Claude Design, Factory $1.5B and Mythos goes federal
Alexandre
··
Reading time: 17 min
This week I had a strange moment: I read the Mythos memo twice to make sure I understood it right. The same model declared too dangerous for the public… plugged into the DoD two weeks later. The rest of the week matched that energy.
In a single day, OpenAI and Anthropic each dropped their major model simultaneously — a classic benchmark war. The day after, Anthropic went again with Claude Design, a visual prototyping tool powered by Opus 4.7. But the story that actually occupied my mind was something else: Bloomberg reveals that Mythos, the model Anthropic just declared too dangerous for the public, is about to be handed to the DoD and Homeland Security. Add Factory raising $150M at $1.5 billion with Morgan Stanley as a client, and Google quietly launching Gemini for Mac with data access on by default — I build Waku with Claude Code, and all five stories touch my tools directly.
The week's radar
Event
Date
Actor
Key figure
Source
Factory raises $150M at $1.5B
Apr. 16
Factory
Morgan Stanley, EY, Palo Alto Networks as clients
TechCrunch
OpenAI Codex Mac — targets Claude Code
Apr. 16
OpenAI
111 new plugins, integrated computer use
The Verge
Claude Opus 4.7 retakes SWE-bench lead
Apr. 16
Anthropic
64.3% SWE-bench Pro vs 57.7% GPT-5.4
VentureBeat
Claude Design launches in research preview
Apr. 17
Anthropic Labs
Prototypes, slides, designs — export to Canva/PPTX, Claude Code handoff
Anthropic
Barbaccia memo — US agencies on Mythos
Apr. 7 (revealed Apr. 16)
White House / Anthropic
6 federal agencies + 3 Congressional committees
Bloomberg / Reuters
Gemini Mac launched + Google I/O announced
Apr. 15–19
Google
I/O on May 19, 2026
9to5Mac / TechCrunch
Sources: TechCrunch, The Verge, VentureBeat, Bloomberg, Reuters, 9to5Mac
Factory raises $150 million at $1.5 billion: enterprise AI coding agents have found their champion
Factory, $150M raised on April 16, $1.5 billion valuation. Khosla leading, Sequoia, Insight Partners, and Blackstone in. Clients: Morgan Stanley, EY, Palo Alto Networks. The company, founded in 2023 by Matan Grinberg, builds autonomous coding agents for enterprises.
Morgan Stanley and EY are not tech early adopters. They are organizations with compliance and audit trail requirements that consumer coding tools simply do not cover. When companies like that are paying for coding agents in production, the market has shifted from pilot to recurring spend.
Definition — Enterprise coding agent: an AI agent specialized in generating and modifying code at organizational scale, with specific constraints: compliance with internal security policies, CI/CD chain integration, action auditability, and multi-team access rights management. Distinct from consumer tools (Claude Code, Codex) by the need to adapt to existing workflows rather than replace them.
The $1.5 billion valuation for a company founded in 2023 is a bet on potential, not on proven revenue. TechCrunch notes that Blackstone is in this round. Blackstone is not a typical VC. When an alternative asset manager backs a coding agent startup, the thesis that "AI agents will replace entire developer teams" has moved from a pitch deck to a capital allocation conviction.
Palo Alto Networks as a client is also worth a pause. Security is a space where code quality has direct consequences on attack surface. The fact that a security firm trusts an external agent to modify its codebase says something about the operational maturity Factory has reached.
My take as a solo dev
I build Waku alone, with Claude Code at €100 a month. I do not need audit trails or SAML integration. Factory's product is not for me.
What this actually says: the tools I use and the tools Morgan Stanley uses are no longer on the same trajectory. Two markets, two pricing models, two roadmaps. For me that is actually good news — competition on the consumer segment stays strong.
What concerns me over the next 18 months: if big enterprises capture a growing share of lab compute through enterprise contracts, will labs start optimizing for their use cases at the expense of solo workflows? Not immediately. But it is a dynamic worth tracking.
And on the same day as the Factory raise, OpenAI was shipping a very direct response to Claude Code.
Click to enlarge
OpenAI Codex Mac: computer use, integrated browser, and 111 plugins targeting Claude Code
Codex Mac launched on April 16. This is not a simple desktop port: background computer use, integrated browser, image generation via gpt-image-1.5, cross-session memory, and 111 new plugins including GitLab, the Atlassian suite, and Microsoft Office. The Verge headlined it as "a direct shot at Claude Code".
The features listed map precisely to Claude Code's known weak spots. Cross-session memory is the most-cited gap: every session starts fresh with no accumulated context. Background computer use lets Codex run tests, verify builds, and interact with GUI apps while you work on something else. The 111 plugins specifically target teams not using GitHub.
Definition — Computer use: an AI agent's ability to control a computer's graphical interface as a human would — clicking, typing, switching windows, reading screen state. Introduced by Anthropic with Claude 3.5 Sonnet in October 2024, it allows agents to automate tasks that require interacting with software that has no API.
The timing is calculated. Anthropic had just shipped Opus 4.7 the same day. OpenAI chose to launch Codex Mac simultaneously, capturing media coverage and forcing direct comparisons. Engadget and The Verge both covered both announcements the same day.
The image generation feature via gpt-image-1.5 inside a coding tool reveals a different strategy. OpenAI is not building a coding tool. It is building a complete development assistant that also covers visual assets, mockups, and icons. An indie dev building an app with Codex can generate the interface, illustrations, and code from a single tool. Claude Code stays narrower in its specialization — and that may be a deliberate choice.
Ars Technica ran a deep test of Codex Mac's computer use: the agent can launch Xcode, wait for compilation to finish, read the errors, and fix the code — without interruption. On native iOS workflows, that is concrete progress.
My take as a solo dev who uses Claude Code every day
This release makes me look at my setup differently.
Cross-session memory is where Claude Code costs me the most time. On Waku, my SwiftUI codebase now runs several thousand lines across dozens of files. Every new Claude Code session needs a context recap I've semi-automated via a CLAUDE.md file — real friction. If Codex maintains context automatically across sessions, that is a meaningful quality-of-life improvement.
The integrated computer use for iOS builds is also what I want. How I use Claude Code to develop Waku still involves a lot of manual back-and-forth with the simulator and Xcode. An agent that can run the build, read compilation errors, and fix them in a loop without me copy-pasting logs: that is the logical next step.
I am not switching tomorrow. Claude Code's architectural understanding of complex codebases, its ability to reason about the implications of a change across multiple files at once, is what convinced me over time. But if Codex Mac delivers on memory and computer use, the distance between the two tools narrows.
Click to enlarge
Claude Opus 4.7: Anthropic retakes the SWE-bench top spot at 64.3%
Opus 4.7 shipped on April 16. 64.3% on SWE-bench Pro, versus 57.7% for GPT-5.4. 3x more difficult production tasks solved versus Opus 4.6, visual resolution 3x better. This is clearly Anthropic's answer to the Opus 4.6 regression criticism.
Definition — SWE-bench Pro: the professional version of the SWE-bench (Software Engineering Benchmark), designed to measure an AI model's ability to resolve real GitHub issues on production open-source repositories. Unlike the standard SWE-bench, SWE-bench Pro excludes issues present in models' training data, making it harder to memorize and more representative of real-world performance.
A 6.6-point gap on SWE-bench Pro is not trivial: we are talking about real GitHub issues, in production codebases, with tests to pass. That is Claude Code's workflow in real conditions. VentureBeat notes the gains are especially strong on tasks the benchmark classifies as "hard": bugs that require cross-file understanding or architecture-level modifications.
The visual improvement (3x) is less spectacular on paper but has practical implications. On workflows that include error screenshots, UI mockups, or architecture diagrams, better vision means fewer hallucinations on details.
The Information reports that the improvement on "production complexity" tasks — modifying multiple files with cross-dependencies — is the most meaningful metric for engineering teams using Claude Code in production. The 3x factor on that criterion is the most concrete data point in this release.
This pattern — reported regression, fix in the next version — is reassuring about Anthropic's ability to listen to power users. Even if a one-month cycle between degradation and fix is long for a daily development tool.
My take as a solo dev
I read SWE-bench numbers with caution. They measure isolated GitHub issues in a controlled environment. My actual Claude Code usage on Waku looks more like "build this feature, maintain consistency with existing components, make sure it compiles on Xcode 16.3, respect the SwiftUI conventions I've been using for six months." That is a much more contextual benchmark by default.
That said, the improvement on hard tasks says something real: Anthropic worked specifically on cases requiring deep codebase understanding. That is exactly where Claude Code was strong before the 4.6 regression.
What actually interests me: if Opus 4.7 genuinely handles complex tasks 3x better, I can send more ambitious prompts in a single pass. Fewer iterations, fewer tokens. I am testing this week on the iOS features in progress — and if you want to cut your Claude Code token usage in half, now is a good time to revisit your prompts.
Claude Design: Anthropic moves into visual prototyping
On April 17, 2026, Anthropic Labs launched Claude Design in research preview for Pro, Max, Team, and Enterprise subscribers. The tool lets you create designs, interactive prototypes, slides, and one-pagers in conversation with Claude — powered by Opus 4.7. Export goes to Canva, PDF, PPTX, or HTML. Claude Code handoff happens in one click via a packaged bundle.
The pitch is straightforward: describe what you need, Claude generates a first version, you refine through conversation, inline comments, or adjustment sliders generated by Claude itself. The tool reads your codebase or design files at setup to build a design system consistent with your brand.
Definition — Research preview: a limited launch phase where a product is available to a subset of existing users (here, paying subscribers), with features that are incomplete or may evolve. Anthropic uses this format to collect real-world feedback before a wider release.
What distinguishes Claude Design from existing tools is the Claude Code handoff. When the prototype is ready, Claude Design generates a bundle you pass to Claude Code in a single instruction to implement. No copying and pasting specs between a design tool and a code tool — context transfers directly.
Anthropic names Datadog, Brilliant, and Canva as early partners. Datadog notes that what used to take a week of back-and-forth between briefs, mockups, and review rounds now happens in a single conversation. Canva is integrated as a native export — you create in Claude Design, you export to Canva for collaborative production.
The stated target is explicitly non-designers: founders, product managers, and marketers who have visual ideas but no Figma background. And designers who want to explore more directions without spending time producing them manually.
First: marketing assets. Landing page, App Store visuals, social posts. I currently spend a lot of time in design tools for tasks that do not require advanced Figma skills. If Claude Design handles that coherently with my design system, that is time recovered.
Second: prototyping new iOS features before coding them. Describe an interface in plain language, get an interactive prototype in minutes, send it to users for feedback, then pass the bundle to Claude Code to implement. That is the workflow I have been trying to optimize for months with patched-together solutions.
The real question: does the output hold up on complex mobile interfaces with different states, SwiftUI animations, and reusable components? The testimonials are all about web and slides. I am testing this week.
Click to enlarge
Claude Mythos enters US federal agencies: the Barbaccia memo
On April 16, 2026, Bloomberg and Reuters revealed that federal CIO Gregory Barbaccia signed a memo on April 7 preparing access for six US federal agencies to Claude Mythos: the DoD, Treasury, Commerce, Homeland Security, Justice, and State. Before this memo, two unnamed agencies and three Congressional committees had already contacted Anthropic directly. The deployment would go through Anthropic's existing FedRAMP agreement.
This is the story of the week. The week of April 10, Anthropic had announced via Project Glasswing that Mythos Preview had discovered thousands of zero-day vulnerabilities in testing — a 27-year-old OpenBSD bug, a 16-year-old FFmpeg flaw. Too dangerous to release publicly. Access limited to 40+ vetted partners including Amazon, Apple, Microsoft, and CrowdStrike.
Now that same model is about to be plugged into the Department of Defense and Homeland Security.
Definition — FedRAMP: Federal Risk and Authorization Management Program. A US government program that standardizes security assessments of cloud services used by federal agencies. FedRAMP accreditation means the provider has passed a rigorous security audit and meets federal data protection requirements. Anthropic had already obtained this accreditation for its commercial models, which simplifies deploying Mythos across agencies.
Reuters details the specific use cases: vulnerability analysis in critical infrastructure systems, monitoring state-level cyber threats, and assisting intelligence analysts on massive document corpora. Not general productivity. Use cases where Mythos's offensive cybersecurity capabilities would be deployed defensively.
Bloomberg reports the decision was facilitated by Mythos's performance in red team exercises organized by CISA (Cybersecurity and Infrastructure Security Agency) in late March. During those exercises, Mythos reportedly identified attack vectors on energy infrastructure systems that human teams had not spotted after six weeks of work.
The timeline is instructive. April 7: Barbaccia signs the memo. Same day: Anthropic announces Project Glasswing. Glasswing is the structure that lets Anthropic deploy Mythos to government actors while maintaining the appearance of multi-party governance. The partner list now looks like a pre-approved list for federal contracts.
My take as a solo dev
This story makes me uncomfortable.
Two weeks after declaring Mythos "too dangerous for the public," Anthropic is about to plug it into the DoD and Homeland Security. Both institutions have a definition of "defensive" that does not necessarily line up with mine.
What bothers me about Anthropic's reasoning is the "in good hands" logic. Glasswing is 40 private companies. Federal access means agencies with documented histories of surveillance abuse (NSA, FBI). "Defensive" in the context of Homeland Security can include use cases that would be classified as offensive anywhere else.
I am not claiming to know what Anthropic should do. But I notice that AI labs consistently navigate between two contradictory positions: "we build for the benefit of humanity" and "we take government contracts for military and national security use cases." Those two positions can coexist — but only with radical transparency on the guardrails. Glasswing is not that.
I use Claude Code to build Waku. Nothing changes in my day-to-day. But knowing that the same lab selling me a dev tool is also signing DoD contracts feeds into my thinking about single-vendor dependency — and about the AI code agent architecture I want to build long-term.
Click to enlarge
Google Gemini Mac and I/O 2026: Mountain View's quiet counter-programming
On April 15, 2026, Google launched the native Gemini app for macOS, with Personal Intelligence on by default — access to your local documents, calendar, emails, and web browsing. In the same move, Google announced Google I/O 2026 for May 19: Gemma, Android 17, Firebase agent-native (an SDK to build AI agents directly in Firebase), and vibe-coding tools. According to 9to5Mac, Gemini for Mac was the third most downloaded app on the Mac App Store within 12 hours of launch.
Google had signaled Gemini Mac's availability for weeks. But Personal Intelligence on by default is a product decision that takes a clear stance. Most AI assistants ask users to explicitly enable data access permissions. Google inverted the logic: access is on, it is up to you to turn it off.
Definition — Personal Intelligence (Google Gemini): a feature that gives Gemini access to the user's personal data stored in Google services (Gmail, Calendar, Drive, Photos) and on the local device, to provide responses contextualized to their real situation. Example: "what do I have scheduled this week?" or "summarize the emails I received from this client last month." The Mac version also includes access to local files with macOS system permissions.
The I/O 2026 announcements deserve attention for independent developers. Firebase agent-native is the most concrete: Google is preparing an SDK to build AI agents that run directly in Firebase infrastructure, with native access to Firestore, authentication, and Cloud Functions. For a solo dev using Firebase for their app's backend, that is potentially a way to build agents without setting up your own infrastructure.
TechCrunch reports that I/O will include a dedicated session on vibe-coding tools. Google devoting a full keynote to the topic confirms that vibe-coding has left the solo dev community and become a full market segment.
9to5Google notes a native Xcode integration: Gemini can be invoked from Xcode via an extension with access to the current project context. That is a direct response to the Claude Code / Xcode integration many indie iOS devs rely on.
My take as a solo dev
While OpenAI and Anthropic are fighting over benchmarks, Google ships a well-executed native Mac app and turns on access to your local data by default. Different strategy — less noise, more quiet infiltration into the daily workflow.
The native Xcode integration interests me more than the rest. I use Claude Code via terminal because it is the most frictionless integration in my current workflow. If Gemini can be invoked directly from Xcode with the project context, that is a friction-reduction argument that could convince devs who do not want to leave their editor.
Google accesses your local files and emails by default. For a company whose business model is targeted advertising, I would rather opt in myself than have to opt out after the fact.
Other AI news this week
News
What happened
Why it matters
Sequoia raises $7B
Sequoia Capital raised its largest late-stage fund ever: $7 billion on April 17. Most of it is earmarked for fast-growing AI companies that need capital before an IPO. (Financial Times)
AI IPOs in 2026–2027 will be massive. Sequoia is writing $500M+ pre-IPO checks.
FluidStack at $18B valuation
FluidStack, which aggregates distributed GPU compute for AI labs, was in talks for a raise valuing the company at $18 billion. (Bloomberg)
On April 13, Booking.com confirmed a breach exposing customer names, emails, and partial card numbers. Attackers reportedly used an AI agent to automate exfiltration via an internal API. (Wired)
First documented case of an AI agent used to automate large-scale data exfiltration. The attack surface for companies exposing internal APIs to agents has fundamentally changed.
Forbes AI 50 2026
Forbes published its AI 50 2026 on April 16. The headline: "From AI Dominance to AI Independence." The list rewards companies building technological sovereignty (own infrastructure, open-source models, local training) over those dependent on lab APIs. (Forbes)
Investors and enterprise clients are pricing in independence from AI labs. A counter-trend worth watching.
OpenAI GPT-Rosalind
OpenAI announced GPT-Rosalind on April 16, a life sciences / biopharma-specialized model trained on clinical and pharmaceutical research corpora. Available via API only for enterprise clients in health and biotech. (STAT News)
The start of a vertical model strategy at OpenAI. Instead of improving generalist models, they are training sector-specific versions. Pattern to watch.
Meta hyperagents
Meta published a research paper on "hyperagents": AI agents capable of modifying their own weights to improve performance on non-coding tasks. (arXiv)
If this ships from the lab, it is a rupture. An agent that can self-improve without human supervision is outside the frame of current AI agents. Watch closely.
What the week of April 14, 2026 changes for independent developers
1. The enterprise AI coding market is diverging from consumer tools
Factory at $1.5 billion with Morgan Stanley and EY confirms that enterprise AI coding is no longer the same market as consumer tools. The two segments will diverge: different pricing, different features, different security requirements. For solo devs, the good news is that competition on the consumer segment (Claude Code vs Codex Mac) stays intact — labs need both markets.
2. Cross-session memory is becoming a standard, not a premium feature
Codex Mac shipped it. Anthropic will have to follow. It is the difference between an assistant that starts from zero every morning and one that remembers what you were working on yesterday. If you have Claude Code workflows that include a CLAUDE.md context file, expect to be able to drop it within six months.
3. Opus 4.7 justifies revisiting your prompts for more ambitious tasks
If Anthropic genuinely tripled performance on hard production tasks, the prompts you were using with Opus 4.6 are undersized. Now is the time to ask for more complex tasks in a single pass. Fewer iterations, fewer tokens, better coherence.
4. Your dev tools and intelligence agencies share the same infrastructure
The Barbaccia memo is not abstract. The same models you use to build your app are being deployed inside US intelligence agencies. If you build for regulated markets or you are based in Europe, the question of jurisdiction over the data you pass to these tools is worth asking.
5. Firebase agent-native at I/O 2026 could change the solo app stack
If Firebase ships a native agent SDK, integrating AI into a mobile app as a solo dev becomes significantly shorter. No need to run your own agent server, no need to stand up LangChain or LlamaIndex. Just Firebase and native agents. Worth watching closely at I/O on May 19.
Conclusion: the week AI coding became geopolitical
Factory at $1.5 billion says the enterprise market exists and is already here. OpenAI and Anthropic are fighting feature by feature over coding tools and benchmarks — and Anthropic is not stopping at code: Claude Design now targets visual prototyping too. And Mythos enters US federal agencies — the same model declared too dangerous for the public two weeks ago.
What I take away concretely: Opus 4.7 to test this week on my hard workflows. Claude Design to evaluate for Waku assets and iOS prototypes. Codex Mac to seriously evaluate on cross-session memory. And Mythos in the DoD reminds me that my dev tools are built by companies that also make geopolitical trade-offs. I am not diversifying tomorrow — but I am thinking about it.
Alex
Key takeaways
Factory raises $150M at $1.5B on April 16, 2026, led by Khosla + Sequoia + Insight + Blackstone. Clients: Morgan Stanley, EY, Palo Alto Networks. Enterprise AI coding is now a distinct market with its own compliance and security requirements.
OpenAI launches Codex Mac on April 16 with background computer use, integrated browser, cross-session memory, image generation via gpt-image-1.5, and 111 new plugins (GitLab, Atlassian, Microsoft Suite). The Verge calls it a direct shot at Claude Code. Cross-session memory is the feature that most distinguishes Codex Mac from Claude Code today.
Claude Opus 4.7 ships April 16: 64.3% on SWE-bench Pro vs 57.7% for GPT-5.4. 3x more difficult production tasks solved vs Opus 4.6. 3x better visual resolution. A direct response to the Opus 4.6 regression criticism.
Claude Design launches April 17 in research preview by Anthropic Labs: interactive prototypes, slides, designs, and one-pagers in conversation with Claude Opus 4.7. Export to Canva, PDF, PPTX, HTML. One-click Claude Code handoff. Target: founders, PMs, and marketers without a design background, and designers who want to explore faster.
Federal CIO Gregory Barbaccia signed an April 7 memo to connect DoD, Treasury, Commerce, Homeland Security, Justice, and State to Claude Mythos, revealed by Bloomberg and Reuters on April 16. The same model declared too dangerous for the public two weeks earlier is entering US intelligence agencies under the FedRAMP framework.
Google launches Gemini Mac on April 15 (Personal Intelligence on by default) and announces Google I/O 2026 on May 19 with Firebase agent-native, Gemma, Android 17, and vibe-coding tools. Firebase agent-native is the most concrete data point for solo devs using Firebase as a backend.
