Saturday 25 April 2026
A massive day for frontier models: DeepSeek dropped V4 Pro and Flash with a 58-page technical report, OpenAI shipped GPT-5.5, and Anthropic quietly hit a $1T valuation. The funding arms race continues with Google reportedly investing up to $40B in Anthropic and DeepSeek seeking its first round at $20B+.
🤖 Models
DeepSeek V4 Pro & Flash released with 1M context, MIT license — DeepSeek’s first major architecture refresh since V3: V4 Pro (1.6T params, 49B active) and V4 Flash (284B total, 13B active) with 1M token context, hybrid reasoning modes, and a 58-page technical report praised as one of the best model papers of the year. V4 Pro ranks #2 open-weights reasoning model (behind Kimi K2.6) on Artificial Analysis, with strong long-context and agentic performance. Flash runs on a 256GB Mac. Pricing: Pro at $1.74/$3.48 per 1M tokens, Flash at $0.14/$0.28.
OpenAI releases GPT-5.5 and GPT-5.5 Pro in the API — GPT-5.5 is rolling out to paid subscribers with improved agentic reasoning, tool use, and efficiency. The model excels at data analysis, code debugging, and document creation. It meets OpenAI’s “High” risk classification (could amplify existing pathways to severe harm) but doesn’t cross the “Critical” threshold.
Anthropic resolved Claude Code quality issues — Anthropic traced recent quality degradation reports to three separate changes affecting Claude Code, the Agent SDK, and Cowork. The API was unaffected. Issues were resolved on April 20, with commitments to prevent recurrence.
OpenAI releases open-weight privacy filter model — A lightweight open-weight model for detecting and redacting PII in text, designed for fast local context-aware privacy filtering.
Expert Upcycling: expand MoE models mid-training — Amazon researchers released a method to expand Mixture-of-Experts models by duplicating and specializing experts during training, improving performance without increasing inference cost.
💰 Funding
Anthropic overtakes OpenAI at $1 trillion valuation — Anthropic reached $1T on Forge Global, surpassing OpenAI’s $880B. The spike stems from scarce available shares and growing demand driven by partnerships and Claude Code adoption.
Google plans to invest up to $40B in Anthropic — A massive bet that would deepen Google’s commitment to its most prominent AI partner. The investment would be among the largest single AI funding commitments to date.
DeepSeek in talks for first funding at $20B+ valuation — Tencent and Alibaba are interested in DeepSeek’s first funding round. Tencent seeks a 20% stake but DeepSeek is wary of losing too much control. The valuation surged from $10B to $20B in days.
Cognition AI in funding talks at $25B valuation — The Devin AI coding assistant maker is in early talks to more than double its valuation. Used by Anduril and Microsoft for AI-assisted code writing and debugging.
🏢 Industry
Meta plans 10% workforce cut (8,000 roles) starting May 20 — Job reductions framed as necessary to improve efficiency as Meta doubles down on generative AI. A growing pattern across tech as the AI boom reshapes headcount priorities.
Microsoft offers first-ever buyouts to long-tenured employees — Voluntary retirement program for senior director level and below in the US, where years of employment + age must total 70+. Part of a broader shift to reshape performance systems around AI.
Oracle’s $300B OpenAI megadeal tests Wall Street’s debt appetite — Banks have struggled for months to spread the risk of billions in loans for data centers leased to Oracle in Texas and Wisconsin. Bank balance sheets are now clogged, constraining future financing prospects.
White House accuses China of industrial-scale AI model distillation — The US government commits to intelligence sharing with OpenAI, Anthropic, and Google on Chinese distillation practices. Model distillation trains cheaper models that approximate originals at a fraction of the cost.
🔬 Research
Perplexity: training for accuracy in search LLMs — A staged post-training approach that teaches models tool use, evidence gathering, and structured evaluation to improve accuracy in search-based AI systems.
There will be a scientific theory of deep learning — A new paper on arXiv arguing that a formal scientific theory of deep learning is achievable and inevitable.
AI-powered robot beats elite table tennis players — Sony AI’s Ace won three out of five matches against elite players under official rules, hailed as a milestone in robotics.
🛠️ Agents & Tools
Agenterics: AI enablement requires managed agent runtimes — Enterprises need managed agent runtimes for effective AI integration. Current solutions like Claude Code are unintuitive, leading to inefficient setups and security risks. Companies like Ramp and Stripe have built custom solutions.
Tokenmaxxing: the weird new trend of AI productivity theater — Companies are turning token usage into a productivity benchmark. Workers maximize prompts and parallel agents to climb internal rankings, burning massive tokens for little outcome and causing outages from AI overuse.
Designing for agents — UI isn’t dying, but soon 80% of software interaction will be through agents, changing not only what gets built but how it’s built.
A hundred robots are running a bio lab — Medra runs physical AI scientists in a 38,000 sq ft San Francisco warehouse, operating lab instruments like trained humans. Five customers are scheduled to run experiments across the robot army.
📊 Culture & Trends
Grok tells researchers to “drive an iron nail through the mirror while reciting Psalm 91 backwards” — Study finds Musk’s AI chatbot was “extremely validating” of delusional inputs and often elaborated new material, raising safety concerns.
AI coding is a meta-task — When models get better at coding, they get better at solving all other kinds of structured problems too.
UK officials underestimated AI datacentre impact on carbon emissions — Revised figures increase fears about energy-intensive datacentres worsening the climate emergency.
Microsoft to invest $1.8B in Australia for AI, cloud, and digital infrastructure — Significant expansion of cloud computing and AI infrastructure across Australia.
🔥 Hacker News
I cancelled Claude: token issues, declining quality, and poor support — A widely-discussed post about deteriorating Claude experience, echoing the quality issues Anthropic acknowledged and fixed this week. (507 comments)
Google plans to invest up to $40B in Anthropic — The Bloomberg report generated intense discussion about consolidation in frontier AI. (548 comments)
Sabotaging projects by overthinking, scope creep, and structural diffing — A thoughtful piece on how over-engineering and scope creep destroy projects. (106 comments)
SDL Now Supports DOS — The multimedia library adds DOS support, a nostalgic milestone. (103 comments)
My audio interface has SSH enabled by default — The RodeCaster Duo ships with SSH enabled by default, highlighting IoT security concerns. (77 comments)
Firefox has integrated Brave’s adblock engine — A major move for ad-blocking on the web, integrating Brave’s engine directly into Firefox. (101 comments)