The Daily Brief · Issue 2 · 17 August 2026SpaceX completes its Cursor takeover as Stripe agrees to buy OpenRouter for more than $7 billion
A weekend of big deals: the most popular AI code editor is now owned by SpaceX, and Stripe is buying the leading AI model router. Z.ai and Alibaba shipped strong new open models, and OpenAI lost more senior leaders before its IPO.
AcquisitionCursor is now officially part of SpaceX
Cursor, the AI code editor, said on 14 August that SpaceX has completed its acquisition. The deal finishes a process that began in April, when Cursor announced a model-training partnership with SpaceXAI, the Elon Musk-owned lab behind Grok. Cursor says the main gain is compute: it will have access to what it calls the world's largest GPU fleet, so it can train stronger models that cost less to run. Cursor points to Grok 4.6, released two days earlier and available in Cursor on day one, as an early result of the tie-up. Financial terms were not given in the post.
Why it matters: The leading independent AI coding tool is now owned by a model maker, so users should watch whether Grok gets favoured over Claude and GPT models inside Cursor.
Source: Cursor
M&AStripe agrees to buy OpenRouter for more than $7 billion, report says
Stripe has finalised a deal to buy OpenRouter for over $7 billion, according to Bloomberg, as reported by TechCrunch on 16 August. OpenRouter is a 'router': one account and one API that lets developers send requests to hundreds of different AI models and switch between them by price or quality. In May it raised a $113 million Series B at a reported $1.3 billion valuation, so the price is more than five times that figure. OpenRouter claimed 8 million users and access to more than 400 models at the time. Neither company had confirmed the price when TechCrunch published.
Why it matters: The biggest neutral marketplace for AI models would sit inside a payments company, which could change how developers pay for and switch between models.
Source: TechCrunch
Model launchZ.ai releases GLM-5.3, with big coding gains and strong hacking skills
Chinese lab Z.ai released GLM-5.3. It uses the same base model as GLM-5.2; every improvement came from extra post-training (further training after the main run). Z.ai reports Terminal-Bench 3.0 rising from 4.6 to 28.3 and DeepSWE 1.1 from 46.2 to 66.9. On the CyberGym security test it scores 84.5%, which Z.ai says is the best result, just ahead of Claude Mythos 5 (83.8%) and GPT-5.6 Sol (83.6%). Z.ai says the model found 2,436 real vulnerabilities across 269 software projects with partner security teams. Thinking can no longer be switched off. All figures are vendor-reported.
Why it matters: A Chinese model now matches US frontier models on at least one hacking benchmark, which adds to the cyber-risk debate that paused OpenAI's Astra.
Source: Z.ai
Open weightsQwen 3.8 27B brings near-flagship quality to a 17 GB laptop download
Alibaba's Qwen lab released Qwen 3.8 27B, a 27-billion-parameter model that can read images, under the open Apache 2.0 licence. Qwen's own benchmarks show it beating both Qwen 3.6 27B and the closed Qwen 3.7-Plus. Developer Simon Willison ran a 17 GB compressed version on a MacBook Pro and an NVIDIA DGX Spark and called it excellent. His main warning: it defaults to the highest reasoning setting ('xhigh'), so one simple drawing task used 22,276 thinking tokens and took 21 minutes. He recommends starting with low or no reasoning. Independent benchmark results were still pending.
Why it matters: Strong models that run offline on one laptop keep getting better, which matters for privacy and for cutting API bills.
Source: Simon Willison
PeopleOpenAI pre-IPO shake-up: revenue chief, former COO and ethics head all leave
OpenAI chief revenue officer Denise Dresser is leaving less than a year into the job, Axios reports, and Dali Rajic, most recently president and COO of Google-owned Wiz, replaces her. The same week, long-time executive and former COO Brad Lightcap announced his departure. The Financial Times reported that head of ethics Chloé Bakalar has also left, following the head of safety systems, Johannes Heidecke, and chief futurist Joshua Achiam. Fidji Simo, Sam Altman's second-in-command, left in July but stays on as an adviser. Co-founder Greg Brockman is taking a bigger role across the company ahead of an expected IPO.
Why it matters: Losing enterprise and safety leaders right before an IPO raises questions about OpenAI's push to catch Anthropic in business sales.
Source: Axios
DealsOpenAI paid about $100 for a 4.2% Cerebras stake before Ultrafast
A Cerebras filing shows OpenAI exercised warrants in July for 10,033,508 non-voting shares at $0.00001 each, about $100 in total. That is a 4.2% stake, worth roughly $2.3 billion at Cerebras's share price on 13 August, the day OpenAI previewed its Cerebras-powered Ultrafast tier for GPT-5.6 Sol. The warrants came with a deal in which OpenAI committed to buy 750 megawatts of Cerebras inference capacity through 2028. Implicator notes that all the speed comparisons in the Ultrafast launch came from Cerebras, and no independent quality tests or prices were published.
Why it matters: OpenAI profits from Cerebras's success, so buyers should wait for independent tests before trusting Ultrafast speed claims.
Source: Implicator.ai