The day broke into two parallel conversations. One was about frontier capability: OpenAI reshaped Codex into a Cowork-class desktop agent, shipped a domain-specific life-sciences model, and moved closer to federal deployment of Anthropic's Mythos. The other, quieter conversation was about whether people actually know how to use any of this — and a controlled experiment with Microsoft Research landed hard on the answer.
- Codex becomes a Cowork competitor. OpenAI expanded Codex from a coding agent into a full desktop workstation with Mac-level computer use, parallel agents, an in-app browser, and 111 new plugins.
- Perplexity drops a 24/7 Mac agent. Personal Computer reads and writes local files and drives iMessage, Mail, and Calendar, rolling out to Max subscribers.
- OpenAI ships a life-sciences model. GPT-Rosalind is the second domain-specific model in three days after GPT-5.4-Cyber, reinforcing a shift away from the one-model-for-everything approach.
- Washington moves to access Mythos. The White House is preparing to give federal agencies use of Anthropic's most powerful model despite the outstanding blacklist dispute.
- Sequoia's services thesis. Services — not software — may be where the next trillion-dollar company gets built, and the implications for SaaS pricing and product strategy are immediate.
- Microsoft Research puts hard numbers on AI training. A controlled experiment at Gap Inc. found that behavioral training doubled the rate of high-quality AI-generated output versus standard feature training.
- AI concierges move into physical retail. Puma, Guitar Center, Vitamin Shoppe, and smart fitting rooms at Victoria's Secret signal a shift from online chatbot to in-store assistant.
- A pro tip worth adopting. Nested CLAUDE.md files at every folder level let Claude Code auto-load business context on every session.
If there is a thread connecting today's items, it is that the gap between what these tools can do and what organizations get out of them is now measurable — and much of the gap has nothing to do with the tools.
Developments
Technology
April 16: OpenAI rebuilds Codex as a Cowork-class desktop workstation
On the same day Anthropic released Opus 4.7, OpenAI shipped a sweeping Codex update that reframes the product as OpenAI's answer to Claude Cowork. The new Codex runs Mac apps in the background via computer use, runs multiple agents in parallel, adds an in-app browser for agent guidance, introduces persistent memory and multi-day automations, and adds 111 new plugins including Atlassian Rovo, CircleCI, and Microsoft Suite. OpenAI disclosed that Codex now has 3 million weekly users growing 70% month over month, and Codex App Lead Thibault Sottiaux framed the direction as “building the super app out in the open.”
So what
The agentic coding market is now a two-horse race with compounding switching costs for any company that standardizes on a single provider. Growth-stage engineering leaders should run a parallel evaluation this quarter rather than default to the incumbent, and CFOs should model a provider-diversification scenario given Anthropic's simultaneous move to consumption-based Enterprise pricing (covered in Edition #8). The harder question for SaaS CEOs is downstream: when both labs ship desktop agents with native browsers, persistent memory, and third-party plugin ecosystems, the category of standalone productivity SaaS gets squeezed from above.
Confidence: Established fact — OpenAI's own announcement, corroborated by independent reporting at TechCrunch, The Deep View, and The Rundown.
Go deeper OpenAI announcement · The Rundown AI
Technology
April 16: Perplexity launches Personal Computer, a Mac-native 24/7 agent
Perplexity released Personal Computer, a Mac app that connects to local folders, reads and writes files, and drives native applications including iMessage, Mail, and Calendar. The agent runs across 20-plus frontier models and can operate continuously, not just in response to prompts. It is available to Max subscribers and the waitlist, and it pairs with the company's existing Comet browser to extend agentic reach from the web to the desktop.
So what
Three of the top AI companies now ship desktop agents with native OS integration (Claude Cowork, OpenAI Codex, Perplexity Personal Computer), and the second-order effects on enterprise procurement are starting to matter. Expect IT and security teams to ask harder questions about what these agents read, write, and retain at the file-system level. For growth-company product teams, the near-term consequence is that the assumption “our customer will open our app to do this task” is weakening; if the task is boring or infrequent, the agent will do it without opening anything.
Confidence: Established fact — Perplexity's product announcement, multiple independent confirmations.
Go deeper Perplexity Personal Computer · The Deep View
Competition
April 16: OpenAI releases GPT-Rosalind, its second domain-specific model in three days
OpenAI released GPT-Rosalind, a life-sciences model purpose-built for drug discovery and biological research, in research preview with Moderna, Amgen, the Allen Institute, and Thermo Fisher. It ships alongside a Life Sciences plugin connecting to more than 50 scientific databases. On blind RNA prediction tasks from Dyno Therapeutics, the model reportedly scored higher than 95% of human experts. Coming three days after GPT-5.4-Cyber, it marks the second specialized model OpenAI has released this week, and access is gated to approved research organizations.
So what
The pattern is more important than either model: the frontier labs are beginning to ship gated, vertical-specific models rather than pushing the one general model further. For growth-company CEOs, this has two implications. First, the defensibility of vertical SaaS is getting more complicated, not simpler: the lab-built vertical model and the SaaS application layer on top of it are now two separate strategic questions. Second, access itself is becoming a moat — Rosalind is not publicly available, and that matters for any portfolio company whose thesis depends on compute or model access at the frontier.
Hype check
The “better than 95% of human experts” figure comes from a single blind RNA prediction test at one lab partner and should not be read as a general claim about scientific reasoning.
Confidence: Informed perspective — announcement-based, performance claims await independent validation.
Go deeper OpenAI announcement · The Deep View
Legal & RegulatoryDeveloping
April 16: White House moves to give federal agencies access to Anthropic's Mythos
Building on the federal-AI coverage running through recent editions, Bloomberg reported on April 16 that the White House is preparing to give select federal agencies access to Anthropic's Mythos model, even as the blacklist dispute between the company and the administration remains unresolved. The Information separately reported that Google is negotiating a classified Gemini deal with the Pentagon. Both moves come the same week OpenAI shipped GPT-5.4-Cyber to a wider audience of verified defenders.
So what
Federal procurement is doing what commercial procurement has not yet: accepting gated, most-capable-model access on defensibly narrow use cases, reportedly on terms that side-step the blacklist. For any growth company selling into regulated industries or federal-adjacent buyers, the signal is that compliance posture around frontier-model usage is about to get more stratified, not more uniform. Engage outside counsel now on what a “Mythos-class” use case looks like under existing contracts, rather than waiting for the rules to catch up.
Confidence: Informed perspective — Bloomberg sourced reporting, policy specifics still being finalized.
Go deeper Bloomberg · The Rundown AI
Case Studies
AI Strategy & Transformation
April 17: Microsoft Research and Gap Inc. study finds behavioral training nearly doubles AI output quality
A controlled experiment run by Microsoft Research with Gap Inc. and AI Mindset enrolled 388 employees, all with full Microsoft Copilot access, and split them into two training arms. The control group received standard feature-and-prompt training. The treatment group received a 30-minute behavioral session focused on the cognitive reasons people default to treating AI like a search engine. Independent blind raters then graded each group's output. The treatment group produced top-quality work 77% of the time, compared to 47% for the control group. Same tool, same task, same data, measurably different result.
So what
This is the study to cite when a board asks why another Copilot or Claude rollout is not producing measurable productivity gains. The gap between 47% and 77% is an adoption gap, not a tool gap, and it closes with 30 minutes of the right training. The practical move for growth CEOs this quarter: audit your current AI enablement program and ask whether it is teaching features or changing mental models. If the agenda looks like “here are the best prompts and use cases,” you are buying the control-group outcome. The cost of re-running the enablement properly is low; the cost of leaving doubled output quality on the table is not.
Confidence: Established fact — controlled experiment, independent blind grading, Microsoft Research-published paper (arXiv:2604.08678).
Go deeper Microsoft interactive data explorer · AI Mindset
Go-to-Market
AI concierges arrive in physical retail, and the underlying Q1 traffic data tells the real story
Puma is deploying Dylan, a seven-foot digital-human running-shoe expert that speaks more than 100 languages. Guitar Center launched Rig Advisor, a QR-code-activated shopping assistant. The Vitamin Shoppe is piloting Shoppe Advisor, an AI touchscreen with real-time inventory. Crave Retail's AI-powered smart fitting room is now live at Victoria's Secret, Under Armour, and Foot Locker. The context that makes these investments legible: Adobe data shows AI-driven web traffic to US retailers rose 393% year over year in Q1 2026, with chatbots like Amazon Rufus and Ask Macy's producing meaningfully higher conversion rates.
So what
Retail is the leading indicator for how mainstream customers will interact with AI in other contexts — and the signal is that the conversational interface is becoming expected, not novel. For growth-company CEOs selling to retail or CPG customers, the 393% Q1 traffic shift means your roadmap for in-application chat and agent-driven guided selling is now table stakes for 2026 RFPs. For B2B SaaS more broadly, the cross-read is that conversion-lift claims for AI-assisted onboarding or support are moving from anecdote to ad-hoc benchmark; start measuring these carefully so you have honest data before your customers ask.
Confidence: Informed perspective — pilot and early deployment data, limited long-run evidence on conversion economics.
Go deeper TechCrunch on Adobe Q1 data · Superhuman AI
