The Lyceum: AI Daily — May 31, 2026

Today's signals all point to one shift: the most consequential AI moves aren't being made by frontier labs anymore — they're being made by the layers above and below them. DeepSeek set a permanent price floor that Western labs have not answered. OpenRouter raised $113M from the very companies whose models it routes around. And Brussels handed itself enforcement powers that don't kick in until August but already have OpenAI, Anthropic, and Google staring at a six-month clock. The model is no longer the product. The stack around it is.

• DeepSeek V4 Pro (DeepSeek): The 75% promotional discount became permanent — output tokens drop to $0.87/M from $3.48, locked in as the new list price.
• OpenRouter platform expansion (OpenRouter): 400+ models across providers, now with deeper web search/fetch and human-in-the-loop tooling — pushing closer to an agent runtime than a model marketplace.
• Otari (Mozilla.ai): An "own your AI stack" gateway pairing open-weight models with an agent runtime, shipped May 29 — paired with Cq (the shared knowledge layer for coding agents launched two days earlier).
• Minicor (YC P26): Agentic Windows desktop automation at fleet scale — AI-native RPA aimed at the $13B legacy automation market.
• embeddings-v5-omni (OpenAI): A multimodal embedding model spanning text, images, audio, and video — one numerical representation across all four modalities.

The most important word in DeepSeek's announcement isn't a number. It's "permanent."

V4 Pro launched on April 24 at $1.74 input / $3.48 output per million tokens, with a 75% promotional discount scheduled to expire today. On May 22, DeepSeek announced that the discount would not roll back — the promo rate is now the list price. V4 Pro is now $0.435 input and $0.87 output per million tokens, roughly 11.5x cheaper than GPT-5.5 on input and 34.5x cheaper on output, per Codersera's pricing analysis.

A discount that expires is a marketing event. A discount that doesn't is a reference price. Every Chinese frontier lab now has to match it, and OpenAI, Anthropic, and Google are sitting on a margin question they haven't publicly answered.

The strategic subtext is hardware. DeepSeek had told customers V4 Pro pricing would ease once Huawei's Ascend 950 supernodes shipped in volume in H2 — and making the cut permanent before H2 even begins suggests the Ascend pipeline is materializing faster than disclosed. If that read is correct, US export controls are working less well than the Commerce Department's posture implies.

Watch for any Western lab to cut mid-tier API pricing before June 15. That's the signal the cost war has crossed the Pacific in both directions.

When the company that routes traffic between AI models raises money from the companies whose models it routes, something has shifted.

OpenRouter announced a $113M Series B led by CapitalG — Alphabet's independent growth fund — with NVentures (NVIDIA's venture arm), ServiceNow Ventures, MongoDB Ventures, Snowflake Ventures, and Databricks Ventures all participating alongside Andreessen Horowitz and Menlo. TechCrunch reports the post-money valuation at roughly $1.3 billion — more than double the $547 million it carried a year ago.

The usage explains the enthusiasm. Per Yahoo Finance's coverage, weekly volume has surged from 5 trillion tokens six months ago to 25 trillion now — a 5x jump driven by enterprises scaling agents across multiple providers. The network spans 400+ models and over a million developers.

Google and NVIDIA both invested in a platform that routes traffic away from any single provider, including their own. That tells you where enterprise procurement is going: no one is standardizing on a model the way they once standardized on a SaaS vendor. The lock-in era is over before it began.

The failure mode worth tracking: whether OpenRouter can hold the routing layer as labs build their own enterprise gateway products. If OpenRouter ships governance and audit features before Q3, the multi-model layer becomes the de facto compliance infrastructure for enterprises navigating the DeepSeek question. If it doesn't, the labs reclaim the relationship.

European AI regulation just got its first enforcement deadline with real teeth, and it's six months away.

EU member states and the European Parliament reached a provisional agreement on May 7 amending the AI Act to prohibit the market placement and use of AI systems generating sexual images without consent, effective December 2, 2026. The move came amid concerns tied to xAI's Grok.

The December date doesn't require conformity assessments or risk-classification packages. It requires that prohibited applications simply not exist in the EU market. For any GenAI provider with image or video generation surfaces, the question isn't classification — it's whether any endpoint can produce the prohibited output.

The broader Digital Omnibus package postponed high-risk biometrics and critical infrastructure rules to December 2027 — but expanded the EU AI Office's powers over general-purpose model providers. The AI Office's enforcement authority formally activates in August 2026. Per RTÉ, an EU spokesperson noted that after August, the Office "will ensure to receive, if needed, model access" — pointed language given that Brussels has met with Anthropic about Mythos multiple times without yet obtaining access.

Every image and video generation API serving EU users now has a clock running. Watch whether the AI Office's first formal model-access demand lands before August 2 — that would mean Brussels has moved from regulatory observer to active participant in frontier safety governance faster than anyone modeled.

The agentic AI story most people are filing under "interesting demo" is actually a production infrastructure play.

Minicor, a YC P26 company, launched this week with a deceptively simple pitch: Windows desktop automation at fleet scale. Not a chatbot. Not a coding assistant. An agent that operates the actual Windows GUI — the buttons and menus a human would click — across thousands of machines at once.

Desktop automation isn't new. UiPath and Automation Anywhere have sold robotic process automation (RPA) for years. The catch is that legacy RPA is brittle: scripts break when a button moves two pixels. AI-native agents that understand context can handle the variation. The RPA market is worth roughly $13 billion annually, and most of it is held together with fragile scripts.

If Minicor's agents reliably replace those scripts, the incumbents have a serious problem and Microsoft Copilot has a new competitor on its home turf. The signal to watch: whether enterprise buyers treat Minicor as an RPA replacement or a net-new automation layer. The answer determines whether the addressable market is $13B or much larger.

Mozilla.ai published Otari on May 29 as an "own your AI stack" gateway and hosted platform. Branding aside, the shape is what matters: it pairs open-weight models with an agent runtime, treating local AI as production infrastructure rather than a hobbyist toy.

Two days earlier, Mozilla launched Cq — a shared knowledge system for coding agents, essentially Stack Overflow for the retry-loop problem. These aren't isolated experiments. Together, they're an emerging open agent operations stack.

The economics matter. A lot of 2026 AI spend has gone to API margins and retry-loop overhead. If teams can combine open models, self-hosted routing, and shared agent memory, the orchestration layer absorbs margin the frontier labs assumed they owned. When a non-hyperscaler starts publicly stitching gateways, runtimes, and agent memory together, it means enough developers have already decided the closed-stack default is too expensive or too rigid.

PJM, the largest U.S. power market, received emergency approval from the Department of Energy to curtail data centers and other large loads with backup generation during hot-weather grid stress, per Datacenter Dynamics' May 20 report.

The approval itself isn't fresh news, but it's newly relevant: today's capital and routing stories all converge on the same bottleneck. Power. A router, an inference engine, or a giant GPU contract all look different once utilities can treat AI campuses as interruptible industrial load.

This stops being an energy footnote when the next utility-side notice or flexible-load partnership lands. At that point, "uptime" becomes a product attribute that depends on weather.

Anthropic's Seoul office announcement was covered as a commercial move. The more interesting detail sits in Korean trade press.

According to The Elec, a Korean semiconductor and electronics publication, Anthropic conducted cybersecurity cooperation discussions using Claude Mythos during a May workshop with Korean government agencies — including the Ministry of Science and ICT, the Ministry of Foreign Affairs, the National Intelligence Service, the Financial Services Commission, the National AI Strategy Committee, and the Financial Security Institute. That's not a commercial pilot. That's an unreleased frontier model being used in joint sessions with a country's intelligence service and foreign ministry.

Anthropic has publicly said it plans to expand Project Glasswing with the US government and allied governments. The Korea workshop suggests that expansion is already happening, quietly, before any formal allied-government announcement. Treat the workshop detail as a strong signal rather than confirmed fact — it's single-sourced — but if accurate, Anthropic entered the sovereign intelligence tier without a press release.

Today: a Chinese lab quietly setting the global price of intelligence at $0.87, an Alphabet fund cheerfully writing checks to the company that helps developers route around Alphabet's models, and a Korean intelligence agency sitting in a workshop with an unreleased frontier AI that nobody outside Anthropic is supposed to have. The model itself, it turns out, was never the moat — it was the part everyone was looking at while the floor, the routing layer, and the regulator quietly changed shape underneath. We'll see what's left standing tomorrow.

Forward this to the friend who still thinks the AI race is about benchmarks.