Monday 25 May 2026

model: xiaomi/mimo-v2.5-pro

A Monday morning grab bag: DeepSeek goes permanent on pricing, AI infra companies printing money, Cursor at $3B ARR, and the “all labs are agent labs” thesis hardening. Plus Iran deal drama, immigration detention failures in Australia, and HN losing its mind over DOS source code.

🤖 Models and Launches

DeepSeek Makes 75% Discount Permanent on V4 Pro — DeepSeek’s flagship model pricing is now $0.435/M input, $0.87/M output permanently. Artificial Analysis estimates running their Intelligence Index on V4 Pro costs ~3x less than Gemini 3.1 Pro, ~12x less than GPT-5.5, and ~19x less than Claude Opus 4.7. The “intelligence too cheap to meter” era accelerates.

Qwen3.7-Max: The Agent Frontier — Alibaba’s Qwen team released Qwen3.7-Max as a proprietary agent-foundation model, posting top scores on Terminal-Bench 2.0, SWE-Pro, SciCode, MCP-Mark, GPQA Diamond, and IMOAnswerBench. Consistent performance across Claude Code, OpenClaw, and custom harnesses.

Gemini 3.5 Flash Tops APEX-Agents-AA — Google’s Gemini 3.5 Flash ranks #1 on the APEX-Agents-AA benchmark, outperforming larger models. Also jumped 16 positions on Design Arena. Builders note benchmark gains don’t always translate to real-world agent usefulness though.

NVIDIA’s Gated DeltaNet-2 Outperforms KDA and Mamba-3 — Decouples erase and write operations in linear attention with channel-wise gates. At 1.3B parameters, beats competitors on language modeling and commonsense reasoning with notable long-context retrieval gains on RULER.

Microsoft Plans to Supply Maia AI Chips to Anthropic — Anthropic faces compute challenges despite Amazon and Google partnerships. Microsoft’s Maia chips offer ~30% improved performance, backed by a $5B investment from November.

💰 Funding and Revenue

Cursor Hits $3B ARR Ahead of SpaceX Deal — Cursor’s annualized revenue hit $3 billion in late April with 3,000+ customers paying at least $100K each. SpaceX has the right to buy Cursor for $60 billion during a 30-day window after it begins trading publicly, expected June 12.

Hark Raises $700M at $6B Valuation — Targeting GPU infrastructure, foundation models, hardware, and multimodal/personal intelligence products. Also reported a 200-hour uninterrupted autonomous run for F.03.

Modal Raises $355M Series C at $4.7B — Rebuilding the cloud stack for AI workloads from the ground up. Strong developer experience and performance positioning it as a core AI cloud winner.

Exa Raises $250M Series C at $2.2B — The AI search/retrieval company joins the unicorn class alongside Modal and TurboPuffer.

TurboPuffer Hits $100M ARR, Profitable — Went from $1M to $100M run-rate in 19 months while raising less than $1M total. “The magic happens with AI when it draws in just the right context.”

OpenAI Q1 Revenue: $5.7 Billion — Surpasses Anthropic. High demand for inference chips has Amazon, Google, and potentially Microsoft vying to supply custom solutions.

🏢 Industry

All Model Labs Are Now Agent Labs — Greg Brockman says “the model alone is no longer the product.” AI21 shuttering its model team to pivot to agents. DeepSeek building a “Harness team” for the first time. The shift from model-as-product to model-plus-harness-plus-workflow is the defining industry move of 2026.

Manus Weighs Raising $1B to Unwind Meta Takeover — China’s regulators ordered Manus to reverse its Meta acquisition. Founders in discussions to raise funds and may chip in their own money. A reversal this large months after completion is virtually unheard of.

Microsoft Cancels Claude Code Licenses — Shifting developers to GitHub Copilot CLI. Likely driven by financial motives given Microsoft’s Anthropic investment and chip deal.

ClickUp Cuts 22% of Staff — Claims strongest position ever, but cutting to “operate at highest level of productivity.” Plans million-dollar salary bands for staff creating outsized impact using AI.

Google Pushes AI-Generated Ads Into Search Results — Testing new ad formats in standard search and AI Mode. AI-generated brand ads appear below AI Mode responses. Google promises “strong guardrails” against hallucinations.

🛠️ Agents and Tools

DeepSeek Reasonix: Native Coding Agent — DeepSeek’s high-caching, low-cost native coding agent. 171 HN comments — the community is paying attention to DeepSeek’s agent play.

Lessons from Building Cloud Agents — Cursor describes durable execution, isolated dev environments, self-healing infrastructure, and cleaner separation between agent state and conversation state.

Constraint Decay: Fragility of LLM Agents in Back-End Code Generation — Research paper examining how LLM agents degrade when generating back-end code. 66 HN comments.

Cheap Code Means Formal Verification Is Reasonable Now — TLA+ and formal spec languages can catch hallucinations that language models create. Comprehensive conformance test suites help guarantee agents produce what developers actually want.

Codex Expands: Appshots, Remote Computer Use, Plugin Sharing — OpenAI’s Codex can now securely use apps on your Mac from your phone even when locked. Appshots capture both screenshot and text from app windows for richer context. Team plugin sharing and org analytics also launched.

🔬 Research

Can SAEs Capture Neural Geometry? — Goodfire AI argues the “models think in curved manifolds, SAEs use straight-line features” critique is only partly right. Fix: cluster SAE features by joint firing patterns to recover geometry through feature groups.

The Best Data Filter May Be No Filter — Surprising scaling result on DCLM: with enough compute, filtering internet-scale data may be unnecessary, with the crossover projected around 1e30 FLOPs.

RAEv2: Representation Autoencoders — >10x faster convergence, better reconstruction and generation. Summing the last K encoder layers instead of only the final layer improves both without added inference cost.

Runway Aleph 2.0 and Edit Studio — Edit a single frame and propagate that edit through the rest of the video. Practical productization of reference-guided edit propagation.

Memory Is Now 63% of AI Chip Component Costs — HBM grew from 52% to 63% of total AI chip component spending from Q1 2024 to Q4 2025. 251 HN comments — the memory wall is real.

🔥 Hacker News

Microsoft Open-Sources Earliest DOS Source Code Discovered to Date — 138 comments. A nostalgia bomb and a piece of computing history.

Claude Is Not Your Architect. Stop Letting It Pretend — 141 comments. The pushback against AI-generated architecture decisions gains traction.

Why Is Vivado 2026.1 Dropping Linux Support for Free Tier? — 159 comments. AMD’s FPGA toolchain alienating the open-source hardware community.

Scammers Abusing an Internal Microsoft Account to Send Spam — 142 comments. Not a great look for Microsoft’s internal security posture.

Greg Brockman Interview — 142 comments. Wide-ranging conversation on OpenAI’s direction and the future of AI.

🌐 Notable

AI Fault Lines on Show at Cannes — Darren Aronofsky among proponents using AI in filmmaking, while Guillermo del Toro says he would “rather die.” The cultural divide over AI in creative industries deepens.

How Big Tech Got Its Way on Trump’s AI Executive Order — The US president’s reversal on calling for a safety review of new AI models is a green light for tech’s unchecked power.

Figure AI’s Humanoid Robots Go Viral — Livestreaming robots placing packages on a conveyor belt since May 13. Originally planned for 8 hours, now running 24/7. Viewers have named the robots.

Eli Lilly’s Retatrutide: Dramatic Weight Loss Results — Participants stopped taking it because they thought they were losing too much weight. Results for the heaviest patients on par with gastric bypass surgery. Not yet applied for regulatory approval.

SpaceX IPO Cements Musk’s Grip — Supervoting shares, mandatory arbitration, stricter rules on shareholder proposals, and Texas corporate law give Musk virtually unchecked executive authority when the company goes public.

🌍 World and Australia

US and Iran Inch Closer to Peace Deal — Trump faces criticism from GOP hawks including Ted Cruz and Lindsey Graham. The US has apparently agreed to unfreeze billions of Iranian assets.

Catastrophic Security Failures in Australia’s Immigration Detention — Prison multinational MTC uses a “minimalist staffing model” resulting in escapes, fires, and a stabbing. Critics say detainees and staff are in serious danger.

Teal MPs Confirm Conversations About New Political Party — Zali Steggall confirms the independents are having “conversations” amid speculation about a formal teal party.

Man Dies in Shark Attack on Great Barrier Reef — The 39-year-old was fishing at Kennedy Shoal between Cairns and Townsville.

Sources: TLDR (General + AI), AINews/Latent Space, The Guardian, Hacker News — 4 newsletters processed, ~45 stories distilled