Fifty-two days of 74 launches! Anthropic has created the craziest team in the AI era!
In March 2026, a map was uploaded on X: someone counted Anthropic from February 1st to March 24th, making a copy of the product release book. Count it:52 days, 74 releasesI don't know. My first reaction was to count it wrong, again, correctly. Instead of 74 minor bug repairs, two completely new models, a desktop-level Agent platform, a mobile phone remote control system, an enterprise application market, a code security scanner, more than 20 Claude Code functionality updates, and a dozen third-party connectors. Every single one, it's enough for a lot of companies to send a PR article.Anthropic. 52 days.I don't know. Behind this is a ship that's raging.
74 press releases. What was it?
These 74 releases cover four main directions and three layers of things grow simultaneously:
First tier: model layer continued upwards
- Opus 4.6.: February 5, Anthropic flagship model
- Sonnet 4.6: 12 days to follow
- Mythos: Leaked at the end of March, code name Capybara, said to be step change level capacity enhancement
Level 2: Product layer independence AI portal
- Claude Code.:AgentTeams (multi-AI collaboration), Voice Mode (voice interaction), Fast Mode (multiplier speed), Auto Mode (auto-review), Dispatch (telephone remote control)
- WorkDesktop-level Agent platform, used at the end of March
- Claude Code Security: scan code holes with reasoning capability and find 22 undiscovered bugs in Firefox
Level three: full ecological layer access
- Marketplace: A third party opened a shop on Claude's platform. Snowflake, GitLab, Refrit are in.
- Claude for Open Source6 months for open source developers. Max, $1,200.
- Third-party connectorA dozen connectors: Google Drive, Slack and others
The third floor is not sequential.It's also pushing.I don't know. The model is stronger and the product is immediately in use; the product is paved with ecological access; data from ecological feedback is fed back to model training.
A machine that accelerates itself.
74 releases, 52 days.Average 1.4 per dayI don't know. This speed is abnormal. Even the most rolling Chinese Internet company, this density is difficult to achieve.
Anthropic is a team of about 2,000 to 3,000 people, and it's a team of about 2,000 to 3,000 people.Less than half of OpenAI(OpenAI currently has about 4,500 people and is scheduled to expand to 8,000 by the end of the year), even less than the 5,000 people in Google DeepMind.Few people, but output density is far higher than peersI don't know.
Why? Wheel effect
Project Compass discloses a data:Anthropic Engineers are now 60% done by Claude.I don't know. That figure was 28 per cent a year ago.Productivity has increased by about 50%.I don't know.
More than that, Work This product. Cowork is Anthropic's main desktop-level Agent platform this year, which helps you process documents, mail, data analysis. This heavyweight product,From zero to release, it took Claude Code 10 days.I don't know.
10 daysI don't know. This is the wheel:
- The stronger Claude Code, the faster the engineer will develop the new function.
- The new feature makes Claude Code stronger and faster.
More than 2,000 people and Claude are working together. And Claude's ability is up every few months, so the machine is accelerating.Fifty-two days, 74 releases, possibly 100 by the next quarter.I don't know.
Build what you Use
But it's not enough for AI to write fast. Many companies are using AI to support their development. Why is it so slow?
The deeper reasons are:Anthropic is probably the only one that's actually working on it in Build.I don't know.
Claude Code is not just a tool for outside users.It's first and foremost Anthropic's own internal infrastructure.I don't know. Agent Teams, Tasks, Git Worktree, all of which are internal, run-through and release. Their daily development is the best test environment in itself, and bugs are exposed faster than any QA team.
Produc Compass also mentioned one data:Anthropic, 60 to 100 internal releases per day.I don't know. We see 52 days of 74 external releases, just the tip of the iceberg.
Ironic contrast.
Google has a Principal Engineer named Jaana Dogan in charge of Gemini API. She put a push on X to the effect that:We discussed the one-year distributed Agent programming system in Google, where there was a lack of convergence and slow progress. She described the problem to Claude Code, which produced a working version in an hour.I don't know.
She's not using her own Gemini, but her rival's product.
Last year, before GPT-5, Anthropic discovered that OpenAI's technical team had been using Claude Code to revoke OpenAI's API access. Anthropic, the spokesman, said something so bad:"Claude Code has become a preferred tool for programmers, so it's not surprising that OpenAI engineers are using it."
Microsoft requires core teams like Windows, 365, Teams to install Claude Code and GitHub Copilot to submit comparative feedback in parallel. Amazon is even more exaggerating.1,500 engineers petitioned in internal forums for Claude CodeAgainst the forced use of their own Kiro.
You don't even use your own engineers. How do you know where to improve?
The exact person is in Ship.
There is also a missing element:It's not an Anthropic official account. It's a specific person.I don't know.
- CEO Dario Amodei 5 times more than 3 times more than 3 times more than 3 times more than 3 times more than 3 times more than 3 times more than 3 times more than 3 times more than 3 times more than 3 times more than 3 times more than 3 times more than 3 times more than 3 times more than 3 times more than 3 times more than 3 times more than 3 times more than 3 times more than 3 times more than 3 times more than 3 times more than 3 times more than 3 times more than 3 times more than 3 times more than 3 times 3 times more than 3 times 3 times more than 3 times more than 3 times 3 times 3 times more than 3 times more than 3 times 3 times more than 3 times 3 times 3 times 3 times more than 3 times 3 times 3 times more than 3 times more than 3 times 3 times 3 times more than 3 times more than 3 times more than 3 times more than 3 times
- Product Manager Noah Zweben. It's a personal understanding of the product.
- Technical VP sent a demo video from Cowork to comment area to answer technical questions in person
This is completely different from the way most technology companies publish it. It is usually a public relations team with a single calibre, a single schedule for the market sector, and a standard "excised to announce" issued by CEO.
Anthropic is more like:Who does it, who does it, what does it say?I don't know. When everyone's name is tied directly to the product, driving you is not just KPI or OKR anymore. You're building your own work and building your own influence.
They're looking at us one step further.
March 26th. Mythos. The new model accidentally leaked. Internal code name Capybara. Fortune reported it was a... Step change in capabilities Level model, much stronger than the current Opus 4.6.
What does that mean?The 74 releases we saw were made from Opus 4.6.I don't know. And the Anthropic internal rate is already developing the next product using Mythos or near Mythos capabilities.
If 60 per cent of the work is done by Claude, every increase in model capabilities is directly equivalent to an increase in productivity across the company.Opus 4.6 has given them 74 features for 52 days. With a stronger model, the machine will turn faster.I don't know.
The worst part of the ship is here:It's not linear acceleration. It's exponential acceleration.I don't know. Every cycle, speed is rising.
It's not a product, it's an operating system.
Put 74 updates together. Anthropic is making one.AI operating system:
| Component | Anthropic Achieved |
|---|---|
| Terminal | Claude Code. |
| Desktop Environment | Work |
| Process management | Agent Teams |
| File System | Google Drive, Slack Connector |
| External driver | Computer Use |
| Applied stores | Marketplace |
| Synchronization across devices | Dispatch. |
| Time job | /Schedule,/loop |
Every part of the traditional operating system, Anthropic, is re-enacted with AII don't know. And fast enough for you to react.
♪ Crazy data ♪
| Indicators | Data |
|---|---|
| Launch speed | 74 out of 52 days, 1.4 per day |
| Group size | 2000-3000, less than half of OpenAI |
| AI Completion | 60% of engineer work is done by Claude. |
| Internal Publishing | 60-100 per day |
| Productivity enhancement | About 50 percent a year. |
| Valuation | $380 billion |
| Annualized income | $19 billion |
| YC W26 Percentage | 52%, first time exceeding OpenAI |
Summary
It's only March.
But I don't think it's these numbers that deserve the most attention, but the logic behind them:
- Make your own AI, and the product itself makes AI stronger and stronger, and then more.
- Those who do these things believe in what they're doing, and everyone's name is written on it.
It's not just "an AI company that's rolling." It's a brand new way of building.
OpenAI is adding, increasing product lines and scenes. Google's doing integration, stuffing AI into existing product matrices.Anthropic is multiplying: a core engine multiplied by an infinite delivery boundary Noodles.I don't know. So 74 releases look a lot, but the bottom is the same thing in the long.
Domestic AI companies are also working hard, publishing sessions one by one, and product lines are expanding. But if you do something that doesn't even work for your team, if your release is just a schedule for the market sector, not an engineer's work, and if you speed up by adding people instead of AI, that gap will only grow.
It's not about rollers. It's about rollers.I don't know.
Fifty-two days of 74 releases, only beginning. Once the wheel turns, it will only get faster!