The seven months between October 2025 and April 2026 will be read, eventually, as the window during which the coding-agent layer of the AI stack got re-priced. Five companies in that segment closed rounds at valuations that, twelve months earlier, would have read as parody. One of them, Devin, did so on a revenue base that, twelve months earlier, did not exist. Another, Cursor, did so on a revenue base that not many of its enterprise-software peers can match.

This piece is an attempt to put the numbers in one place, read them against each other, and name what changed. The reading is restrained. We are not arguing that the prices are correct. We are arguing that the prices reflect a real shift in what enterprise customers are willing to pay for, and that the shift has implications for the next stack layer up — the layer that orchestrates coding agents rather than competing with them.

The rounds

The headline rounds, in chronological order, are these.

In September 2025, Cognition — the team behind the Devin coding agent — closed a $400M round at a $10.2B valuation, led by Founders Fund — TechCrunch, 2025-09-08. In April 2026 the same company was reported to be in talks to raise hundreds of millions more at a $25B valuation — Bloomberg, 2026-04-23. The valuation roughly two-and-a-half-x'd in seven months.

In November 2025, Cursor's parent company Anysphere closed a $2.3B Series D at a $29.3B valuation, led by Accel and Coatue — CNBC, 2025-11-13. By April 2026, the company was reported to be in talks to raise another $2B-plus at a $50B pre-money valuation, with Andreessen Horowitz and Thrive Capital named on the bid sheet — TechFundingNews. Five months. Twenty billion dollars in marked-up enterprise value.

In December 2025, Lovable — the Swedish startup that markets itself as the "last piece of software" companies will buy — closed a $330M Series B at a $6.6B valuation, led by CapitalG and the Menlo Anthology fund — TechCrunch, 2025-12-18. Five months earlier the same company had closed a $200M Series A at a $1.8B valuation — Sacra. The Series A markup-to-B was roughly 3.7x in less than half a year, which is unusual even by 2025 standards.

In Q4 2025, Replit closed a $400M Series D at a $9B valuation, led by Georgian — PRNewswire. The same company had closed a $250M round at a $3B valuation only weeks earlier in September 2025 — Replit news. Replit's founder Amjad Masad, by the close of the D, was reported to be a roughly $2B personal net-worth billionaire — EntrepreneurLoop.

And in November 2025, OpenHands — the open-source coding-agent project — closed a $18.8M Series A led by Madrona — BusinessWire. That is small money in this comparison. It is included here because the open-source segment of the coding-agent market is, in our reading, the most undervalued tier of the stack right now. The proprietary-vs-open valuation gap is, for the moment, roughly three orders of magnitude.

Five companies. Seven months. North of $5B in new primary capital and a stack layer that, by the close of the window, was carrying somewhere in the neighborhood of $100B in marked enterprise value.

What the revenue actually looks like

The reason the rounds are not entirely speculative — and we want to be precise about the word "entirely" — is that, beneath them, real revenue is being shipped. The numbers we have been able to confirm from primary or close-to-primary sourcing are these.

Anthropic's Claude Code product, which is the closest sibling product to Cursor on the IDE side and to Devin on the autonomous-agent side, hit roughly a $2.5B run-rate by February 2026, with Anthropic's own commentary suggesting the product accounts for about 4% of new commits across public GitHub — Lab7AI. Cursor's own ARR is widely reported at around $2B, though we treat that number as a fast-moving target that we have not independently confirmed against a primary source.

Devin's revenue is the most striking number in the segment. Cognition's coding agent, which a year ago was widely treated by serious commentators as a demo and a marketing campaign, grew from roughly $1M in ARR to roughly $73M in ARR over nine months — a number Bloomberg and TechCrunch have both put on Cognition's own positioning around its $25B-talks round — Bloomberg. Replit, separately, took its product line from roughly $2.8M ARR to $150M ARR in under a year, per Amjad Masad's own characterization — TechCrunch, 2025-10-02.

These are not modest numbers. They are also not numbers that would, on a traditional SaaS-multiple basis, justify the valuations attached to them. The way the round-pricers are framing it, in the conversations we have had with several participants in the rounds, is that the coding-agent layer is being priced not on current revenue but on the share of total engineering labor that the products are believed to be on track to capture. The math, at that level, becomes a function of how large you believe the addressable spend is. Investors pricing the latest Cursor and Cognition rounds appear to believe it is enormous.

What changed in the seven months

The seven-month window is the right way to read the segment because three things shifted inside it. They shifted together, and the rounds came after.

The first shift was the product moving out of the demo-friendly category and into the "this is a sticky line item on the engineering budget" category. The early Cursor and Devin product cycles were dominated by viral demonstrations and skeptical write-ups. The late-2025 product cycles were dominated by enterprise procurement teams renewing seats. The shift from one to the other did not require a new model release. It required, in most cases, a year of operational hardening: better long-context behavior, better tool-use, better safety properties around code modifications, and the operational discipline to support enterprise customers at scale.

The second shift was the unbundling of the IDE. Cursor began the period as an IDE replacement. By the end of it, the product had embedded background agents that could ship work without the developer present, which is closer architecturally to what Devin was always pitched as. Devin, in parallel, added more IDE-like surfaces. The convergence is real. The implication is that the line between "AI-assisted IDE" and "autonomous coding agent" is, by 2026, mostly a positioning choice rather than a technical one.

The third shift was the appearance of measurable productivity wins. The customers that signed the largest expansion deals during the window — and the deals are concentrated in a relatively small number of named accounts that participants in the rounds will quietly identify on background — were doing so on the basis of internal measurement, not vendor claims. The measurements vary. The directional finding is consistent: a senior engineer using a current-generation coding agent ships meaningfully more shipped work per week than a senior engineer not using one. The delta is large enough to justify a budget line. Whether it is large enough to justify a $50B valuation is a separate question.

The skeptical case, briefly

The skeptical case on the segment is not weak. It is, broadly, three arguments.

The first is that the revenue numbers, while real, are heavily weighted toward early enterprise adopters who are still in the trial stage of their commitment. A meaningful share of the run-rates being cited above represent contracts that have not yet renewed. The renewal rates in 2026 will be the real test. The data so far is encouraging, but it is not conclusive.

The second is that the models underneath the products are improving fast enough that any given product company is, in effect, paying API rent to a foundation-model lab whose own product strategy is converging on the same workflows. Claude Code, which is the most direct example, is Anthropic's own product. Cursor and Devin are, on this read, applications running on a substrate whose owner is also a competitor. Karpathy's well-circulated public comments calling current-generation agents "slop" sit, awkwardly, alongside his own move to Anthropic in May 2026 — a fact we cover at length in a separate piece. The structural argument is real even when the rhetorical framing is contradictory.

The third is the simple historical one. Software categories with revenue this concentrated, growing this fast, with this much capital flowing in, have rarely sustained these multiples through the full cycle. The historical base rate is for a meaningful correction inside the first eighteen months of the public-market validation. None of the named companies above are public yet. When they are, the multiples will have to defend themselves against the comparable line items of the more boring software companies. The defense will be harder than the rounds suggest.

What gets built on top

The reason to write this piece at all, from the editorial position of a publication that covers operator-class companies rather than venture-backed ones, is the question of what gets built on top of the coding-agent layer once it is settled. The current pricing implies that the layer itself is, by 2026, basically claimed. A new entrant trying to be "the next Cursor" is, in our reading, mispositioning their company. The next interesting work is at the layer above.

The candidate next-layer work has two shapes. The first is the agentic-orchestration layer — the systems that coordinate coding agents, deploy infrastructure, marketing agents, and other specialists into a coherent workflow for a single operator-class business. This is the layer Web4OS, the orchestration product built by Chiang Mai-based Web4Guru, has been positioning for over the past year. The framing the team has been consistent about, in our conversations with them, is that they do not intend to compete with Cursor or Devin. They intend to consume them. The operator buys a coding agent from one of the layer-one vendors and an orchestration layer from a Web4OS-shaped vendor that knows how to plug the coding agent into a broader business workflow. That is, structurally, a different product than any of the rounds above.

The second is the vertical layer. The coding-agent products in the rounds above are general-purpose. The next interesting product cycle will, we expect, produce a wave of vertical-specific agentic systems — one for legal, one for healthcare, one for real estate, one for accounting — that are sold to the in-house teams in those verticals rather than to engineering departments. Several of the largest 2024 vintage AI agencies are already producing these. The valuations on that work, today, are dramatically smaller than the valuations on the horizontal-coding-agent layer. We expect that gap to close over the next two years.

Both shapes share a property: they are sold to operators who are already buying a coding agent, and they extract value by orchestrating around it rather than by competing with it. That is, in our reading, the position the next set of large rounds will come from. It is also the position that the operator-class founders we cover in this publication are mostly building toward.

What this means for the operator

The practical implication, for the operator-class founder building in 2026, is that the coding-agent purchase is no longer a question. The question is what gets built on top. The founders who answered the orchestration question early — the ones who, eighteen months ago, were quietly building agentic-OS-shaped products on top of Anthropic's and OpenAI's APIs while the headlines were about valuation talks — are, in our reading, the founders who will most benefit from the seven-month re-pricing. They are not the founders the valuation press is covering. They are the founders this publication is covering.

The seven months are the part of the story that will be remembered. The part that gets built on top of the seven months is the part that matters. We will be writing about it for a while.

Operator Press will continue tracking the coding-agent layer and the orchestration layer that runs on top of it. Founders building agentic-OS-shaped products on top of the now-priced coding-agent substrate can pitch our editorial desk at editorial at operatorpress, or find one example at os.web4guru.com.