Overview of Anthropic's Labs Lead On Fable's Capabilities + Building AI-Native Products — With Mike Krieger
This conversation with Mike Krieger, co-founder of Instagram and now head of Anthropic Labs, focuses on how Anthropic turns frontier model capabilities into products like Claude Code, why Labs exists, and how the company thinks about safety, product strategy, pricing, and the future of AI-native software. Krieger explains how more capable models change the way teams work, why “closing the gap” between what models can do and how people actually use them is the core mission of Labs, and why Anthropic is trying to build products that move the industry forward rather than simply copy what already exists.
What Anthropic Labs Does
The purpose of Labs
- Labs was created to help Anthropic stay ahead of the model curve and build products that showcase what new models can do.
- Krieger describes Labs as a place to:
- Prototype frontier products
- Explore future use cases
- Close the gap between model capability and real-world product use
Why Labs mattered earlier vs. now
- When Krieger first started Labs, Anthropic had:
- A small product team
- Models that were improving, but not yet strong enough to fully inspire new product forms
- Today, Anthropic has a much larger product surface area, so Labs is less about filling a product vacuum and more about:
- Anticipating what models will be good at in the next 6 months
- Designing products that unlock those capabilities in useful ways
Fable, Model Capability, and the White House Backlash
What happened with Fable
- Krieger says Fable was the best model he had personally used, but it was only available briefly before access had to be pulled back.
- He notes the public reaction was immediate and intense, but warns against judging a model solely by the first few days of reaction.
Why the backlash mattered
- The U.S. administration’s concern appeared to be tied to capability uplift and access patterns, not just a single demo.
- Krieger frames this as part of a broader AI trend:
- Model capability is advancing quickly
- Risks scale with usefulness and accessibility
- Safety reviews need to happen in real time
His view on “real concern” vs. “marketing”
- Krieger pushes back on the idea that Anthropic is simply doing hype marketing.
- He argues the company is genuinely trying to:
- Surface real risks early
- Work with partners on safety-relevant capabilities
- Be transparent about what the models can do
How More Capable Models Change Work
Delegation gets bigger
- With stronger models, Krieger says his workflow shifted from:
- Small tasks and edits
- To large, goal-oriented delegation
- Examples he gives include:
- Asking a model to plan and execute a large Python-to-TypeScript code conversion
- Having the model break down work into subtasks
- Running overnight tasks that previously would have taken much longer and required more manual review
What improves with frontier models
- According to Krieger, better models are not just faster:
- They are more accurate
- More reliable
- Better at “seeing around corners” in software engineering
- But they still have limitations:
- Weaknesses in vision/UI details
- Gaps in debugging
- Inconsistent common sense
Anthropic’s Product Strategy
Building what’s next, not what’s already crowded
- Krieger says Anthropic only wants to enter areas where it can help push the industry forward.
- If Anthropic is just doing the same thing as everyone else, he considers that a poor use of the team’s time.
- He also emphasizes that Anthropic should not build everything itself:
- The company wants to remain a platform
- But also a product company that creates new categories
Why companies worry about Anthropic building adjacent products
- The discussion references concerns from ecosystem players like Cursor and Figma:
- If Anthropic creates products on top of its own models, will it compete with partners?
- Krieger’s answer:
- Be transparent
- Use shared building blocks
- Build where Anthropic can move the field forward
- Avoid turning Anthropic into a company that merely clones existing products
Future Directions: Agency, Environment, and Self-Knowledge
A major theme: give Claude more environment awareness
Krieger says one of the biggest product opportunities is giving Claude:
- More agency
- More self-knowledge
- Better understanding of the environment it is operating in
This would reduce friction like:
- Downloading files manually
- Moving zip files between tools
- Failing to understand what context or permissions it already has
Why this matters
- Krieger believes this could transform Anthropic’s products “from head to toe.”
- The goal is to make Claude:
- More useful to non-technical users
- More autonomous in complex tasks
- Better at solving repeatable workflow problems
Closing the gap between intent and execution
- Another major theme is making software reflect what users already know in their heads.
- Example:
- A privacy worker had to manually move a ticket across multiple queues through many copy-paste steps.
- Labs helped automate that workflow, making the system finally match the user’s mental model.
- Krieger wants this kind of “what’s in my head is now what I’m using” experience to become much more common.
Pricing, Tokens, and Product Metrics
Token usage is not a perfect proxy for value
- Krieger says Anthropic has found little correlation between the people using the most tokens and the people producing the most value.
- He argues that token maxing is not the right way to evaluate quality or usefulness.
What Anthropic actually optimizes for
Anthropic looks at:
- Model intelligence
- Efficiency
- Task-specific token use
- Outcome quality
Pricing may shift toward outcomes
- Krieger says he has thought about outcome-based pricing as a more logical model for some AI products.
- That works more naturally in cases where success is concrete, like support ticket resolution.
- It’s harder for more subjective tasks like strategy critique or brainstorming.
- One practical step in that direction is Claude Managed Agents, which lets customers define outcomes and rubrics rather than just consume tokens.
Consumer AI, App Breakouts, and the Bigger Market
Why AI consumer breakouts are still rare
- Krieger thinks it’s harder than ever to build a breakout consumer app because:
- The market is more consolidated
- Users are locked into existing platforms
- Data gravity makes switching difficult
His take on productivity charts and app launches
- He suggests the industry may be generating many more app releases without a proportional boom in durable, high-usage products.
- This doesn’t mean AI is failing; it means making something people want is still very hard.
Instagram as a comparison
- Krieger says Instagram could likely have been built much faster with today’s tools, perhaps with far fewer people.
- But the challenge of making a product people love would still have remained.
Culture, Safety, and Anthropic’s Mission
Why he joined Anthropic
- Krieger says the company’s internal culture is shaped by a real belief in making AI go well for humanity.
- He contrasts that with product-first companies that spend most of their time on execution metrics.
- At Anthropic, mission and safety are central to the company identity.
Anthropic’s broader impact on Silicon Valley
He points to a few possible effects Anthropic may have on the Valley:
- Renewed interest in philanthropy
- More serious, real-time discussion of AI governance
- Stronger emphasis on considering societal impact before release
Safety and harmful products
- Krieger acknowledges that some AI products may be useful in isolation but harmful in practice.
- He says Labs actively evaluates whether a product might:
- Nudge people in the wrong direction
- Be “hypey” and bad for the world
- Create unacceptable risk if released broadly
Key Takeaways
- Anthropic Labs exists to turn model advances into new product forms, not just incremental features.
- Model capability is changing how work gets done, especially for engineering and product-building tasks.
- Anthropic wants to lead the market without becoming a copycat platform, even if that means navigating conflicts with partners.
- Token usage is an imperfect metric; outcome quality and efficiency matter more.
- The future of Claude may be more agentic, more environment-aware, and more deeply integrated into real workflows.
- Safety is not an afterthought at Anthropic—it is built into how the company thinks about product strategy.
Notable Insight
“Closing the gap” is the recurring theme:
the gap between model capability and product usage, and the gap between what users intend and what software makes possible.