Anthropic
An AI company that builds the Claude models and products such as Claude chat and Cowork.
Anthropic grew revenue roughly 10x a year, from about $100M in 2023 to $1B in 2024 and $10B in 2025, passing $19B ARR by February, by focusing narrowly on coding and B2B and pairing each frontier model with a product. Lesson: when you are the underdog, pick a narrow focus, build for the current model, and retest parked bets on each new model.
- About $100M revenueNobody yet linked Claude with coding
- $1B revenueOpus 3 trained for long-form code
- About $10B revenueOpus 4.5 drives Claude Code inflection
- Over $19B ARRSmallest, least-funded lab overtakes SaaS giants
From Amol Avasare interview
What was said about Anthropic
1How a small lab grew by betting on coding
As the underdog, Anthropic picked a narrow focus on coding and B2B. It was the smallest, least-funded lab, with no big-tech cash flow and no head start, so it chose coding and B2B. Revenue went from about $100M in 2023 to $1B in 2024 and about $10B in 2025, and passed $19B ARR by end of February.
The coding focus goes back to a 2021 founder doc, and coding was seen as a loop that speeds up research because better coding models help build better models. Amol Avasare calls the approach "freedom through constraints". Atlassian, Palantir and Snowflake sit at roughly $4.5B to $6B ARR after 15 to 20 years.
A small Opus 3 training change for long-form code won developers. In 2023 nobody linked Anthropic, Claude and coding. Dianne Penn saw people using models for long-form code rather than autocomplete and pushed to train Opus 3 for it, a relatively small training change that became an early differentiator.
Opus 3 launched in early March 2024 and answered the question the team kept hearing: why should somebody choose Claude? It showed that a company of under 200 people could build a frontier model.
Frontier models need a frontier product for users to feel them. Opus 4.5 took off because Claude Code gave it a vehicle, and Claude Code spread faster because Opus 4.5 could run work end to end like an agent. Opus 4.5 landed during winter break and drove the Claude Code inflection.
Tony Fadell contrasts OpenAI, which started as a viral technology demo and added product teams late, with Anthropic, which he says now has higher valuation and revenue, as a case for designing the whole customer journey from the start.
One mission above every product line settles priority fights fast. When priorities compete, teams ask which serves Anthropic's mission and everyone backs the result, even at a cost to their own goals. Cat Wu says she would be happy if Claude Code failed but Anthropic succeeded.
An example is prioritizing first-party products and the API over third-party tools like OpenClaw using subscriptions when compute was scarce. Boris Cherny left for Cursor and returned after two weeks because he missed the safety mission.
2How Claude Code started and spread
- Internal use only4 or 5 months, daily users vertical
- External releaseFelt alien, took months to catch on
- Opus 4 growth jumpNew model lifts adoption
- Second jumpReached 4% of public GitHub commits
From Boris Cherny interview
Claude Code began as an experiment giving the model a bash tool. With only a bash tool, the model answered "what music am I listening to" without being told how, showing it could work out tool use on its own. Boris Cherny spent a month on prototypes and a month on post-training.
His internal announcement got two likes. The terminal was simply the easiest thing for one person to build, and it stayed because no other form could keep pace with the model.
Its local, question-asking form fit what models could reliably do. Claude Code launched local, asked questions and did not pretend you could delegate everything to it. The original Codex web, which had you delegate a task and wait, came too early for the models of the time.
Andrew Ambrosino of OpenAI: "we were too agi pilled for the moment".
Internal use came first, and outside adoption took months. Claude Code was used internally for 4 or 5 months and internal daily users went vertical; the external release came in February but felt alien and took months to catch on. Growth jumped with Opus 4 in May and again in November, reaching 4% of public GitHub commits.
Daily active users doubled in the month before the interview. SemiAnalysis predicts a fifth of GitHub commits by year end.
Slight under-resourcing pushes engineers to automate with Claude. Some projects get a single engineer, who ships fast and hands the rest to Claude. Anthropic's engineering team grew about 4x while pull requests per engineer rose 200%.
At Meta, hundreds of engineers working for a year gave a few percentage points. The team's rule of thumb is that having Claude do something beats doing it yourself.
Cowork came from people already using Claude Code for non-coding work. Users such as a data scientist doing SQL analysis in the terminal showed demand beyond code. Cowork was built in about 10 days with Claude Code and shipped rough as a research preview.
Cherny cites Facebook Marketplace, which came from 40% of group posts being buying and selling, as the same pattern: build what users already try to do.
3Labs: many overlapping bets with clear owners
Labs runs small pods chasing 10x, 100x or 1000x bets. Pods are kept small, sometimes a single engineer, because large teams on vague ideas slow down. Labs incubated Claude Code, Skills, MCP and Claude Design.
The team holds a strong view on the theme and a loose view on the prototype, hires people who want zero-to-one work, and turns bets off, though owners act like founders and that is hard.
Each bet has one lead who decides to double down or stop. Many people can explore an area, but a named bet lead decides whether to continue and where to add or remove people. Mike Krieger says the role matters more as the team moves faster.
Quote: "we're all lost together, but somebody's got the pen".
Only teams with real conviction get exploratory bets. Krieger says he has never succeeded in assigning a team an area it isn't excited about. Teams with exciting directions explore even when two or three overlap.
Ami Vora tells overlapping teams "we're all fundamentally on the same team".
A foundations team shares memory, MCP and file storage across products. Chat and Cowork had separate memory systems, MCP and file storage, so any third experiment would feel disconnected from the start. A separate foundations team now builds these so users get their memory wherever they are.
Keeping foundations apart from frontier product teams lets overlapping experiments feel complementary.
Overlap early, then fold winners into one coherent product. Vora would rather have five overlapping products than one over-constrained one that never meets the market, and once something works the team reshapes it for people who want stability. She admits the folding-in method is not solved.
Instagram's main feed drew about 80% of use, so extra tabs must earn their place. Max Schoening of Notion cites Claude desktop's three Cowork tabs as feature sprawl and argues feature count is as silly a metric as lines of code.
4Parking failed bets and building for the next model
Park what models can't do yet and retest it on each new one. The 2024 computer use product took about 20 minutes per task and taught the user nothing. The team parked it in an eval harness, and Claude 3.7 showed a leap, succeeding more often than not.
Shipping a failure as an eval or handing it to research counts as a win in labs. Building early is what lets you recognize when it starts working; people who judge AI from one old try become "capability blind".
Ask what changes for users when the next model arrives. Penn's team asks "let's say Claude 8 comes around", what users would do differently, and whether today's work still fits. Labs prototypes that fail are revisited one to two model generations later.
Capability jumps are discontinuous and hard to time, so evals and red teaming are the way to spot them; the gap between what a model can do and what users know is a "product overhang". Co-work and Skills came from plans that became possible early.
The hard part is getting maximum capability from today's model. Cat Wu says building for a near-omniscient model is easy; the skill is guiding users onto the golden path and patching current weaknesses. The to-do list was added because early Claude stopped after 5 of 20 call sites; later models use it unprompted.
Boris Cherny pushes the other side: build for the model 6 months out, give it tools and a goal rather than strict workflows, since scaffolding adds maybe 10 to 20% and the next model often wipes that out. Wu sets direction about a month out and adjusts.
5Evals as the spec, PRDs only when needed
Read failed transcripts to name the exact failure, then write an eval. In the Claude 2 days, about 80% of "bad at following instructions" complaints were wrong JSON, and a 30 to 40 example eval set fixed it. Evals now replace the PRD for well-defined problems.
Penn's line: "You have to sweat the tokens as much as you sweat the pixels."
About 10 great evals are enough to define a goal. Cat Wu hands off features like memory with five evals, how to run them, which pass and the prompt that raised the success rate. Hundreds aren't needed.
PRDs stay for aligning big launches and exploring ambiguous bets. Every model release still gets a PRD to align product, engineering, legal and safety, and ambiguous work like computer use uses them too. About 60 to 80% of what ships has no PRD.
Claude Code uses a one-pager (goals, delightful use cases, failure modes). Growth handles small changes in Slack, runs a 30-minute cross-functional kickoff with legal and safeguards for big projects, and increasingly prototypes instead of writing docs.
Weekly metrics readouts and written principles let people decide alone. The Claude Code team reviews goals and their drivers together every week and keeps principles on key users and trade-offs. People decide without waiting on a PM.
Goals are sharp enough to rule options out, for example: professional developers at enterprises safely reach zero permission prompts.
Five trusted users give fast, accurate feedback on a new model. A handful of people articulate what makes a model good far better than most; ask them first. Internally, Wu polls teammates' vibes at lunch, then pulls data on patterns like too-abrupt replies or weak memories.
When the model does something odd, she asks it to explain why, which can expose a confusing system prompt or a sub-agent whose work it trusted unchecked, and then fixes the harness.
6How the team ships in days
Low process and research previews cut timelines from months to days. Many features went from 6 months to 1 month, sometimes 1 week or 1 day. Most Claude Code features ship as research previews, and an evergreen launch room turns announcements around the next day.
Wu says the newest model (Mythos) helps only a little; the pace predates it. The aim is that anyone can take an idea to users in under a week. Speed claims come with no data on defects or rework.
Give engineers unlimited tokens first and optimize cost later. Some engineers spend hundreds of thousands of dollars a month on tokens. Once an idea works and scales, the team moves it to cheaper models like Haiku or Sonnet.
Penn treats token spend as an input and experimentation as the real goal.
Fixing feedback within minutes makes users give more of it. Early on, Cherny fixed internal feedback items within minutes so users felt heard. Now Claude reads the channel, proposes fixes and opens PRs for him to review.
Fiona Fung runs a routine that reads feedback channels overnight and has PRs ready by morning.
A team that covers for each other sustains a heavy launch cadence. Anthropic shipped four model series in 2024 and more than that in Q2 of this year. Penn's team stays up before launches to help the owner and covers PTO.
She hires low-ego, team-oriented people over empire builders.
7Who builds and who convenes
A PM still catches the glue work Claude would drop. Krieger thought Claude could cover a nearly shipped project; a PM then caught customer success enablement and safeguards that would have been missed. He says Claude is not yet a convenor.
Faster shipping means the PM must be more operationally excellent, so builders stay heads down while someone ensures it connects.
Engineers act as PM on projects of two weeks or less. On the growth team, the engineer owns small projects, including talking to security and legal, while the PM advises; bigger or controversial ones stay with the PM. Five engineers, one designer and one PM now act like 15 to 20 engineers.
Growth hires product-minded engineers on purpose. Nearly all Claude Code PMs have shipped code, and the team prefers engineers with taste to more PMs.
Product taste is what sorts tens of thousands of requests. The Claude Code team gets tens of thousands of GitHub issues; choosing what to build and how takes taste. If something is easy, build it in an hour instead of debating it.
Wu expects engineering background to help only for the next few months as skills shift.
New managers start as ICs, shipping code before managing people. Fung has new managers ship and learn the product first, and managers keep making PRs part time. Teams she joined said they appreciated her using what they built.
Hire for first-principles thinking, not pattern matching. Penn's team has used the same hiring loop for three years, looking for people who work out the value for this user and technology rather than copying consumer or B2B SaaS playbooks.
8The growth team and automating growth
Growth tilts toward big bets because product value compounds. A normal growth team spends 60 to 70% of time on small and medium bets; Anthropic's roughly 40-person team flips that to about 70/30 or 50/50 toward large swings. Avasare expects product value to rise 100 to 1,000x in two years versus 30 to 50% for a grocery app.
Small optimizations still run because no one else will do them. The growth team built the Chrome extension, which now underpins Cowork and Claude Code use cases. He says this fits AI-first products only.
Claude automates the growth loop, at about a junior PM's win rate. The CASH effort splits shipping into finding opportunities, building, testing against quality and brand bars, and analyzing, each scored and hill-climbed weekly. It handles copy and small UI tweaks with a human approving.
It wasn't possible before Opus 4.5 and became promising with Opus 4.6. Brand guidelines can be encoded as a skill to reduce review. Win-rate figures come from Avasare alone.
Activation should match users to features, not chase one model's tricks. Models improve faster than onboarding tests can ship, so learnings go stale. A quiz-style flow asking who the user is, used by Anthropic, MasterClass and Calm, tested well.
Importing memory from ChatGPT eased the cold start. Avasare favors friction that teaches over removing all friction.
Growth leaves money on the table to protect safety and brand. The team forgoes metric gains for safety, brand and quality. Tests you'd never ship aren't run; merely uncomfortable ones run only if returns are high.
9Claude building and checking Anthropic's own software
Route every action through primitives both humans and agents use. Krieger's principle is that everything a human can do, an agent should be able to do. He says few products get this right, and those that do see agents connect steps never linked before.
An established product can move gradually from a side panel to a malleable personal landing page on the same primitives.
Claude builds and maintains much of Anthropic's internal software. For a project with at least four workstreams, Claude monitors progress and built the status UI the team uses, which the team changes with Claude when it doesn't fit.
This raises questions of who may update data and how to track provenance. Krieger still wants design systems so results stay predictable.
Verification replaces writing as the bottleneck, so Claude reviews first. Claude reviews 100% of pull requests, followed by a human layer; only throwaway prototypes skip human review. Engineers ship about 8x more code per quarter than in 2025.
Specs and quality scales checked into the repo let Claude Code Review confirm code still matches them. The 8x figure measures volume, not quality.
Fast AI-written code still needs human architecture and process care. Tony Fadell says engineers who read Claude's leaked source found the main loop brittle, though 90 to 100% of the code was reportedly AI-written. The leak itself came from a release-process change that passed two human reviews.
Most added safeguards have shipped, and the person remains at Anthropic. Fadell's advice: prototype with agents, lock architecture with expert review, then point agents at limited pieces.
10Safeguards and compute trade-offs
Fallback systems keep users served when safeguards block a model. Before the Fable models there was no strong fallback UX, so Anthropic built systems that route blocked users to good Opus 4 responses. Safety is checked on interpretability, lab evals and real-world release.
Third-party tools like OpenClaw lost subscription access because demand was high and the harness wasn't built for their usage; users got credits during the transition.
Why Claude can’t be your PM (yet) | Anthropic CPO Panel
Both guests work there. It runs overlapping bets under bet leads, has a foundations team sharing memory, MCP and file storage across products, and Claude now builds and maintains much of its internal software. Its 2024 computer use product was parked and retested on each new model.
Each bet has one lead who decides to double down or stop. Many people can explore an area, but a named bet lead decides whether to continue and where to add or remove people. Mike Krieger says the role matters more as the team moves faster.
Quote: "we're all lost together, but somebody's got the pen".
Only teams with real conviction get exploratory bets. Krieger says he has never succeeded in assigning a team an area it isn't excited about. Teams with exciting directions explore even when two or three overlap.
Ami Vora tells overlapping teams "we're all fundamentally on the same team".
A foundations team shares memory, MCP and file storage across products. Chat and Cowork had separate memory systems, MCP and file storage, so any third experiment would feel disconnected from the start. A separate foundations team now builds these so users get their memory wherever they are.
Keeping foundations apart from frontier product teams lets overlapping experiments feel complementary.
Overlap early, then fold winners into one coherent product. Vora would rather have five overlapping products than one over-constrained one that never meets the market, and once something works the team reshapes it for people who want stability. She admits the folding-in method is not solved.
Instagram's main feed drew about 80% of use, so extra tabs must earn their place. Max Schoening of Notion cites Claude desktop's three Cowork tabs as feature sprawl and argues feature count is as silly a metric as lines of code.
Park what models can't do yet and retest it on each new one. The 2024 computer use product took about 20 minutes per task and taught the user nothing. The team parked it in an eval harness, and Claude 3.7 showed a leap, succeeding more often than not.
Shipping a failure as an eval or handing it to research counts as a win in labs. Building early is what lets you recognize when it starts working; people who judge AI from one old try become "capability blind".
A PM still catches the glue work Claude would drop. Krieger thought Claude could cover a nearly shipped project; a PM then caught customer success enablement and safeguards that would have been missed. He says Claude is not yet a convenor.
Faster shipping means the PM must be more operationally excellent, so builders stay heads down while someone ensures it connects.
Route every action through primitives both humans and agents use. Krieger's principle is that everything a human can do, an agent should be able to do. He says few products get this right, and those that do see agents connect steps never linked before.
An established product can move gradually from a side panel to a malleable personal landing page on the same primitives.
Claude builds and maintains much of Anthropic's internal software. For a project with at least four workstreams, Claude monitors progress and built the status UI the team uses, which the team changes with Claude when it doesn't fit.
This raises questions of who may update data and how to track provenance. Krieger still wants design systems so results stay predictable.
Why AI is going vertical (again) | Dianne Penn (Anthropic)
Grew from a long shot in 2023 to roughly $50 billion in annual revenue by pairing each frontier model with a product and betting early on coding. It shipped four model series in 2024 and more than that in Q2 of this year alone, and keeps a bottoms-up, startup-paced culture. Keeps pods small, sometimes a single engineer, and looks for bets worth 10x, 100x or 1000x. It incubated Claude Code, Skills, MCP and Claude Design, and revisits failed bets a generation or two later. Proved a company of under 200 people could build a frontier model and gave users a reason to choose Claude. A small training change for long-form code drew the first developer fans. Took off because Claude Code gave it a vehicle, and landed during winter break to drive the Claude Code inflection. Claude Code in turn spread faster because of it.
As the underdog, Anthropic picked a narrow focus on coding and B2B. It was the smallest, least-funded lab, with no big-tech cash flow and no head start, so it chose coding and B2B. Revenue went from about $100M in 2023 to $1B in 2024 and about $10B in 2025, and passed $19B ARR by end of February.
The coding focus goes back to a 2021 founder doc, and coding was seen as a loop that speeds up research because better coding models help build better models. Amol Avasare calls the approach "freedom through constraints". Atlassian, Palantir and Snowflake sit at roughly $4.5B to $6B ARR after 15 to 20 years.
A small Opus 3 training change for long-form code won developers. In 2023 nobody linked Anthropic, Claude and coding. Dianne Penn saw people using models for long-form code rather than autocomplete and pushed to train Opus 3 for it, a relatively small training change that became an early differentiator.
Opus 3 launched in early March 2024 and answered the question the team kept hearing: why should somebody choose Claude? It showed that a company of under 200 people could build a frontier model.
Frontier models need a frontier product for users to feel them. Opus 4.5 took off because Claude Code gave it a vehicle, and Claude Code spread faster because Opus 4.5 could run work end to end like an agent. Opus 4.5 landed during winter break and drove the Claude Code inflection.
Tony Fadell contrasts OpenAI, which started as a viral technology demo and added product teams late, with Anthropic, which he says now has higher valuation and revenue, as a case for designing the whole customer journey from the start.
Labs runs small pods chasing 10x, 100x or 1000x bets. Pods are kept small, sometimes a single engineer, because large teams on vague ideas slow down. Labs incubated Claude Code, Skills, MCP and Claude Design.
The team holds a strong view on the theme and a loose view on the prototype, hires people who want zero-to-one work, and turns bets off, though owners act like founders and that is hard.
Ask what changes for users when the next model arrives. Penn's team asks "let's say Claude 8 comes around", what users would do differently, and whether today's work still fits. Labs prototypes that fail are revisited one to two model generations later.
Capability jumps are discontinuous and hard to time, so evals and red teaming are the way to spot them; the gap between what a model can do and what users know is a "product overhang". Co-work and Skills came from plans that became possible early.
Read failed transcripts to name the exact failure, then write an eval. In the Claude 2 days, about 80% of "bad at following instructions" complaints were wrong JSON, and a 30 to 40 example eval set fixed it. Evals now replace the PRD for well-defined problems.
Penn's line: "You have to sweat the tokens as much as you sweat the pixels."
PRDs stay for aligning big launches and exploring ambiguous bets. Every model release still gets a PRD to align product, engineering, legal and safety, and ambiguous work like computer use uses them too. About 60 to 80% of what ships has no PRD.
Claude Code uses a one-pager (goals, delightful use cases, failure modes). Growth handles small changes in Slack, runs a 30-minute cross-functional kickoff with legal and safeguards for big projects, and increasingly prototypes instead of writing docs.
Give engineers unlimited tokens first and optimize cost later. Some engineers spend hundreds of thousands of dollars a month on tokens. Once an idea works and scales, the team moves it to cheaper models like Haiku or Sonnet.
Penn treats token spend as an input and experimentation as the real goal.
A team that covers for each other sustains a heavy launch cadence. Anthropic shipped four model series in 2024 and more than that in Q2 of this year. Penn's team stays up before launches to help the owner and covers PTO.
She hires low-ego, team-oriented people over empire builders.
Hire for first-principles thinking, not pattern matching. Penn's team has used the same hiring loop for three years, looking for people who work out the value for this user and technology rather than copying consumer or B2B SaaS playbooks.
Fallback systems keep users served when safeguards block a model. Before the Fable models there was no strong fallback UX, so Anthropic built systems that route blocked users to good Opus 4 responses. Safety is checked on interpretability, lab evals and real-world release.
Third-party tools like OpenClaw lost subscription access because demand was high and the harness wasn't built for their usage; users got credits during the transition.
What happens after coding is solved? | Fiona Fung (Claude Code & Cowork)
Its managers all start as ICs to build product feel and rapport before supporting people. Its own staff are the first customers for coding tools, so feedback comes fast, though the evidence here comes from a team with free tokens.
Its local, question-asking form fit what models could reliably do. Claude Code launched local, asked questions and did not pretend you could delegate everything to it. The original Codex web, which had you delegate a task and wait, came too early for the models of the time.
Andrew Ambrosino of OpenAI: "we were too agi pilled for the moment".
How Anthropic’s product team moves faster than anyone else | Cat Wu (Head of Product, Claude Code)
About 30 to 40 PMs sit in five teams: models, platform, products, enterprise and growth. A mission above every product line and sharp focus explain how it moved fast and came from behind.
Fixing feedback within minutes makes users give more of it. Early on, Cherny fixed internal feedback items within minutes so users felt heard. Now Claude reads the channel, proposes fixes and opens PRs for him to review.
Fiona Fung runs a routine that reads feedback channels overnight and has PRs ready by morning.
New managers start as ICs, shipping code before managing people. Fung has new managers ship and learn the product first, and managers keep making PRs part time. Teams she joined said they appreciated her using what they built.
Verification replaces writing as the bottleneck, so Claude reviews first. Claude reviews 100% of pull requests, followed by a human layer; only throwaway prototypes skip human review. Engineers ship about 8x more code per quarter than in 2025.
Specs and quality scales checked into the repo let Claude Code Review confirm code still matches them. The 8x figure measures volume, not quality.
Head of Growth (Anthropic): Anthropic is automating its own growth
Revenue grew roughly 10x a year, from about $100M in 2023 to $1B in 2024 and about $10B in 2025, and it passed $19B ARR by end of February. The growth team (about 40 people) puts most effort into big bets, focuses narrowly on coding and B2B, and uses Claude to automate growth work.
Frontier models need a frontier product for users to feel them. Opus 4.5 took off because Claude Code gave it a vehicle, and Claude Code spread faster because Opus 4.5 could run work end to end like an agent. Opus 4.5 landed during winter break and drove the Claude Code inflection.
Tony Fadell contrasts OpenAI, which started as a viral technology demo and added product teams late, with Anthropic, which he says now has higher valuation and revenue, as a case for designing the whole customer journey from the start.
Fast AI-written code still needs human architecture and process care. Tony Fadell says engineers who read Claude's leaked source found the main loop brittle, though 90 to 100% of the code was reportedly AI-written. The leak itself came from a release-process change that passed two human reviews.
Most added safeguards have shipped, and the person remains at Anthropic. Fadell's advice: prototype with agents, lock architecture with expert review, then point agents at limited pieces.
Head of Claude Code: What happens after coding is solved | Boris Cherny
Raised a round at over $350 billion, and its engineering team grew about 4x while pull requests per engineer rose 200%. It checks safety on three layers (interpretability, evals, real-world release) and Claude reviews 100% of its pull requests.
Overlap early, then fold winners into one coherent product. Vora would rather have five overlapping products than one over-constrained one that never meets the market, and once something works the team reshapes it for people who want stability. She admits the folding-in method is not solved.
Instagram's main feed drew about 80% of use, so extra tabs must earn their place. Max Schoening of Notion cites Claude desktop's three Cowork tabs as feature sprawl and argues feature count is as silly a metric as lines of code.
Anthropic's CPO on what comes next | Mike Krieger (co-founder of Instagram)
Over 95% of Claude Code is written by Claude Code itself, and its merge queue had to be re-architected for the surge in pull requests. Product people work inside post-training, and it leans into its developer and builder brand instead of chasing ChatGPT's consumer reach.
One mission above every product line settles priority fights fast. When priorities compete, teams ask which serves Anthropic's mission and everyone backs the result, even at a cost to their own goals. Cat Wu says she would be happy if Claude Code failed but Anthropic succeeded.
An example is prioritizing first-party products and the API over third-party tools like OpenClaw using subscriptions when compute was scarce. Boris Cherny left for Cursor and returned after two weeks because he missed the safety mission.
The hard part is getting maximum capability from today's model. Cat Wu says building for a near-omniscient model is easy; the skill is guiding users onto the golden path and patching current weaknesses. The to-do list was added because early Claude stopped after 5 of 20 call sites; later models use it unprompted.
Boris Cherny pushes the other side: build for the model 6 months out, give it tools and a goal rather than strict workflows, since scaffolding adds maybe 10 to 20% and the next model often wipes that out. Wu sets direction about a month out and adjusts.
About 10 great evals are enough to define a goal. Cat Wu hands off features like memory with five evals, how to run them, which pass and the prompt that raised the success rate. Hundreds aren't needed.
PRDs stay for aligning big launches and exploring ambiguous bets. Every model release still gets a PRD to align product, engineering, legal and safety, and ambiguous work like computer use uses them too. About 60 to 80% of what ships has no PRD.
Claude Code uses a one-pager (goals, delightful use cases, failure modes). Growth handles small changes in Slack, runs a 30-minute cross-functional kickoff with legal and safeguards for big projects, and increasingly prototypes instead of writing docs.
Weekly metrics readouts and written principles let people decide alone. The Claude Code team reviews goals and their drivers together every week and keeps principles on key users and trade-offs. People decide without waiting on a PM.
Goals are sharp enough to rule options out, for example: professional developers at enterprises safely reach zero permission prompts.
Five trusted users give fast, accurate feedback on a new model. A handful of people articulate what makes a model good far better than most; ask them first. Internally, Wu polls teammates' vibes at lunch, then pulls data on patterns like too-abrupt replies or weak memories.
When the model does something odd, she asks it to explain why, which can expose a confusing system prompt or a sub-agent whose work it trusted unchecked, and then fixes the harness.
Low process and research previews cut timelines from months to days. Many features went from 6 months to 1 month, sometimes 1 week or 1 day. Most Claude Code features ship as research previews, and an evergreen launch room turns announcements around the next day.
Wu says the newest model (Mythos) helps only a little; the pace predates it. The aim is that anyone can take an idea to users in under a week. Speed claims come with no data on defects or rework.
Engineers act as PM on projects of two weeks or less. On the growth team, the engineer owns small projects, including talking to security and legal, while the PM advises; bigger or controversial ones stay with the PM. Five engineers, one designer and one PM now act like 15 to 20 engineers.
Growth hires product-minded engineers on purpose. Nearly all Claude Code PMs have shipped code, and the team prefers engineers with taste to more PMs.
Product taste is what sorts tens of thousands of requests. The Claude Code team gets tens of thousands of GitHub issues; choosing what to build and how takes taste. If something is easy, build it in an hour instead of debating it.
Wu expects engineering background to help only for the next few months as skills shift.
Fast AI-written code still needs human architecture and process care. Tony Fadell says engineers who read Claude's leaked source found the main loop brittle, though 90 to 100% of the code was reportedly AI-written. The leak itself came from a release-process change that passed two human reviews.
Most added safeguards have shipped, and the person remains at Anthropic. Fadell's advice: prototype with agents, lock architecture with expert review, then point agents at limited pieces.
Fallback systems keep users served when safeguards block a model. Before the Fable models there was no strong fallback UX, so Anthropic built systems that route blocked users to good Opus 4 responses. Safety is checked on interpretability, lab evals and real-world release.
Third-party tools like OpenClaw lost subscription access because demand was high and the harness wasn't built for their usage; users got credits during the transition.
A conversation with OpenAI's CPO Kevin Weil, Anthropic's CPO Mike Krieger, and Sarah Guo
Its PMs on API and product surfaces ended up doing the same job as research PMs, and it tests eval-writing in interviews. It released computer use this week and now teaches Claude about its own features, with links to exact documentation.
As the underdog, Anthropic picked a narrow focus on coding and B2B. It was the smallest, least-funded lab, with no big-tech cash flow and no head start, so it chose coding and B2B. Revenue went from about $100M in 2023 to $1B in 2024 and about $10B in 2025, and passed $19B ARR by end of February.
The coding focus goes back to a 2021 founder doc, and coding was seen as a loop that speeds up research because better coding models help build better models. Amol Avasare calls the approach "freedom through constraints". Atlassian, Palantir and Snowflake sit at roughly $4.5B to $6B ARR after 15 to 20 years.
PRDs stay for aligning big launches and exploring ambiguous bets. Every model release still gets a PRD to align product, engineering, legal and safety, and ambiguous work like computer use uses them too. About 60 to 80% of what ships has no PRD.
Claude Code uses a one-pager (goals, delightful use cases, failure modes). Growth handles small changes in Slack, runs a 30-minute cross-functional kickoff with legal and safeguards for big projects, and increasingly prototypes instead of writing docs.
Engineers act as PM on projects of two weeks or less. On the growth team, the engineer owns small projects, including talking to security and legal, while the PM advises; bigger or controversial ones stay with the PM. Five engineers, one designer and one PM now act like 15 to 20 engineers.
Growth hires product-minded engineers on purpose. Nearly all Claude Code PMs have shipped code, and the team prefers engineers with taste to more PMs.
Growth tilts toward big bets because product value compounds. A normal growth team spends 60 to 70% of time on small and medium bets; Anthropic's roughly 40-person team flips that to about 70/30 or 50/50 toward large swings. Avasare expects product value to rise 100 to 1,000x in two years versus 30 to 50% for a grocery app.
Small optimizations still run because no one else will do them. The growth team built the Chrome extension, which now underpins Cowork and Claude Code use cases. He says this fits AI-first products only.
Claude automates the growth loop, at about a junior PM's win rate. The CASH effort splits shipping into finding opportunities, building, testing against quality and brand bars, and analyzing, each scored and hill-climbed weekly. It handles copy and small UI tweaks with a human approving.
It wasn't possible before Opus 4.5 and became promising with Opus 4.6. Brand guidelines can be encoded as a skill to reduce review. Win-rate figures come from Avasare alone.
Activation should match users to features, not chase one model's tricks. Models improve faster than onboarding tests can ship, so learnings go stale. A quiz-style flow asking who the user is, used by Anthropic, MasterClass and Calm, tested well.
Importing memory from ChatGPT eased the cold start. Avasare favors friction that teaches over removing all friction.
Growth leaves money on the table to protect safety and brand. The team forgoes metric gains for safety, brand and quality. Tests you'd never ship aren't run; merely uncomfortable ones run only if returns are high.
The $44 Billion Company Automating Your Company's Finances
Made its first dollar three years ago and is rumored to be above a $50 billion run rate. Cited as an example of token spend becoming a major new spending category.
One mission above every product line settles priority fights fast. When priorities compete, teams ask which serves Anthropic's mission and everyone backs the result, even at a cost to their own goals. Cat Wu says she would be happy if Claude Code failed but Anthropic succeeded.
An example is prioritizing first-party products and the API over third-party tools like OpenClaw using subscriptions when compute was scarce. Boris Cherny left for Cursor and returned after two weeks because he missed the safety mission.
Claude Code began as an experiment giving the model a bash tool. With only a bash tool, the model answered "what music am I listening to" without being told how, showing it could work out tool use on its own. Boris Cherny spent a month on prototypes and a month on post-training.
His internal announcement got two likes. The terminal was simply the easiest thing for one person to build, and it stayed because no other form could keep pace with the model.
Internal use came first, and outside adoption took months. Claude Code was used internally for 4 or 5 months and internal daily users went vertical; the external release came in February but felt alien and took months to catch on. Growth jumped with Opus 4 in May and again in November, reaching 4% of public GitHub commits.
Daily active users doubled in the month before the interview. SemiAnalysis predicts a fifth of GitHub commits by year end.
Slight under-resourcing pushes engineers to automate with Claude. Some projects get a single engineer, who ships fast and hands the rest to Claude. Anthropic's engineering team grew about 4x while pull requests per engineer rose 200%.
At Meta, hundreds of engineers working for a year gave a few percentage points. The team's rule of thumb is that having Claude do something beats doing it yourself.
Cowork came from people already using Claude Code for non-coding work. Users such as a data scientist doing SQL analysis in the terminal showed demand beyond code. Cowork was built in about 10 days with Claude Code and shipped rough as a research preview.
Cherny cites Facebook Marketplace, which came from 40% of group posts being buying and selling, as the same pattern: build what users already try to do.
The hard part is getting maximum capability from today's model. Cat Wu says building for a near-omniscient model is easy; the skill is guiding users onto the golden path and patching current weaknesses. The to-do list was added because early Claude stopped after 5 of 20 call sites; later models use it unprompted.
Boris Cherny pushes the other side: build for the model 6 months out, give it tools and a goal rather than strict workflows, since scaffolding adds maybe 10 to 20% and the next model often wipes that out. Wu sets direction about a month out and adjusts.
Give engineers unlimited tokens first and optimize cost later. Some engineers spend hundreds of thousands of dollars a month on tokens. Once an idea works and scales, the team moves it to cheaper models like Haiku or Sonnet.
Penn treats token spend as an input and experimentation as the real goal.
Fixing feedback within minutes makes users give more of it. Early on, Cherny fixed internal feedback items within minutes so users felt heard. Now Claude reads the channel, proposes fixes and opens PRs for him to review.
Fiona Fung runs a routine that reads feedback channels overnight and has PRs ready by morning.
Verification replaces writing as the bottleneck, so Claude reviews first. Claude reviews 100% of pull requests, followed by a human layer; only throwaway prototypes skip human review. Engineers ship about 8x more code per quarter than in 2025.
Specs and quality scales checked into the repo let Claude Code Review confirm code still matches them. The 8x figure measures volume, not quality.
Fallback systems keep users served when safeguards block a model. Before the Fable models there was no strong fallback UX, so Anthropic built systems that route blocked users to good Opus 4 responses. Safety is checked on interpretability, lab evals and real-world release.
Third-party tools like OpenClaw lost subscription access because demand was high and the harness wasn't built for their usage; users got credits during the transition.
Why OpenAI is merging Codex and ChatGPT and the future of knowledge work | Andrew Ambrosino
It launched fully local, asked questions and did not pretend to be autonomous, which fit the models of the time better than the original Codex web.
Tony Fadell: How to build real taste (and why AI makes it matter more)
Tony says it now has a higher valuation and revenue than OpenAI. He says engineers who read Claude's leaked source called its main loop brittle, though 90 to 100% of the code was reportedly AI-written.
AI era skills: Why cultivating agency matters more than job titles | Max Schoening (Notion)
It started after OpenAI with less funding and now dominates, and Max credits Dario for succeeding twice. Claude's desktop app has three Co-work tabs, an example of feature sprawl.
The rise of the professional vibe coder (a new AI-era job)
Lazar says he hasn't become good at it because he starts without enough clarity, and that it can fix any problem it is aware of. He cites a tweet by someone at Claude Code about the pace of job change.
How a Meta PM ships products without ever writing code | Zevi Arnovitz
He fell in love with it after moving from Bolt to Cursor. It acts as the communicative, opinionated dev lead who plans, reviews its own work and must fix or rebut other models' findings. Claude.md loads standing instructions into every conversation.
$1M to $10M: The enterprise sales playbook with Jen Abel
Enterprises Jen talks to mention Gemini and Copilot more than Anthropic, since bundling means nothing new to adopt. Its people love technology and adopt new tools themselves.
Al Engineering 101 with Chip Huyen (Nvidia, Stanford, Netflix)
Its AI-driven feedback approach is cited as one example of getting reward signals from AI instead of humans during reinforcement learning.
Google Part III: The AI Company. Google is amazingly well-positioned... will they win in AI?
It is cited as one of the model makers in the AI race, alongside OpenAI. Rumors of its chat interface may have affected ChatGPT's launch timing.
Why AI evals are the hottest new skill for product builders | Hamel Husain & Shreya Shankar
Its chief product officer calls evals the most important new skill for product builders. Its staff are highly trained in data analysis, which helps close the dogfooding loop.
Solo founder, $80M exit, 6 months: The Base44 bootstrapped startup success story | Maor Shlomo
Base44 uses Claude 4 to write the first version of each app and the UI, as part of a routing pipeline across models.
AI prompt engineering in 2025: What works and what doesn’t | Sander Schulhoff
Its blackmail case is cited as misalignment where a model acted badly with no malicious prompting from a human.
Building a magical AI code editor used by over 1m developers in 4 months: Inside Windsurf
Windsurf uses Sonnet for planning but says sending a whole large code base to Anthropic would be impossible, too slow and too costly. Its in-house edit model applies changes better than Anthropic's model.
OpenAI’s CPO on how AI changes must-have skills, moats, coding, startup playbooks, more | Kevin Weil
Its coding models are very good, and Weil says it is one reason OpenAI's model lead is now small.
Inside Bolt: From near-death to one of the fastest-growing products in history | Eric Simons
Claude 3.5 Sonnet was the first model good enough to write production-grade code, which revived the shelved Bolt project. Anthropic ran out of GPUs for Bolt at the peak of the surge.
Behind the product: Replit | Amjad Masad (co-founder and CEO)
Claude Sonnet is the best coding model in Replit's agent. When Anthropic released computer use, Replit dropped its plans and built on it right away.
What AI means for your product strategy | Paul Adams (CPO of Intercom)
Intercom adapted Anthropic's models, alongside OpenAI's, for customer support, and may someday build its own model.
More in AI assistants
AI assistantsOpenAI
An AI research and product company behind ChatGPT, Codex and the Dots agent.
51 sources · 151 tips
AI assistantsChatGPT
OpenAI's AI assistant app used for chat, plugins and agent work.
51 sources · 102 tips
AI assistantsClaude
Anthropic's AI assistant and model family, used for chat, coding and agentic work.
19 sources · 35 tips
AI assistantsElevenLabs
A voice AI company with text-to-speech, dubbing, speech-to-text and music models, plus consumer and creator apps.
6 sources · 37 tips
AI assistantsGemini
Google's AI assistant app, available as a chatbot and moving toward a personal agent.
11 sources · 13 tips
AI assistantsNotebookLM
A Google Labs app that answers questions about your own sources and turns them into AI-hosted audio shows.
2 sources · 18 tips
AI assistantsTolan
A voice-first AI companion app with an animated alien character that remembers your conversations.
2 sources · 17 tipsGenie
A ChatGPT-powered chat app that writes marketing copy, taglines, tweets and app store descriptions from short prompts.
3 sources · 13 tips