ALETH / AI-BRIEF / 2026-04-25 / SPACEX TAKES A $60BN OPTION ON CURSORThe Aleth Briefs link to the sources behind the stories and show how the week unfolded.
The week in five lines:
SpaceX took a $60bn option to buy Cursor.
China moved to block US investment in leading AI startups.
DeepSeek V4 closed the gap with frontier models at a fraction of the cost.
Canada’s Cohere agreed to buy Germany’s Aleph Alpha in a sovereign-AI deal.
Browse by day:
Weekend: NSA Mythos · Beijing robots
Monday: Kimi K2.6 · Amazon-Anthropic
Tuesday: Facial recognition · Florida · Firefox · Mythos breach · SpaceX-Cursor
Wednesday: Google TPUs/agents · OpenAI agents
Thursday: Distillation memo · Meta job cuts · GPT-5.5 · CAISI · Tesla
Friday: DeepSeek · Cohere-Aleph · China · Anthropic · DOJ-xAI · Altman · Meta
Weekend (18-19 April)
The NSA is using Claude Mythos Preview even though the Pentagon, which oversees it, calls Anthropic a supply chain risk, Axios reported.
One source said the model is also used more widely in the Defense Department. Anthropic has limited Mythos to about 40 organisations and named only 12, and the NSA is among the unnamed.
The White House is also seeking access, the FT reported. Dario Amodei met chief of staff Susie Wiles and Treasury Secretary Scott Bessent on Friday to discuss its use, and both sides called the meeting productive, Axios reported.
NSA using Anthropic’s Mythos despite Defense Department blacklist · Axios
Anthropic CEO met White House chief of staff on Mythos access · Financial Times
A humanoid robot won Beijing’s second robot half-marathon on Sunday in 50 minutes 26 seconds, inside the human world record.
More than 100 teams ran, up from 20 last year, when the winner needed 2 hours 40 minutes. Several robots finished >10 minutes ahead of the human winners. The winner, Lightning, was built by smartphone maker Honor, Caixin reported.
Champion Crowned for Robot Half Marathon · Beijing government
Chinese-Built Robot Wins Beijing Half-Marathon · Caixin Global
Monday 20 April
Moonshot AI released Kimi K2.6 as open weights, and Artificial Analysis ranked it the leading open-weights model.
Moonshot says K2.6 runs swarms of 300 sub-agents across 4,000 steps, up from 100 in Kimi K2.5, with gains in long-horizon coding. It is a 1T-parameter mixture-of-experts (MoE) model with 32B active parameters, and scores 54 on the Artificial Analysis Intelligence Index, behind top Anthropic, Google and OpenAI models at 57.
Kimi K2.6: Advancing Open-Source Coding · Moonshot AI
Kimi K2.6: The new leading open weights model · Artificial Analysis
Amazon will invest $5bn in Anthropic now and up to $20bn more, and Anthropic committed >$100bn over ten years to AWS.
The deal secures up to 5GW of capacity, including nearly 1GW of Trainium2 and Trainium3 by the end of 2026, across Graviton and Trainium2 to Trainium4 chips. Amazon had already invested $8bn. Anthropic said run-rate revenue had passed $30bn, from about $9bn at the end of 2025, and that consumer growth had hurt reliability at peak hours.
Anthropic and Amazon expand compute collaboration · Anthropic
Amazon announces $5B Anthropic investment, up to $20B more · Amazon
Tuesday 21 April
The High Court in London rejected a challenge to the Metropolitan Police’s use of live facial recognition.
Big Brother Watch supported the case brought by anti-knife crime campaigner Shaun Thompson, whom the system falsely flagged near London Bridge. The court held that the Met’s policy met the requirements of the European Convention on Human Rights, and dismissed the claim. Thompson plans to appeal, The Register reported. The Met says the technology has led to >2,100 arrests since 2024.
Thompson v Met Police Commissioner [2026] EWHC 915 · National Archives
Summary of judgment · Judiciary of England and Wales
High Court approves Met Police’s facial recog after dispute · The Register
Florida’s attorney general opened a criminal investigation into OpenAI over ChatGPT’s chats with the Florida State University gunman.
James Uthmeier said that “if ChatGPT were a person, it would be facing charges for murder”. State prosecutors subpoenaed OpenAI for its policies and internal training materials from 1 March 2024 to 17 April 2026, among other records.
Uthmeier Launches Criminal Investigation into OpenAI · Florida Attorney General
Mozilla fixed 271 Firefox vulnerabilities found with Claude Mythos Preview.
The fixes ship in Firefox 150, after an early evaluation of Anthropic’s model. An earlier scan with Claude Opus 4.6 produced 22 security fixes in Firefox 148. Mozilla’s Bobby Holley wrote that the team has found “no category or complexity of vulnerability that humans can find that this model can’t”.
The zero-days are numbered · Mozilla
A small group of unauthorised users gained access to Claude Mythos Preview on the day Anthropic announced it, Bloomberg reported.
Members of a private Discord channel got in with the credentials of a member who works as an Anthropic contractor, plus data from a leak at Mercor, and used it for simple tasks such as building websites. Anthropic said it was investigating and had seen no sign that its own systems were compromised.
Anthropic’s Mythos AI Model Is Being Accessed by Unauthorized Users · Bloomberg
SpaceX secured the right to buy Cursor for $60bn.
SpaceX can buy Cursor later this year for $60bn or pay $10bn for their joint work instead. The partnership pairs Cursor’s product with SpaceX’s Colossus supercomputer, and Cursor said compute had been its training bottleneck.
Cursor had been hours from closing a $2bn round at a $50bn valuation, TechCrunch reported, and SpaceX will delay any purchase until after its IPO this summer. Microsoft had looked at buying Cursor but made no offer, CNBC reported.
SpaceX on X · SpaceX
Cursor partners with SpaceX on model training · Cursor
How SpaceX preempted a $2B fundraise with a $60B buyout offer · TechCrunch
Microsoft looked at buying Cursor before SpaceX deal · CNBC
Wednesday 22 April
Google split its eighth-generation TPUs into two chips: TPU 8t for training and TPU 8i for inference.
A TPU 8t superpod scales to 9,600 chips and 121 exaflops, nearly 3x the last generation. TPU 8i pairs 288 GB of high-bandwidth memory with 384 MB of on-chip SRAM. Both arrive later this year with up to 2x Ironwood’s performance per watt. The same Google Cloud Next event introduced Gemini Enterprise Agent Platform, bringing agent development, deployment and governance into one system.
Two chips for the agentic era · Google
Introducing Gemini Enterprise Agent Platform · Google Cloud
OpenAI introduced shared workspace agents in ChatGPT for recurring team workflows.
The research preview lets organisations build agents that use connected tools, run multi-step tasks and share the results across a workspace, with approval and governance controls. It brings longer-running agent workflows directly into the ChatGPT enterprise product.
Introducing workspace agents in ChatGPT · OpenAI
Thursday 23 April
The White House accused mostly China-based actors of “deliberate, industrial-scale campaigns” to distil American frontier AI models.
Michael Kratsios, director of the White House Office of Science and Technology Policy, told agency heads in a memo that the actors use proxy accounts and jailbreaks. The administration will share intelligence on the campaigns with US AI companies.
Adversarial Distillation of American AI Models (NSTM-4) · White House
U.S. accuses China of “industrial-scale” campaigns to steal AI secrets · Axios
White House accuses China of ‘industrial-scale’ AI theft · Financial Times
Meta told staff it will cut about 8,000 jobs, 10% of its workforce, from 20 May, and leave 6,000 open roles unfilled, Bloomberg reported.
Meta said the cuts are meant to make it more efficient and offset its heavy AI spending. Reuters reported on 17 April that further cuts could follow later in the year.
Meta Tells Staff It Will Cut 10% of Jobs in Push for Efficiency · Bloomberg
Meta targets May 20 for first wave of layoffs, additional cuts later in 2026 · Reuters
OpenAI released GPT-5.5.
OpenAI calls GPT-5.5 its smartest model, with the biggest gains in agentic coding, computer use, knowledge work and early scientific research, at GPT-5.4’s per-token latency. OpenAI reports 82.7% on Terminal-Bench 2.0, against 69.4% for Claude Opus 4.7 and 82.0% for Claude Mythos Preview.
API prices are $5/m input and $30/m output tokens, double GPT-5.4’s, and $30/m and $180/m for GPT-5.5 Pro. Both reached the API on Friday. Artificial Analysis puts GPT-5.5 3 points clear at the top of its Intelligence Index. On its AA-Omniscience benchmark, GPT-5.5’s hallucination rate is 86%, against 36% for Opus 4.7.
Chief scientist Jakub Pachocki told reporters on a call to expect “extremely significant improvements in the medium term”, and called recent progress “surprisingly slow”.
Introducing GPT-5.5 · OpenAI
GPT-5.5 narrowly beats Mythos Preview on Terminal-Bench · VentureBeat
OpenAI’s GPT-5.5 is the new leading AI model · Artificial Analysis
OpenAI Unveils GPT-5.5. Company Says Expect a Faster Release Pace · Tae Kim
The US Commerce Department dropped its plan to have former Anthropic researcher Collin Burns lead the Center for AI Standards and Innovation.
Burns was still onboarding when the department chose Chris Fall, who ran the Energy Department’s Office of Science in Trump’s first term, The Daily Signal reported.
Commerce Department Changes Direction on AI Safety Leader · The Daily Signal
Tesla disclosed in its Q1 2026 10-Q that it agreed in April to buy an unnamed AI hardware company for up to $2bn in stock and equity awards.
About $1.8bn depends on service conditions or performance milestones tied to deploying the company’s technology.
Tesla Q1 2026 Form 10-Q · Tesla
Tesla quietly discloses $2 billion AI hardware company acquisition · Electrek
Friday 24 April
DeepSeek released preview versions of DeepSeek-V4-Pro and DeepSeek-V4-Flash as open weights.
DeepSeek-V4-Pro has 1.6T total and 49B active parameters and V4-Flash 284B and 13B, and both take 1M tokens of context. At that length V4-Pro needs 27% of the single-token compute and 10% of the KV cache of DeepSeek-V3.2, DeepSeek says.
Artificial Analysis scores V4-Pro 52 on its Intelligence Index, up from 42 for V3.2 and second among open-weights models to Kimi K2.6 (Moonshot AI) at 54. API prices are $1.7/m input and $3.5/m output tokens for Pro and $0.14/m and $0.28/m for Flash, the cheapest in their classes, Simon Willison noted.
DeepSeek V4 Preview Release · DeepSeek
deepseek-ai/DeepSeek-V4-Pro · Hugging Face
DeepSeek is back among the leading open weights models · Artificial Analysis
DeepSeek V4: almost on the frontier, a fraction of the price · Simon Willison
Canada’s Cohere plans to acquire Germany’s Aleph Alpha in a sovereign-AI deal that remains subject to regulatory conditions.
Companies of Schwarz Group, Aleph Alpha’s backer, intend to lead Cohere’s Series E with a €500m ($600m) structured financing commitment. The FT’s headline called it a $20bn transatlantic tie-up.
Cohere and Aleph Alpha Join Forces · Cohere
Cohere and Aleph Alpha agree $20bn transatlantic AI tie-up · Financial Times
Cohere to acquire German AI company Aleph Alpha · CNBC
China’s National Development and Reform Commission has told several private tech companies to reject US money, Bloomberg reported.
Moonshot AI, StepFun and ByteDance are among them. The move followed Meta’s $2bn purchase of Manus.
China to Curb US Investment in Tech Companies After Meta Deal · Bloomberg
Google will invest up to $40bn in Anthropic.
Google is investing $10bn now at a $350bn valuation, with $30bn more to follow if Anthropic hits performance targets, according to Anthropic. Google Cloud will also provide 5GW of new capacity over the next five years.
Google Plans to Invest Up to $40 Billion in Anthropic · Bloomberg
Google to invest up to $40B in Anthropic in cash and compute · TechCrunch
The US Justice Department moved to intervene in xAI’s lawsuit against Colorado’s law on algorithmic discrimination.
It argues that the law, SB24-205, breaches the Equal Protection Clause by requiring AI companies to prevent unintentional disparate impact while exempting discrimination meant to advance “diversity”. xAI sued on 9 April.
Justice Department Intervenes in xAI lawsuit · US Department of Justice
Sam Altman apologised to the community of Tumbler Ridge, British Columbia, because OpenAI did not alert police to an account it had banned in June.
He wrote: “I am deeply sorry that we did not alert law enforcement to the account that was banned in June”. Altman had promised the apology to Premier David Eby and Mayor Darryl Krakowka a month earlier, and it came the day after Eby said the Royal Canadian Mounted Police investigation into the shooting was in its final stages.
OpenAI apologizes to Tumbler Ridge · Tumbler RidgeLines
Meta agreed to deploy tens of millions of Amazon Graviton CPU cores to support agentic AI workloads.
The AWS deal covers CPU-intensive tasks such as coordinating agents, running tools and managing data alongside accelerator-based model inference. It points to another part of the AI infrastructure buildout: demand for general-purpose computing as agent workflows expand.
Meta agreement with AWS to power agentic AI on Amazon’s Graviton chips · AWS
ALETH · OPEN ACCESS RESEARCH