ALETH / AI-BRIEF / 2026-04-18 / CEREBRAS TAKES ITS NVIDIA CHALLENGE TO THE NASDAQThe Aleth Briefs link to the sources behind the stories and show how the week unfolded.
The week in five lines:
Cerebras filed for a Nasdaq listing after 2025 revenue rose 76% to $510m.
Anthropic shipped Claude Opus 4.7 with its cyber skills deliberately reduced.
The UK’s AI Security Institute saw Mythos complete a 32-step network attack.
OpenAI answered Mythos with GPT-5.4-Cyber, released to vetted defenders only.
GPT-5.4 Pro solved Erdős problem #1196, and Terence Tao praised its method.
Browse by day:
Weekend: Berkeley benchmarks
Monday: OpenAI London · AISI · Erdős
Tuesday: Nscale · GPT-5.4-Cyber · Meta-Broadcom · Gemini Robotics
Wednesday: Jane Street · WhatsApp
Thursday: Qwen · Opus 4.7 · Codex · Callosum · Rosalind · π0.7 · London
Friday: Claude Design · Cerebras · Kevin Weil
Weekend (11-12 April)
Researchers at UC Berkeley showed that eight leading AI agent benchmarks can be gamed to near-perfect scores without solving a single task.
Their agent found that a 10-line Python file marks every SWE-bench Verified instance resolved, and that a fake curl wrapper scores 100% on all 89 Terminal-Bench tasks.
How We Broke Top AI Agent Benchmarks · Berkeley RDI
Monday 13 April
OpenAI signed a lease for its first permanent London office, with room for >500 staff.
The 88,500 sq ft office is in King’s Cross, and OpenAI employs about 200 people in London. The lease came days after OpenAI paused its Stargate UK project.
OpenAI announces first permanent London office · CNBC
The UK AI Security Institute saw Claude Mythos Preview complete a 32-step simulated network attack.
Anthropic’s withheld model is the first to complete The Last Ones, AISI’s 32-step simulated attack that runs from reconnaissance to full network takeover and takes a human ~20 hours. It finished in 3 of 10 attempts and averaged 22 steps, while Claude Opus 4.6, the next best, averaged 16. On expert-level capture-the-flag tasks, which no model could complete before April 2025, it succeeds 73% of the time. AISI directed the model and gave it network access, with a budget of up to 100M tokens per run.
Our evaluation of Claude Mythos Preview’s cyber capabilities · AI Security Institute
GPT-5.4 Pro (OpenAI) solved Erdős problem #1196, a conjecture about primitive sets.
Liam Price posted the proof in the Erdős Problems forum on Monday. Terence Tao wrote on Tuesday that its method would be “a meaningful contribution to the anatomy of integers that goes well beyond the solution of this particular Erdos problem”. Math, Inc. had formally verified the proof by Thursday.
1196 Discussion Thread · Erdős Problems
Tuesday 14 April
Microsoft will take >30,000 Nvidia Rubin GPUs from Nscale at a Norwegian site once earmarked for OpenAI’s Stargate.
The GPUs go into Nscale’s 230 MW campus at Narvik, with capacity arriving in 2027. OpenAI failed to conclude an agreement for the site, Bloomberg reported.
Nscale announces expanded deal with Microsoft in Norway · Nscale
Microsoft contracts 30,000 GPUs from Nscale at Norway data center · DCD
OpenAI released GPT-5.4-Cyber to vetted security defenders.
GPT-5.4-Cyber is a version of GPT-5.4 fine-tuned for cyber work, with fewer restrictions and a lower refusal boundary for legitimate security tasks. It adds binary reverse engineering, so analysts can examine compiled software for malware and vulnerabilities without the source code. OpenAI is starting with a limited deployment to vetted security users in the top tiers of its Trusted Access for Cyber programme.
Trusted access for the next era of cyber defense · OpenAI
Meta committed to >1 GW of custom MTIA chips co-developed with Broadcom, and Broadcom chief executive Hock Tan will leave Meta’s board.
The 1 GW is the first phase of a multi-gigawatt roll-out across several chip generations. Broadcom says the MTIA will be the industry’s first 2nm AI accelerator. Tan becomes an adviser to Meta.
Meta Partners With Broadcom to Co-Develop Custom AI Silicon · Meta
Broadcom Announces Extended Partnership with Meta · Broadcom
Google DeepMind released Gemini Robotics-ER 1.6, a reasoning model that plans robot tasks, through the Gemini API.
DeepMind says it beats Gemini Robotics-ER 1.5 and Gemini 3.0 Flash on spatial reasoning such as pointing, counting and success detection. It can also read gauges and sight glasses, a capability developed with Boston Dynamics.
Gemini Robotics-ER 1.6: Enhanced Embodied Reasoning · Google DeepMind
Wednesday 15 April
Jane Street committed about $6bn to CoreWeave’s AI cloud, including Nvidia Vera Rubin capacity, and bought $1bn of CoreWeave stock.
It paid $109 a share for the Class A stock.
Jane Street Signs $6 Billion AI Cloud Agreement With CoreWeave · CoreWeave
The European Commission told Meta it intends to order WhatsApp reopened to rival AI assistants on the terms in place before October 2025.
The Commission’s supplementary statement of objections says Meta’s March switch to paid access is in effect equivalent to its earlier ban. It also extended the investigation to Italy, so it now covers the whole EEA.
Commission sends Meta fresh WhatsApp charge sheet · European Commission
Thursday 16 April
Alibaba’s Qwen team released Qwen3.6-35B-A3B, the first open-weight Qwen3.6 model.
It has 35B parameters, 3B active. Alibaba says it scores 73.4 on SWE-bench Verified and 51.5 on Terminal-Bench 2.0, against 52.0 and 42.9 for Gemma 4 31B (Google).
Qwen/Qwen3.6-35B-A3B · Hugging Face
Anthropic released Claude Opus 4.7 and said it had deliberately held back its cyber capabilities.
Anthropic calls the model a notable improvement on Claude Opus 4.6 in advanced software engineering, with the biggest gains on the hardest tasks. It is less broadly capable than Claude Mythos Preview, and has reduced cyber capabilities, shipping with safeguards that block high-risk cybersecurity requests.
Prices stay at $5/m input and $25/m output tokens. Artificial Analysis scores it 57.3 on its Intelligence Index, tied for first with Gemini 3.1 Pro at 57.2 and GPT-5.4 at 56.8.
Introducing Claude Opus 4.7 · Anthropic
Opus 4.7: Everything you need to know · Artificial Analysis
OpenAI gave Codex control of the user’s computer, an in-app browser, image generation, memory and >90 new plugins.
Codex, used weekly by >3 million developers, can now schedule its own work and resume long tasks. Computer use starts on macOS and reaches EU and UK users later.
Codex for (almost) everything · OpenAI
The UK’s £500m Sovereign AI Unit made its first equity investment, in Callosum, which builds software that makes different kinds of chip work together.
Six more start-ups get up to 1 million GPU hours each on the UK’s supercomputers. Technology Secretary Liz Kendall announced the awards.
AI firms get first backing through UK’s Sovereign AI · GOV.UK
The White House budget office began work on safeguards before possibly giving federal agencies a modified Claude Mythos Preview.
Federal chief information officer Gregory Barbaccia’s Tuesday email said OMB is working on safeguards before “potentially releasing a modified version of the model to agencies”, Bloomberg and Nextgov/FCW reported. On Friday an OMB spokesperson said “OMB is not giving access to anything to agencies”.
OMB is ‘not giving access to anything to agencies,’ official says · Nextgov/FCW
OpenAI launched GPT-Rosalind, a reasoning model for research, drug discovery and translational medicine, as a research preview for vetted customers.
Amgen, Moderna, the Allen Institute and Thermo Fisher Scientific are among the first users, through OpenAI’s trusted access programme, Axios reported.
Introducing GPT-Rosalind for life sciences research · OpenAI
OpenAI launches new AI model for life sciences research · Axios
Physical Intelligence released π0.7, a robot model it says can recombine skills learned in training to do tasks it was never taught.
The company says the model matched its own task-specific models at making coffee, folding laundry and assembling boxes. Coached step by step in plain language, it handled an air fryer far better. The paper claims only “early signs” of generalisation.
A Steerable Model with Emergent Capabilities · Physical Intelligence
Physical Intelligence says its robot brain can do tasks never taught · TechCrunch
Anthropic said it was taking office space for 800 people in London’s Knowledge Quarter, days after OpenAI’s London lease.
It has >200 staff in the city.
Anthropic unveils plans for major London expansion · CNBC
Friday 17 April
Anthropic launched Claude Design, a Labs product for making prototypes, slides and one-pagers with Claude - Figma and Adobe shares fell.
Claude Design runs on Claude Opus 4.7 and is in research preview for Pro, Max, Team and Enterprise subscribers. Work exports to Canva, PDF, PPTX or HTML.
Introducing Claude Design by Anthropic Labs · Anthropic
Anthropic launches “Claude Design,” sending Figma shares down · Sherwood News
Cerebras filed to list on Nasdaq after revenue rose 76% in 2025.
Cerebras’s S-1 shows 2025 revenue of $510m and net income of $238m, against a $482m net loss in 2024. It will list under the ticker CBRS, with Morgan Stanley, Citigroup, Barclays and UBS among the lead underwriters. Two UAE customers dominate. Mohamed bin Zayed University of Artificial Intelligence was 62% of 2025 revenue and G42 24%. OpenAI has lent Cerebras $1bn and holds a warrant to buy its stock, and in January agreed a multi-year deal worth >$20bn to deploy 750 MW of Cerebras compute. The Information reported an indicative valuation of >$35bn.
Cerebras - S-1 (April 2026) · SEC
Cerebras Prepares Public Listing, Eyes $35 Billion-Plus Valuation · The Information
Kevin Weil left OpenAI, which is splitting up his OpenAI for Science group and closing Prism, its AI workspace for scientists.
Prism’s team of about 10 moves under Codex chief Thibault Sottiaux. Bill Peebles, who led Sora, also announced his departure on Friday.
Today is my last day at OpenAI · Kevin Weil on X
OpenAI Executive Kevin Weil Is Leaving the Company · WIRED
Kevin Weil and Bill Peebles exit OpenAI as company sheds ‘side quests’ · TechCrunch
ALETH · OPEN ACCESS RESEARCH