Martin De Saulles, Principal Analyst, and Fiona, AI Research Analyst — Information Matters, May 2026
The coding-agent category has the cleanest growth story in agentic AI. The numbers are not in dispute. Cursor reached $2 billion in run-rate revenue by February 2026. Claude Code was disclosed at $2.5 billion run-rate alongside Anthropic’s February 2026 Series G announcement. GitHub Copilot ended Microsoft’s Q2 FY2026 with 4.7 million paid subscriptions, against the approximately 90 per cent Fortune 100 adoption figure Satya Nadella had disclosed on the preceding earnings call. The Cursor curve from zero to $2 billion ARR is the fastest such climb the analyst community has documented.
Our new report (#IM104) is not about those numbers. It is about the question they don’t answer.
The missing question
The headline case for the coding-agent category rests on revenue scale and growth velocity. Both are real. But the question that determines whether the equity is worth what it is currently priced at — gross margin — has not been answered by any of the major vendors. Until it is, the sector is being underwritten on an assumption that an application-layer agent making heavy frontier-model inference calls will arrive at SaaS-grade margins.
Our reading is that it will not. The first IPO filing or formal earnings break-out which discloses gross margin will be the trigger that prices that distinction in. The timing is closer than the consensus narrative implies.
The behavioural data points the wrong way for the SaaS comparator. Anthropic and OpenAI have each disclosed flagship-model price cuts in the 60-80 per cent range over the eighteen months ending Q1 2026. That price-curve compression helps application-layer vendors who can stabilise per-user inference cost faster than per-user revenue compresses; it hurts those whose users learn to consume more tokens per session as the agents get more capable. JetBrains’ April 2026 workplace AI tools survey found developers who had adopted agentic coding tools were running them on a meaningfully larger share of their working time than developers who had adopted only autocomplete-style tools twelve months earlier. The unit of consumption has changed. A token is not the right unit for a margin model when the user’s session length is doubling.
The closest indirect signals support the direction without supplying the number. Microsoft’s Q2 FY2026 earnings called out commercial cloud gross margin in the high 60s percentage and explicitly noted “AI infrastructure scaling” as a margin headwind without quantifying the per-product impact. Salesforce has flagged consumption-based agent economics as a structural input to gross margin in its FY26 Q1 commentary, without disclosing per-product COGS for Agentforce.
The trigger event
The cleanest test is Cursor. Anysphere is the only major coding-agent vendor whose business is plausibly close to a public-equity event during 2026, and Cursor’s product is monolithically a coding agent — there is no enterprise platform layer obscuring the unit economics, no bundled cloud business, no foundation-model training revenue subsidising the application surface. An S-1 filing from Anysphere would publish a coding-agent gross margin without ambiguity for the first time.
CEO Michael Truell’s March 2026 Fortune interview is the data point against this thesis: he told the magazine the company is focused on long-term mission rather than IPO. We have read that interview as a deflection of the timing question, not a denial of the eventual filing. The pattern of advisory engagement consistent with S-1 preparation — the breadth of legal counsel, the cadence of bank introductions — is what an outside observer would expect at Anysphere’s scale and revenue trajectory, although none of it is public-disclosed and this part of the prediction is qualitative.
Anthropic via Claude Code is the candidate the consensus narrative reaches for first. We think it is the wrong candidate. Anthropic is plausibly preparing for a public-equity event during late 2026 or 2027 — Wilson Sonsini representation was reported by the Financial Times in early December 2025 — but the disclosure surface in any Anthropic S-1 will not be a coding-agent margin in isolation. The Anthropic filing is consequential for the category — it will disclose the margin profile of the underlying inference infrastructure, which is the input to every coding-agent vendor’s COGS — but it will not be the coding-agent margin reset itself. The reset requires a pure-play disclosure.
We hold the view that the trigger event lands in the third quarter of 2026: an Anysphere (Cursor) S-1 filed during Q3, with the road-show during Q4, and the listing during Q4 2026 or the early weeks of Q1 2027. A one-quarter delay is the contingent reading if the macro window deteriorates.
The number, and the gap
The disclosed margin in that filing will be the data point that matters. We expect it to land in the 50-60 per cent range on a properly-loaded gross-margin calculation that includes the inference-cost component at non-discounted commercial rates. That is materially below the 75-85 per cent range that a SaaS comparator would price into the equity. A ~20-point gap, priced when the disclosure event lands.
The first read on the filing will be the margin number. The second, more consequential read will be the rate of margin compression or expansion implied by the cohort-level usage trajectory disclosed in the same document. Cursor’s per-user inference consumption has been growing — that has been our read of the developer-survey data and the company’s own pricing-tier evolution — and the S-1 will quantify it.
The category bifurcates
The contrarian read of this report is not that coding agents are going to fail. It is that the category is going to bifurcate, that the mechanism of bifurcation is the public disclosure of gross margin, and that the timing is closer than the equity-market positioning currently implies.
The specialists survive. Companies that have chosen a vertical, a workflow layer, or a deployment posture outside the autocomplete primitive — Cognition, Sourcegraph, Lovable, Tabnine in its self-hosted segment, the QA-and-debugging tier — operate independently into 2028. The revenue ceiling for survivors is lower than the headline cohort but the margin profile is structurally better.
The middle is repriced or absorbed. The autocomplete and generic-agent layer consolidates into Microsoft via GitHub, Google, the foundation-model providers’ first-party coding offerings, and JetBrains for the IDE-resident audience. Several names in the second cohort are acquired during 2027, on terms set against the post-disclosure comparable rather than the pre-disclosure one.
The headline names diverge. Cursor and Cognition do not converge. The independent IPO path requires a sustainable margin trajectory — either pricing power that compresses inference cost as a share of revenue, or the kind of model-tier specialisation that only a foundation-model owner can provide. We expect Cursor to remain independent through 2027 with margin pressure visible in subsequent filings, and Cognition to either find a strategic acquirer or continue private at higher capital-cost burn.
Discipline of the house
Our 50-60 per cent range is reasoned from the inference-economics evidence in our Q1-2026 Agentic AI Market Report and the public price-curve disclosures from Anthropic and OpenAI. It is not sourced from a primary disclosure by a coding-agent vendor, because no such disclosure exists at the date of this report. That is the point of the report. When the disclosure arrives, this estimate will be tested. We will publish the reconciliation in the next quarterly update.
The single highest-signal observation point in 2026 is the Q3 disclosure window. If a major independent coding-agent vendor files an S-1 between July and October 2026, the prediction in this report is testable on the day the filing is made public. The position holds in either case; what changes is when the market reprices, not whether.
Read the full report (10 pages, ~30 minutes). Coding agents: the margin reset is coming — Information Matters, Report #IM104, May 2026.
