CHONGQING, CHINA - FEBRUARY 1: In this photo illustration, a smartphone displays the Moltbook website homepage, describing the platform as a social network for artificial intelligence agents, with a Moltbook character graphic visible in the background, on February 1, 2026, in Chongqing, China. Moltbook is a newly emerged social network exclusively for artificial intelligence agents, where autonomous AIs can post, comment and interact with one another without human participation, drawing widespread attention and debate in global technology and ethics communities about the implications of AI-to-AI communication and autonomy. (Photo illustration by Cheng Xin/Getty Images)

Getty Images

Two groups of people are arguing about AI capabilities, and they are describing different products. One group tried a free chatbot and found it unreliable. The other is using frontier agentic coding tools and watching them complete days of engineering work in hours. Both are right. They are just not talking about the same thing.

Investors have already priced the difference. Anthropic raised $30 billion in February 2026 at a $380 billion valuation, driven largely by Claude Code, its terminal-based coding agent, which reached $2.5 billion in annualized revenue within nine months of its public launch. Enterprise customers account for more than half of that revenue. Anysphere (Cursor) and Lovable raised at 30x and 33x revenue multiples in the same period. That kind of valuation premium does not attach to a glorified autocomplete tool.

The consumer data tells a different story. An NBER working paper by OpenAI researchers and Harvard economist David Deming, drawing on 1.5 million conversations, found that coding accounts for just 4.2% of ChatGPT consumer messages. More than 70% of usage is non-work-related. The three most common topics are practical guidance, information lookup, and writing. For the majority of the 800 million weekly users on ChatGPT’s free tier, AI is a better search engine. That is a useful product but it isn’t a product enterprise engineering teams are rebuilding their workflows around.

Andrej Karpathy, former Tesla AI director and OpenAI research scientist, explained his thoughts yesterday. The models improving fastest are improving in coding and technical domains because those domains offer verifiable rewards: code passes its tests or it does not. That makes them tractable for reinforcement learning in a way that writing and general advice are not. The same dynamic concentrates enterprise investment, which funds further improvement. Consumer AI and enterprise coding AI are now on different capability curves, running at different speeds, even when they share a company name on the label.

GitHub Copilot, Claude Code, and Anysphere now hold over 70% of the AI coding market, which CB Insights values at $4 billion. Gartner projects that 40% of enterprise applications will include task-specific AI agents by the end of 2026, up from less than 5% today. Software stocks have shed roughly $2 trillion in market cap from peak as institutional investors reprice the sector against what agentic coding tools can actually do to engineering headcount. The investors are clearly not waiting for the perception gap to close.

The perception gap will not close for everyone at the same time. It will close fastest for companies and investors who stop treating AI as a single category and start treating it as two: a consumer utility optimized for reach, and an enterprise agent optimized for verifiable output. The winners in the second category are already separating. Three platforms now control over 70% of the AI coding market, switching costs are rising as engineering teams rebuild entire workflows around specific tools, and enterprise contracts are locking in multi-year commitments.

Late entrants will find the moat wider than benchmark charts suggest. For founders, the remaining opportunity is vertical: domain-specific coding agents in legal tech, biotech, and financial engineering, where verifiable reward signals are strong and incumbent tooling is weakest.

For LPs, the signal is simpler; firms whose portfolio companies have already integrated agentic coding into production are running at a structurally different cost base than those still in pilot mode. That gap compounds. Closing it requires actually using the product, not reading about it.