The AI coding market in 2026 — where the money is, and the wedge

01 / THE SPLIT

Two markets wearing one name.

"AI coding" is really two economies. One is proven and profitable-shaped; the other is enormously funded and quietly unproven.

$2B

Cursor ARR by Feb 2026, from $100M a year earlier. $29.3B valuation, 1M+ paying. Simple per-seat.

~$15M

CodeRabbit ARR (code review), 20% month-over-month, ~$24-48 per seat. Bounded, human-reviewed.

~15%

Independent task-completion rate for Devin (3 of 20, Answer.AI). Headline ARR is huge; reliability is not.

02 / WHAT MAKES MONEY

The proven money is bounded and supervised.

Proven

In-editor assistants + code review

Cursor $100M to $2B ARR in ~13 months, $2.3B Series D at $29.3B.
CodeRabbit ~$15M ARR, $60M Series B at $550M.
Graphite $52M Series B. Greptile ~$30M Series A at $180M.
Pattern: a bounded workflow, a human in the loop, billed per seat.

Funded, unproven

Autonomous agents

Cognition / Devin $1M to a reported $492M run-rate, $26B valuation.
Growth is enterprise compute spend, not proven reliability.
~15% task success; falls into unrecoverable error loops.
2026 consensus: you cannot leave it running unsupervised.

Funded is not the same as profitable. No gross-margin or profitability data survived fact-checking for any player. The whole category is burning cash; ARR figures are self-reported annualized run-rates.

03 / THE AUTONOMOUS REALITY

The ceiling nobody puts on the billboard.

Independent testing (Answer.AI, Jan 2025) had Devin finish 3 of 20 tasks: 14 failures, 3 inconclusive. Agents lack operational awareness (running Linux commands in PowerShell, declaring failure before a command finishes) and loop without recovering. You cannot submit a task Friday and trust the code Monday.

Fair caveat: that number is single-lab, n=20, ~17 months old in a field that moves weekly. Current agents are better. But the structural point holds: blind, long, unsupervised autonomy is the unreliable lane, and it is the one selling on compute.

04 / THE WEDGE

The market is moving away from full autonomy.

The smart money is funding the opposite of "more autonomous." YC backed Compyle ("a less autonomous coding agent that asks before it acts") and infrastructure for agents that ship reviewable pull requests (Amika, Dedalus). The underserved gap is reliability + oversight + one narrow workflow.

Upgrades

Dependency and framework version bumps. High-tedium, recurring, easy to verify.

Test backfill

Turn untested code into covered code. The suite is the oracle.

Migrations

Large mechanical refactors applied consistently across a whole repo.

The play: take the long-horizon, multi-agent engine and point it at ONE bounded, high-tedium, human-verifiable job, and always hand back a reviewable PR. The ~15% blind-autonomy ceiling stops mattering because a human reviews the output, exactly why code review and per-seat assistants make money and blind autonomy does not. The "runs for hours across a repo" capability becomes the engine, not the pitch.

05 / FOCUSED vs BROAD

For a small entrant, focused beats a do-everything assistant.

Every reliable money-maker sells one bounded job well. A broad "AI assistant for everything" is a worse business at small scale: harder to position, harder to trust, harder to price. The differentiator that is actually defensible here is self-hostable / privacy (your code never leaves your environment) plus depth on the long jobs, not "we can do more."

06 / MONETIZATION

Subscription for the value, credits for the cost.

The model you floated, credits via OpenRouter plus a subscription, is the right hybrid, and it is what the market leader actually does: Cursor's per-seat tiers now bundle usage credits with overage billing. Here is the honest breakdown.

Layer	What it is	Why
Subscription	A flat per-seat or per-workspace plan.	The proven model. Predictable revenue, predictable bill (buyers fear metered surprises). This is where durable margin lives.
Credits (via OpenRouter)	Metered model usage; heavy jobs draw down credits, overage tops up.	Passes variable model cost through with a markup and caps your downside, so a token-hungry long job can never burn your margin.

Margin truth: reselling model credits is thin. OpenRouter takes its cut, providers cut prices constantly, and users can route around you. So treat credits as cost-pass-through plus a light markup, not the profit center. The subscription carries the business; credits keep heavy users from eating it. Avoid pure usage-based pricing, it is the thing that scares dev-tool buyers off.

07 / RECOMMENDATION

The bet, in one screen.

Keep the long-horizon multi-agent engine. Change the packaging, the runtime, and the promise.

ProductA supervised long-horizon agent for one narrow workflow (upgrades, test backfill, or migrations) that hands back a reviewable PR.

Pitch"Give it the job, get a reviewable PR." Supervised, you keep ownership. Not "autonomous overnight magic."

PricingPer-seat subscription as the base, plus usage credits (via OpenRouter) for heavy jobs and overage.

EdgeSelf-hostable / private, plus real depth on the big jobs. Not "we do more."

DropThe full general assistant, and the "runs unsupervised for days" framing.

FirstValidate which workflow buyers will pay for before over-building. Reliability is the precondition.

Why it could fail: the specific-workflow willingness-to-pay is unproven, the self-hosted segment is unsized, incumbents can add the feature, and reliability is not there yet. None of this works until the engine can run unattended without breaking, which is the hardening work in flight.