Tokens become units of account — Opus 4.8, OpenRouter, Cognition, Snowflake, and Pope’s warning

• Continuing our series on transformer alternatives.

• This week’s AI section covers Opus 4.8.

• This week’s opinion discusses the strategic differences between companies like Google, NVIDIA, Microsoft, OpenAI, and Anthropic when it comes to ownership of different areas of the AI ​​stack.

For two years, the AI ​​boom has been a conversation about the future told in benchmarks and term sheets. This week’s discussion focused on the current state of affairs as told by income.

Start with the board. Anthropic has shipped Claude Opus 4.8 and at the same time disclosed that it is tracking towards its first operating profit (expected second-quarter revenue of approximately $10.9 billion, up approximately 130% sequentially) while closing a $65 billion round. So sit down. The laboratory, which is still conducting frontier training, is approaching operational surplus. The belief that “research laboratories are structurally unprofitable”, which was at the root of all bearish incidents, has simply lost its load-bearing wall.

Anthropic refreshingly described the model itself as a “modest but visible improvement.” For the same price as 4.7, agent coding is nudged from ~64% to ~69% and tool inference is nudged from ~55% to ~58%. The interesting part is below the benchmarks. Three changes are important. First is effort control, which adjusts how much effort the model takes to think for each task. This is explicit governance of the compute-quality trade-off that all agent builders have been hacking with prompting tricks. Number 2, dynamic workflow: Claude code functionality where the model plans a large task, spins up parallel subagents to attack parts, and validates and reports its output. Combined with the Messages API, it accepts live edits to the running message array without corrupting the prompt cache, allowing you to work with long jobs without having to destroy and restart the job. Third, honesty as a measured feature: 4.8 is about a quarter as likely as 4.7 that a flaw in the code itself will go unflagged and its uncertainties surface more easily. When you stack these things up and run the model unattended for hours on end, you discover what actually matters. That is, the model cannot plan, parallelize, check its own work, and read all the differences, so it is trained not to trust itself. It also burns tokens by doing everything with fists.

Next, pay attention to where the money is flowing. You can see that the account unit is a token. OpenRouter raised $113 million for $1.3 billion by doing something almost embarrassingly simple: routing across over 400 models and capturing up to 5% of the inference spend that passes through it. The weekly throughput increased from 5T tokens to 25T tokens in 6 months, a 5x increase. It’s not a prediction. It’s a meter. Cognition raised $1 billion for $26 billion, but embedded in that announcement was a line that said they needed to reorganize what they had done so far. 89% of code committed within Cognition is now written by Devin, up from 13% in December. Run-rate revenue increased from $37 million to $492 million over the year. Autonomous Software Engineering has stopped being a demo and has become the default committer.

Snowflake closes the loop on the public side. Product revenue increased 34%, guidance was raised, and inventory increased ~36% in a single session. Two of them are the $6 billion AWS compute deal and the acquisition of Natoma, an MCP platform for managing agent access. The data tier is re-pricing around agents. consumenot the analyst doing the query. The entire stack (model, router, agent, board) is converging into one business model. In other words, it’s billing with tokens because tokens are work.

That is precisely the moment Pope Leo XIV chose to publish it. Magnifica Humanitashis first encyclical presented with Anthropic’s Chris Oller. Stripped of its theology, this argument is a critique of engineering that the field should take seriously. Technology is never neutral. Because technology inherits the incentives of those who build and fund it. And the danger is not malicious intent, but quiet disintermediation, as decision-making slips out of human hands, one delegated commit at a time.

Make sure you grasp these two facts. 89% of the recognition is an indexical restatement of the thesis of the encyclical. The meter that makes economics work is the same meter that measures how much judgment we delegate. The bull case and the moral case read the same numbers.

The flywheel is no longer a slide. It’s on the income statement. The open question is: what are we optimizing for?

AI lab: CMU and Amazon

summary: To address the challenge of assessing whether AI agents can accurately translate informal programming intentions into formal specifications, researchers introduced the VERUS-SPECBENCH benchmark and the VERUS-SPECGYM agent environment. By extending the execution mechanism to test the generated specifications against both formal tests and adversarial “hacks,” this study reveals that automatic formalization of specifications is still highly vulnerable even for models that can generate correct code.

AI lab: Tsinghua University, NVIDIA, University of Toronto, Vector Institute

summary: Gamma-World provides a scalable and generative multi-agent world model that goes beyond traditional single-agent simulation by leveraging simplex rotary agent encoding for permuted symmetric IDs and sparse hub attention for efficient inter-agent communication. The framework enables real-time action response rollouts at 24 FPS through teacher and student conditional distillation and KV cache streaming, maintaining strong consistency across virtual game environments and physical robot environments.

AI lab: Harvard University & MIT

summary: Bidirectional evolutionary search (BES) overcomes the limitations of sparse validation signals and narrow autoregressive extensions by combining forward candidate evolution and backward target decomposition. By recombining trajectory segments to avoid narrow probability distributions and scoring against fine-grained subgoals, BES significantly outperforms existing open-source frameworks on complex logical reasoning and open problem-solving tasks.

AI lab:Meta AI

summary: MobileMoE introduces a family of less than 1 billion active parameters Mixture-of-Experts (MoE) language models that are specifically optimized for efficient deployment on edge devices such as smartphones. Guided by new on-device scaling laws and supported by custom fused MoE kernels, these models achieve state-of-the-art performance while delivering significantly faster prefill and decode speeds compared to dense baselines of similar memory footprint.

AI lab: Google Deep Mind

summary: Gemini Embedding 2 is a native multimodal embedding model that seamlessly maps text, image, audio, and video inputs into a single unified representation space without relying on intermediate transcription. Trained via extensive contrastive learning in a multi-task setting, the model establishes new state-of-the-art performance across unimodal, cross-modal, and multimodal search benchmarks, while demonstrating robust zero-shot generalization across diverse enterprise and professional domains.

summary: The MiniMax-M2 series introduces a highly efficient 22.99 billion parameter expert mixture model that activates only 9.8 billion parameters per token, and is specifically designed for complex and long-running agent workflows. By leveraging an agent-driven data pipeline, a specialized reinforcement learning system called Forge, and autonomous self-evolution capabilities, the model achieves frontier-level performance across coding, deep search, and inference benchmarks while maintaining a minimal computational footprint.

human We have released a new version of the marquee modelwith powerful agent and coding capabilities.

1. Anthropic raises $65 billion in Series H at $965 billion post-money valuation — Anthropic raised $65 billion in Series H funding (co-led by Altimeter, Dragoneer, Greenoaks, and Sequoia) at a post-money valuation of $965 billion, revealed earlier this month that run-rate revenue exceeded $47 billion, and previously committed $15 billion in hyperscaler investments (including $5 billion from Amazon) in addition to Micron, Samsung, and SK Hynix as strategic memory/storage partners.

2. Cognition raises $1 billion at $25 billion pre-money valuation — Cognition, creator of AI software engineer Devin, raised more than $1 billion (led by Lux Capital, General Catalyst, and 8VC) at a post-money valuation of approximately $26 billion, and reached an annual revenue run rate of $492 million, more than doubling in eight months.

3. Robinhood allows AI agents to trade stocks — Robinhood launches Agentic Trading and Agentic Credit Card in beta. This allows customers to connect a third-party AI agent to another funded account (via MCP) and trade and buy stocks autonomously. Original source:

4. OpenRouter’s valuation doubles to $1.3 billion — The multi-model AI inference routing startup raised $113 million in Series B led by Alphabet’s CapitalG, at a valuation of approximately $1.3 billion, more than double the level from a year ago as weekly trading volume increased from 50 million tokens to 250 million tokens.

5. Hark raises $700 million in Series A — Brett Adcock’s secretive AI startup has raised $700 million (led by Parkway Venture Capital) at a $6 billion post-money valuation to build a “universal” agent AI assistant with a unique multimodal model and custom hardware, with the first model expected to be released in summer 2026.

6. Mistral signs deals with Airbus and BMW — Mistral AI expands into “Physical AI” for manufacturing, announcing partnerships and new data center in France to apply its models to Airbus (aircraft design, flight safety, defense/space) and BMW’s “Large Industrial Model” crash simulation initiatives.

7. MiniMax doubles sales ahead of new model — China’s AI developer’s annual revenue more than doubled in two months to at least $300 million, driven by a five-fold increase in enterprise users ahead of the launch of the M2.7 model and its next flagship product. There is no primary source of information that can be replaced. The numbers come from a Bloomberg TV interview with co-founder Yun Yei, so Bloomberg is the original source.

8. SK Hynix joins the trillion dollar club — Driven by HBM’s AI demand, the South Korean memory maker’s stock price rose about 9% to 15%, joining rivals Samsung and Micron to surpass $1 trillion in market value for the first time. There are no company announcements regarding stock price milestones. This event was brought to you by Reuters: https://money.usnews.com/investing/news/articles/2026-05-26/sk-hynix-joins-1-trillion-club-after-samsung-micron-on-ai-chip-boom

9. Pope Leo warns that AI should not rule humanity — in his first encyclical, Magnifica HumanitasPope Leo Original source: The encyclical itself issued by the Vatican (vatican.va) — this is the document on which the coverage is based.

10. Snowflake signs $6 billion AWS deal for Graviton chips — Snowflake commits $6 billion over five years to AWS. This is our largest infrastructure commitment in history, expanding our use of Amazon’s ARM-based Graviton CPUs and GPUs to power agent-based AI workloads.