's 88.0% on the same benchmark. On ExploitBench, Sol matches Mythos Preview using roughly one-third the output tokens. OpenAI also introduced explicit cache breakpoints and a 30-minute minimum cache life to improve cost predictability for agentic workflows.

The government review ended in roughly 13 days, well under EO 14409's 30-day window, after Commerce signed off. OpenAI has been explicit that it views the current review mechanism as a transitional arrangement, not a permanent default. First covered in Vol. I, No. 44 (July 8, 2026).

SpaceXAI's launch page reports that Grok 4.5 resolves SWE-Bench Pro tasks using an average of 15,954 output tokens, compared with 67,020 for Claude Opus 4.8 at max settings, a 4.2× gap. At Grok 4.5's $6 per million output tokens versus Opus 4.8's $25, the combined pricing-and-efficiency math is the product's main commercial argument. Artificial Analysis independently placed Grok 4.5 at 54 on its Intelligence Index on launch day, ranking it fourth overall behind Fable 5 (60), Opus 4.8 (56), and GPT-5.5 (55), suggesting the model is in but not leading the frontier tier.

The benchmark picture from SpaceXAI's own data is split: Grok 4.5 leads on SWE Marathon (29.0% vs. Fable 5's 24.0% and Opus 4.8's 26.0%) and is near-tied with GPT-5.5 on Terminal-Bench 2.1 (83.3% vs. 83.4%), but trails meaningfully on DeepSWE 1.1 (53% vs. Fable 5's 70% and GPT-5.5's 67%) and on SWE-Bench Pro resolve rate (64.7% vs. Opus 4.8's 69.2% and Fable 5's 80.4%). The token-efficiency figure is vendor-stated on the SWE-Bench Pro task set; no independent third party has reproduced it as of publication.

Caveat Token efficiency figure sourced from SpaceXAI's own launch benchmarks; no independent verification as of publication date.

The European Commission published its Action Plan on Cybersecurity and Artificial Intelligence on July 7, 2026, targeting three interlocking objectives: promoting safe use of advanced AI, protecting EU digital infrastructure against AI-enabled attacks, and building sovereign European AI capacity for cybersecurity. The plan builds on the AI Act, Cyber Resilience Act, NIS2 Directive, DORA, and the Cyber Solidarity Act rather than introducing new legislation.

The two most operationally significant elements: First, the Commission will build an EU evaluation capacity for AI models by 2027, third-party assessment before frontier models reach the EU market, complementing the AI Office's regulatory function. Second, ENISA (the EU Agency for Cybersecurity) will develop a structured-access blueprint defining the conditions under which European public and private organizations can reach the most advanced AI systems for defensive cybersecurity work. A Joint Research Centre testing platform for AI in critical sectors, energy, transport, health, finance, and public administration, is also planned for launch by end of 2026. Executive VP Henna Virkkunen led the announcement, saying advanced models "can now build cyber exploits in minutes or hours at a fraction of the cost of vulnerability discovery by trained humans."

The honest constraint: the plan is a policy document, not a rulebook. It attaches no funding figures to the evaluation capacity initiative, gives no hard timeline for when the JRC platform will be operational, and does not name which model providers will accept EU structured access on European terms. Leading AI labs including OpenAI and Anthropic have historically preferred evaluation relationships with the UK AI Security Institute, which holds no regulatory authority. Whether Brussels can compel meaningful pre-market access under existing AI Act authority, without a new regulation, remains the open legal question.

SpaceXAI released Grok 4.5 on July 8, 2026, the first public model from its 1.5-trillion-parameter V9 foundation, roughly 3× the parameter count of the v8-small architecture behind Grok 4.3. The model was trained on tens of thousands of NVIDIA GB300 GPUs at the Colossus cluster in Memphis, with Cursor developer session data incorporated in supplemental training. xAI engineers have acknowledged that supplemental inclusion is "not quite as good as having it in initial training," and say the next V9-generation model will bake Cursor data in from the start of pre-training. Grok 4.5 runs at approximately 80 tokens per second and is live in Grok Build, in Cursor on all plans, and via the SpaceXAI API console.

Musk's positioning, "an Opus-class model, but faster, more token-efficient and lower cost", is a defensible label on four vendor-published benchmarks where Grok 4.5 splits roughly evenly against Opus 4.8: it wins on SWE Marathon (29.0% vs. 26.0%) and Terminal-Bench 2.1 (83.3% vs. 78.9%), and loses on DeepSWE 1.1 (53% vs. 59%) and SWE-Bench Pro resolve rate (64.7% vs. 69.2%). Artificial Analysis independently placed Grok 4.5 at 54 on its Intelligence Index, ranked fourth. Musk's own clarification on X puts it more precisely: "our internal assessment is that Grok 4.5 is roughly comparable to Opus 4.7, but much faster." The combination of lower per-token price and claimed 4.2× output-token efficiency on SWE-Bench Pro tasks is the commercial argument, not the benchmark ceiling. EU availability is expected mid-July.

The strategic context: Grok 4.5 is the first visible product of the February 2026 SpaceX–xAI merger and the $60B Cursor acquisition announced in June, which has not yet closed. SpaceXAI plans to ship newly pre-trained foundation models monthly through end of 2026, starting from this release. The compute the model trained on is shared with competitors, SpaceXAI currently leases Colossus capacity to Anthropic and Google, a tension Axios noted will force a capacity allocation choice as SpaceXAI's own model compute needs grow.

The FTC published a proposed policy statement on July 1, 2026, arguing that AI companies that steer their systems' outputs toward undisclosed objectives, whether for profit, opinion-shaping, or state-law compliance, may be engaging in deceptive practices under Section 5 of the FTC Act. The statement was issued pursuant to Executive Order 14365, which directed the FTC to clarify how Section 5 applies to AI models and, specifically, to explain when state laws requiring alterations to accurate AI outputs are federally preempted. The Commission vote was 2–0. Public comment closes July 31.

The preemption argument is the statement's most consequential and most contested element. The FTC asserts that state law is "impliedly preempted to the extent it conflicts with a federal regulatory scheme," specifically citing Colorado's Artificial Intelligence Act as an example of a law that could pressure AI companies to suppress output accuracy in order to avoid disparate-impact liability. Legal analysts at TechFreedom and Stanford Law's CodeX have noted that the Supreme Court applies a "presumption against preemption" and that Section 5's general language was "deliberately framed in general terms", meaning conflict preemption under Section 5 would likely require the FTC to complete a full formal rulemaking under the Administrative Procedure Act, not a policy statement. A policy statement does not carry the force of law.

The practical disclosure safe harbor: the statement says AI companies can avoid Section 5 liability by making "clear, conspicuous, and adequate disclosures" that a model is configured to prioritize objectives different from what users requested. Companies subject to state AI fairness mandates who want to comply without triggering federal deception claims will need documented disclosure architecture before the final statement, whenever issued, takes effect. The comment period is the window to engage on the definitions of "adequate disclosure," "expected objectives," and the scope of the preemption analysis.