Anthropic: the model they were afraid to release

The main intrigue of the week is Claude Mythos, a new model from Anthropic that the company showcased but did not release publicly. The reason is unexpected: cybersecurity concerns. According to benchmarks, the model shows a significant improvement over previous versions, but that is precisely what became the problem — too powerful a tool without proper security guarantees.

Instead of a public release, Anthropic launched Project Glasswing — a closed program for about 40 organizations that will use Mythos to find vulnerabilities in critical software. In effect, the company turned a potential risk into a defensive tool.

The financial market reaction was telling: the US Treasury Department and the Federal Reserve urgently notified the heads of the largest banks about the cyber risks associated with this release. Skeptics accused Anthropic of hype — saying the best way to draw attention to a model is to announce that it is too dangerous. But the very fact of emergency warnings from regulators shows that the issue is being taken seriously.

What this means for the market: the precedent of “we built it, but won’t release it” may become a new norm for frontier models. The question is how long companies will be able to sustain such decisions under competitive pressure.

Model market: new players and new records

Z.ai GLM-5.1: China’s answer without Nvidia

The most technically interesting news of the week came from the Chinese company Z.ai. Their open model GLM-5.1 — a MoE architecture with 744 billion parameters (40 billion active) under the MIT license — took first place on the SWE-Bench Pro benchmark, beating GPT-5.4 and Claude Opus 4.6.

But even more important is something else: the model was trained entirely on Huawei chips, without a single Nvidia GPU. Under US export restrictions, this is not just a technical fact — it is a demonstration that Chinese AI development has not stopped. The open MIT license makes GLM-5.1 available for any use, including commercial use.

Meta Muse Spark

Meta introduced Muse Spark — the first model from the new Superintelligence Labs division. The model is proprietary and, according to benchmarks, sits between Claude Sonnet 4.6 and Opus 4.6. Meta's new division is actively recruiting researchers, poaching specialists from DeepMind and OpenAI.

OpenAI: new pricing plan

OpenAI launched Pro Lite for $100 per month — an intermediate option between Plus and Pro. Increased limits, 10x higher, are in effect until the end of May. The company is also testing Image V2 — the next generation of image generation, three variants of which have already appeared on the LM Arena platform.

Microsoft Harrier: SOTA for multilingual embeddings

Microsoft open-sourced Harrier — a multilingual embedding model under the MIT license that ranks first on MTEB-v2 across more than 100 languages. For RAG system developers, this is an important event: there is now a free SOTA model for semantic search in any language, including Russian.

AI agents are taking over dev infrastructure

Platform data shows that agentic development has ceased to be an experiment and has become a production-scale reality.

GitHub records a 14x increase in commits from agents. Over six months, the number of pull requests from AI agents has grown by 4.25x — with occasional platform overloads. Vercel says that more than 30% of all deployments today are made by agents, and 75% of them are Claude Code.

Against this backdrop, Anthropic released Managed Agents — a cloud environment for deploying long-lived agents with isolated sandboxes, checkpointing, and scoped permissions. At the same time, it published a description of the Advisor Strategy: Opus 4.6 acts as an advisor for Sonnet/Haiku, which reduces the cost of agent pipelines without sacrificing reasoning quality.

Telegram updated the Bot API, allowing bots to communicate with each other — technically, this opens up possibilities for building multi-agent systems directly inside the messenger.

AWS launched S3 Files — mounting S3 buckets as a file system via NFS with a latency of about 1 millisecond.

Big money: ARR, investments, manifestos

Anthropic surpassed OpenAI in revenue

Anthropic reached $30 billion ARR — up from $9 billion at the end of 2025. For comparison: OpenAI's ARR is $25 billion. The company also found itself at the center of a geopolitical game: amid the conflict with the Pentagon, London invited Anthropic to expand its presence in the UK and consider a dual listing on the stock exchange.

Perplexity: $450M ARR and banking data

Perplexity grew revenue by 50% in a month to $450 million ARR and launched Personal Finance based on integration with Plaid — real-time AI analysis of bank accounts and budgets. The product raises sharp privacy concerns, but demand is clearly there. In addition, the company announced a Billion Dollar Build hackathon with a prize of up to $1 million in investment.

Altman’s Manifesto

Sam Altman published a 13-page document on the economy in the age of AI. Key points: shift taxes from labor to capital, create a Public Wealth Fund from AI revenues, and test a four-day workweek. In essence, it is a call for a new social contract. The document has already sparked debate: some see it as genuine concern, others as an attempt to shape the narrative ahead of regulatory pressure.

The rest of the important news of the week

Generative video. Alibaba turned out to be behind the viral Happy Horse model (15 billion parameters, open-source, No. 1 on Video Arena). Runway launched Characters — real-time avatars based on its GWM-1 world model.

Google is not slowing down: Gemini received Notebooks projects with NotebookLM integration, learned to generate interactive 3D models directly in chat, and also released the free Eloquent AI dictaphone for iOS based on Gemma.

MemPalace and Karpathy. Actress Milla Jovovich released the open-source MemPalace tool based on the “memory palace” method — 23,000 GitHub stars in two days. Andrey Karpathy independently proposed a similar approach: knowledge bases with topic grouping, summarization, and backlinks. The topic of AI memory management is clearly gaining momentum.

Bitcoin and Satoshi. The NYT published an investigation, in which it used AI algorithms to compare Satoshi Nakamoto’s correspondence with Adam Back’s letters and found matches in grammar and stylistic constructions. Satoshi’s wallet holds about $78 billion.

The labor market: Jevons paradox in action

In the context of all of the above, an important social backdrop. According to Fortune data, 80% of white-collar workers are quietly sabotaging corporate AI adoption. Among Gen Z, that figure is 44%. People bypass corporate AI tools, use personal accounts, or simply ignore the new workflows.

At the same time, the number of job openings for programmers is now higher than at any point in the past three years — despite the fact that AI tools should, in theory, reduce the need for developers. This is the classic Jevons paradox: greater technological efficiency increases demand for it rather than reducing employment.

Key takeaways of the week

For businesses and leaders:

• The question of “whether to adopt AI” has shifted to “how to implement it so employees don’t sabotage it.” Change management is becoming more critical than technology.
• Anthropic’s Managed Agents is the first enterprise-ready environment for production agents with proper governance. Worth studying.
• Z.ai’s GLM-5.1 is the first open frontier-level model. For Russian companies under restrictions, this is an interesting alternative.

For developers:

• Advisor Strategy (Opus advises Sonnet) — a ready-made pattern for reducing the cost of agent pipelines.
• Microsoft Harrier is the best open model for multilingual embeddings, relevant for Russian RAG.
• The Telegram Bot API now supports bot-to-bot communication — multi-agent systems in Telegram have become easier.

For those following the trends:

• “We built it, but won’t release it” (Claude Mythos) — a new type of event in the AI market. The frontier of model capabilities is beginning to outpace society’s readiness to use them.
• Anthropic has overtaken OpenAI by ARR. This changes the balance of power and intensifies competition.

FAQ

What is Claude Mythos and why wasn’t it released?

Claude Mythos is a new language model from Anthropic that showed significant benchmark gains. The company decided not to release it publicly due to cybersecurity concerns. Instead, a closed Project Glasswing program has been launched for around 40 organizations that use the model to find vulnerabilities in critical software.

What is GLM-5.1 from Z.ai and what makes it notable?

GLM-5.1 is an open (MIT) MoE model with 744 billion parameters from the Chinese company Z.ai. It stands out for two things: first place on SWE-Bench Pro (programming) and full training on Huawei chips without using Nvidia GPUs. It is the first Chinese model to top this benchmark.

What is Anthropic’s Advisor Strategy?

An architectural pattern for agent systems: a more powerful model (Opus 4.6) acts as an advisor for a faster and cheaper one (Sonnet or Haiku). Opus analyzes the task, formulates a strategy or warnings, and Sonnet executes. It makes it possible to reduce costs without sacrificing reasoning quality.

Is it true that 80% of office workers are sabotaging AI?

According to Fortune (April 2026), yes — 80% of “white-collar workers” are bypassing corporate AI tools in one form or another. Among Zoomers, the figure is somewhat lower — 44%. The reasons vary: concerns about jobs, distrust of corporate oversight, and inconvenient tools.

Has Anthropic really overtaken OpenAI?

By ARR (Annual Recurring Revenue), yes. Anthropic reached $30 billion ARR, while OpenAI reached $25 billion. At the same time, OpenAI remains larger in absolute revenue and in the number of ChatGPT users.