Overview of the Big Technology Podcast episode with Boris Cherny
Alex Kantrowitz interviews Boris Cherny, head of Claude Code at Anthropic, about the product’s explosive growth, how it’s being used by developers and non-developers alike, what Anthropic is building next, and whether the current surge in AI-agent usage is sustainable. The conversation covers token “maxxing,” model efficiency, rate limits, competition with OpenAI, and the broader future of software, work, and AI agents.
Key Takeaways
-
Claude Code has become a breakout product for Anthropic
- Cherny says growth has been “insane” and repeatedly accelerating since launch.
- He describes the product as many users’ first real experience with Anthropic, not just a coding tool.
- Anthropic’s products and API are both growing quickly; he would only say the mix is changing, not which is bigger.
-
The product’s core value is agentic tool use
- Claude Code differs from chatbots because it can use tools: browsers, computers, files, and connected apps.
- Cherny frames it as a shift from autocomplete-style AI to systems that can actually do work.
- This enables tasks like building software, organizing files, and handling workflows across apps.
-
Non-engineers are increasingly adopting AI agents
- Early usage was dominated by engineers, but now marketers, accountants, operations teams, and solo operators are finding uses.
- Cherny shared examples of booking flights, managing calendars, generating presentations, and even doing bookkeeping-like tasks.
- He argues the most valuable uses often come from unexpected people inside organizations.
Topics Discussed
Growth and demand
- Anthropic’s demand has reportedly surged dramatically year over year.
- Claude Code’s growth inflected hard after model releases like Opus and Sonnet updates.
- Cherny says the team has never seen this level of growth in prior hypergrowth products.
Token maxxing and usage incentives
- Kantrowitz raised concerns about “token maxxing,” where companies encourage employees to burn tokens to hit AI adoption targets.
- Cherny said this is not a major share of demand.
- He sees most usage as genuine productivity-seeking behavior, not artificial inflation.
Productivity gains at work
- Cherny claims code written per engineer at Anthropic has increased by roughly 250% since Claude Code launched, without major drops in quality or reliability.
- He advises companies to:
- give employees access to tokens,
- create psychological safety for experimentation,
- and avoid optimizing too early.
- His broader point: real gains come from letting people discover new workflows.
Efficiency, loops, and model behavior
- Kantrowitz pressed on whether models waste tokens by getting stuck in loops or overthinking simple tasks.
- Cherny’s response:
- intelligence is the first priority,
- efficiency comes next,
- and users can already control effort levels.
- He pointed to model settings like:
- model choice: Opus, Sonnet, Haiku
- effort levels: low to maximum
- He rejected the idea that these inefficiencies are permanent limitations.
Auto Mode and safer automation
- Anthropic’s Auto Mode reduces constant permission prompts.
- Instead of asking the user every time, one Claude checks whether another Claude’s tool use is safe.
- Cherny said this is safer and less fatiguing than repeated approvals.
Rate limits and capacity
- Some users complain about hitting rate limits and then switching to alternatives.
- Cherny said:
- only a small percentage of users hit limits,
- Anthropic has doubled some rate limits,
- and it is expanding capacity to keep up with demand.
- Heavy users can also move to API-based usage for more scale.
Competition with OpenAI Codex
- Cherny said competing products are flattering and push the field forward.
- He emphasized that Anthropic’s focus is simply serving users well.
- He also noted that the growth of Claude Code continues to accelerate despite competition.
The future of chatbots and agents
- Cherny agreed that the future of chat is likely proactive:
- not just answer the question,
- but take the next action for the user.
- He sees agents as the future of the product experience.
Moats, software, and the “SaaSpocalypse”
- Cherny discussed which business moats matter more in an AI era:
- network effects likely get stronger,
- switching costs may weaken because AI can help migrate systems.
- He doesn’t think one universal app will replace all software, but he does think AI will reshape how software is used.
AI safety and self-improving models
- On the question of models improving themselves, Cherny sounded open to fast takeoff scenarios.
- He emphasized that Anthropic exists because of AI safety concerns.
- The company is thinking seriously about the long-term risk of recursive improvement.
World models vs. LLMs
- Kantrowitz raised the debate over whether reliable agents require a separate “world model.”
- Cherny did not take a strong theoretical stance, but pointed to surprising planning and reasoning abilities already observed in Claude.
- His view: the practical results matter, and current models are already doing more than many expected.
Notable Insights
-
“Give people tokens, but don’t token max.”
- Cherny’s advice: let employees experiment freely, but don’t turn token usage into a blunt competitive metric.
-
Productivity gains often come from unexpected places
- He repeatedly stressed that the best AI use cases are often discovered by people management wouldn’t have predicted.
-
AI adoption is forcing organizational redesign
- Cherny compared the current AI transition to the early computer era: companies only gain real value when they restructure workflows around the new technology.
-
The user experience is changing from passive to agentic
- The old model was: ask a question, get an answer.
- The emerging model is: ask for help, and the system carries out actions on your behalf.
Bottom Line
Boris Cherny portrays Claude Code as part of a larger shift from chat-based AI to agentic AI that can operate tools, automate workflows, and take real actions. He argues the product’s growth is real, not just gamified, and that most of the industry’s concerns about token waste or model inefficiency are solvable engineering and product problems—not fundamental limits. The big picture: Anthropic is betting that AI’s next frontier is not just better answers, but systems that can reliably do the work.