Ship week, and a leap toward security

Big launches from Anthropic, OpenAI, and Microsoft, and the AI labs lean into security.

Plus, the fight for “default AI” is here.

Here’s the rundown:

1. Claude rolls out Marketplace to simplify enterprise AI procurement

Anthropic introduced Claude Marketplace, allowing enterprise customers to apply their existing Anthropic spending commitments to Claude-powered third-party tools like GitLab and Snowflake without separate procurement. Anthropic’s pitch: give customers the flexibility to consolidate AI spend across various vetted partner tools. Early partners include GitLab, Harvey, Lovable, Replit, Snowflake, and Rogo, a genAI platform for finance. Marketplace is in limited preview, currently available to select Anthropic enterprise customers.

Why this matters: Marketplace is Anthropic’s procurement and distribution play for enterprise lock-in. By letting customers spend their existing Anthropic commitments across a partner ecosystem, Anthropic makes itself harder to churn and gives Claude-powered partners a direct pipeline to customers already in the stack. The parallel to OpenAI's agentic commerce push is telling: while OpenAI is embedding itself into consumer transactions through PayPal's checkout infrastructure, Anthropic is embedding itself into enterprise procurement.

2. OpenAI launches GPT 5.4 for professional work

OpenAI released GPT-5.4, available in ChatGPT (as GPT-5.4 Thinking), the API, and Codex. The model consolidates OpenAI's recent advances in reasoning, coding, and agentic workflows into a single model, bringing in GPT-5.3-Codex’s coding capabilities and stronger performance on spreadsheets, presentations, and documents. On GDPval, OpenAI’s benchmark evaluating knowledge work across 44 occupations, GPT-5.4 matches or exceeds industry professionals in 83% of comparisons, up from 70.9% for GPT-5.2. This is OpenAI's first general-purpose model with native computer-use capabilities, enabling agents to operate computers, and its most token-efficient reasoning model yet. A Pro variant is available for users who need maximum performance on complex tasks.

Why this matters: As frontier models continue to multiply, consumers might be building model rosters, deploying different models for different jobs. Reporting on a poll of his Substack audience, Alex Banks wrote that the consensus was that GPT-5.4 is a "workhorse, not a thoroughbred"— faster and more usable than prior GPT models, but still inferior to Opus 4.6 for deep research and complex reasoning. One user summed it up neatly: GPT-5.4 is what you use to "save your precious Opus tokens, not replace them.” The competitive dynamic is shifting from "which model wins" to "which model for which job."

3. Microsoft releases Copilot Cowork to manage workflows

Microsoft launched Copilot Cowork, an agentic feature within Microsoft 365 Copilot. Cowork converts a user prompt into a workflow, grounded in Microsoft 365 context (email, meetings, files, data), and runs it in the background with checks for user approval at key steps. Use cases include calendar triage, meeting prep, company research, and product launch planning. It is currently in research preview, with broader availability expected later this month. Microsoft says Copilot includes technology from Claude Cowork, reflecting its push toward a multi-model approach.

Why this matters: Copilot Cowork is a snapshot of where enterprise AI might be headed: up the org chart in how it’s sold, and multi-model under the hood. When we started this Digest, the AI labs and big tech were speaking of AI teammates and coworkers. Now, Microsoft has shipped something closer to an executive delegate, an agent that manages workflows around tasks. Notably, this executive uses Claude technology. As of 2025, Microsoft holds a ~$135 billion stake in OpenAI, yet chose Claude to power this product – a reminder that in enterprise AI, platforms don't have to be loyal to one model, and the best model for the job can still get the call.

In other news

1. Anthropic sues U.S. government over supply chain risk designation (Bloomberg)
2. OpenAI to acquire AI security platform Promptfoo (OpenAI)
3. Anthropic and Mozilla partner to enhance Firefox security (Anthropic)
4. OpenAI exec leaves company citing Pentagon deal (Politico)
5. AI service agent Fin raises $250M to build customer agent (Eoghan McCabe)

Trending: AI's labor market impact in the U.S.

Anthropic’s new study on the labor market impacts of AI finds no effect on unemployment rates for workers in the most exposed occupations, though there’s “tentative evidence” that hiring has slowed slightly in the professions for workers aged 22-25. The study introduces a new AI displacement risk metric: “observed exposure.” This metric combines theoretical LLM capability and real-world, work-related usage of the model. The most exposed roles under this metric include computer programmers, customer service representatives, and financial analysts. In the most exposed professions, workers tend to be older, female, more educated, and higher-paid. The report also flags that AI usage currently covers only a fraction of what’s feasible, given AI’s theoretical capability.

Bonus: the fight for “default AI”

In its new top 100 GenAI consumer apps report, a16z argues that the AI race is shifting from who has the most users to who is hardest to replace.

The report elaborates that the AI labs are creating lock-in through context, personalization, and connector ecosystems. But for the most part, the labs are placing distinct bets. ChatGPT is aggressively expanding its app ecosystem in consumer (travel, shopping, food, health), while Claude is focusing on professional integrations (PitchBook, Snowflake, Databricks).

A16z summarizes how this fight could play out: “If the AI assistant becomes not just a chat window but an operating environment, this race may end up looking less like the search wars — where one player took 90% of the market — and more like the mobile OS wars, where two platforms with very different philosophies both built trillion-dollar ecosystems.”

5 new AI tools to try

1. Codex Security, security agent in research preview – learn more here (OpenAI)
2. Uni-1, reasoning and image generation model – try here (Luma AI)
3. autoresearch, open-source repo for agentic LLM training experiments – try here (Andrej Karpathy)
4. Loop command, feature to schedule recurring tasks on Claude Code – try here (Anthropic)
5. Scheduled tasks, feature to create task schedule on Claude Code desktop – try here (Anthropic)

Recommended reading

1. How retail banks can put AI agents to work (Stiene Riemer, Marc Pauly, Bharat Poddar, Dan Sack)
2. The Auton agentic AI framework (Sheng Cao, Zhao Chang, Chang Li, Hannan Li, Liyao Fu, Ji Tang)
3. Diagnosing retrieval vs. utilization bottlenecks in LLM agent memory (Boqin Yuan, Yue Su)
4. Bayesian teaching enables probabilistic reasoning in Large Language Models (Linlu Qiu, Fei Sha, Kelsey Allen, Yoon Kim, Tal Linzen, Sjoerd van Steenkiste)
   

See you next week!