AI Coding Agent Pricing: The Seat-to-Token Shift in 2026

The AI coding agent market repriced itself in the space of a single quarter. GitHub Copilot dropped its flat per-seat subscription for usage-based billing on June 1, 2026. Devin was already metering by compute unit. Cursor is reportedly in talks at a $50B valuation while running $2B in annual run rate. The signal isn't which vendor is "winning." It's that the entire vertical just moved the cost of autonomy from the vendor's balance sheet to the buyer's.

What changed in AI coding agent pricing on June 1, 2026

For most of Copilot's life, the deal was simple: pay a fixed monthly fee per developer, use it as much as you want. On June 1, 2026, GitHub moved every plan to usage-based billing built on AI Credits: Pro includes $15 a month in credits, Pro+ includes $70, Max includes $200. Premium model calls draw down the pool. TechCrunch's headline on the reaction, quoting a developer who called the change "a joke," captures the mood better than any press release does.

The mechanics matter more than the backlash. A flat seat is a budgeting instrument: one number, no surprises, and the vendor absorbs the risk that some users barely touch the product while others hammer it daily. Usage-based pricing flips that. The buyer now carries the cost of every planning loop, every sub-agent delegation, every retry on a failed test. Devin priced this way from day one, calling its unit an ACU (Agent Compute Unit) rather than a seat. Copilot just joined it.

• GitHub Copilot: $10/mo entry (Pro), usage-based AI Credits since June 2026, compute risk on the buyer.

• Devin (Cognition): $500/mo per team, usage-based ACU billing since launch, compute risk on the buyer.

• Cursor (Anysphere): $20/mo entry (Pro), included usage plus overage, compute risk on the buyer but softened by a shared pool.

• Replit Agent: $25/mo entry (Core), included credits plus effort-based overage, compute risk on the buyer.

None of the four majors still sell a true unlimited flat seat for agentic work. That's the actual news. The vendor comparison chart is a distraction from the structural point: nobody is subsidizing your heaviest engineer's usage anymore.

The reason is structural, not competitive. Older AI assistants offered single-line autocomplete: small context in, small suggestion out, cheap to serve at scale. Modern coding agents read large swaths of a codebase, hold long context windows, plan across multiple steps, and call tools repeatedly, sometimes running for minutes per task. Every one of those steps consumes tokens, and the more autonomous the agent, the more tokens it burns per unit of delegated work. A vendor that keeps a flat seat price on top of that cost curve is either subsidizing heavy users out of its own margin or has not yet scaled to the point where the subsidy hurts. Gartner's own guidance to enterprise buyers now treats a consumption-cost model, run against your last 90 days of real usage, as a mandatory step before any agent contract renewal.

The valuation gap that revenue doesn't explain

Line up the money and the picture gets stranger before it gets clearer. Cursor's parent Anysphere closed a $2.3B Series D at a $29.3B valuation in November 2025 and was reportedly in talks near $50B by April 2026, backed by roughly $2B in annualized revenue. Cognition raised past $1B at a $26B valuation. Replit tripled its own valuation from $3B to $9B in a March 2026 round.

Absolute revenue explains part of the gap: Cursor's run rate is close to four times Cognition's. It does not explain all of it. Cognition's ARR moved from $73M annualized in mid-2025 to roughly $492M by mid-2026, a delta that outpaces Cursor's growth rate over the same window even though the base is smaller. Replit's ARR moved from $2.8M to over $150M before settling near $240M, then targeting $1B by year-end. Read those two numbers side by side and the story isn't "who is biggest." It's "whose delta compounds faster from here."

That distinction is the one that should actually inform how an operator underwrites this vertical, whether the lens is a competitive audit, a partnership decision, or a build-versus-buy call for an internal engineering org. Augment and Poolside, two other well-capitalized entrants in the same category, have not published revenue figures that hold up against any of the three leaders above, which is itself a signal: in a vertical this well-funded, silence on ARR usually means the number isn't ready to be compared yet, not that it doesn't exist.

Devin priced by compute before anyone else had to

Cognition built Devin's pricing around Agent Compute Units from the start, which looks less like a pricing quirk in hindsight and more like an early read on where the whole category was headed. The Team plan runs roughly $500 per seat per month, with no per-user cap on who can use the shared allotment, and extra ACUs billed on top when a job runs long.

What that price buys is narrow by design: PR review, code migrations, issue triage, scheduled maintenance, not open-ended research or slide generation. Cognition reports Devin can learn a specific codebase's conventions over repeated sessions rather than starting cold each time, and that enterprise usage grew roughly 50% month over month for six consecutive months before the company's ARR crossed $492M. The $500/seat floor puts it out of reach for a solo developer. It's built for an engineering org that can already point to a migration backlog and put a dollar figure on the hours it costs to clear it manually.

Replit's bet on consumption pricing at consumer scale

Replit took the opposite entry point on the same underlying shift. Replit Agent sits inside Replit Core at $25 a month, roughly 5% of Devin's list price, and scaffolds a full-stack app, wires a database, and deploys it from a natural-language prompt with no local setup.

The catch is the same compute economics everyone else is now navigating: a subscriber who runs Agent 3 at maximum autonomy for a long session can burn $5 to $15 of credits in a single workflow, because the agent spins up sandboxed environments and calls premium frontier models on every step. Replit's bet isn't that usage-based billing goes away. It's that a consumer-friendly entry price plus transparent credit consumption converts a much larger top of funnel than an enterprise-only sales motion does, and the ARR trajectory (over fifty million registered users, a stated $1B ARR target for year-end 2026) is the evidence that bet is paying off.

What Manus's acquisition signals about the broader agent economy

Not every agent that mattered to this category shipped as a standalone coding tool. Manus operated as a fully autonomous general-purpose agent, running inside its own virtual browser, terminal, and file system to plan and execute multi-step tasks end to end. Meta acquired the startup behind it in 2026, folding a credible autonomous-agent product into a much larger balance sheet rather than letting it keep raising and scaling independently.

That outcome is worth tracking alongside the pure-play coding agents because it's a second, distinct exit pattern in the same broader category: instead of a $9B or $26B independent valuation, an incumbent buys the capability outright. For anyone mapping which AI coding agent and general-agent startups are structurally durable versus acquisition targets, an M&A exit from a well-funded general agent is a data point that belongs in the same model as the funding rounds.

The open-source counter-argument: what Suna proves about the token math

Suna, built by Kortix, runs the same category of task (browser, shell, file system, multi-step execution) as an open-source, self-hostable project with roughly 20,000 GitHub stars, often cited as the closest open-source answer to Manus. The core software is free. The real cost shifts entirely to compute plus whichever model API you connect, which is the same usage-based math every proprietary vendor just adopted, minus the vendor's margin on top of it.

The trade is explicit rather than hidden behind a pricing page: you own uptime, security patching, and prompt-engineering effort yourself, in exchange for full inspection of what the agent is actually doing and no mandatory subscription. For a platform team already running its own model infrastructure, that math can beat $500 a seat. For most teams, the setup cost is exactly why the hosted vendors still have a business.

What the human-in-the-loop pushback tells procurement teams

Cognition's own CEO, Scott Wu, argued publicly that AI coding agents shouldn't replace human engineers, a notably cautious position from the person selling the agent. The same week, TechCrunch reported the opposite pressure from the other direction: some developers now refuse to work without AI tools at all, a dependency the outlet flagged as a risk in its own right.

Read next to the billing shift, those two data points aren't a philosophical debate. They're an operational one. Usage-based pricing means the cost of a team that leans hard on agentic execution is now variable and visible on the invoice, not fixed and buried in headcount. A CTO who treats "how much should we delegate to the agent" as a productivity question without treating it as a budget question is going to be surprised by the token line the same way GitHub Copilot's user base was surprised in June.

Should you sign a seat-based contract before your next renewal?

Skip it if the vendor still lets you. A flat seat is the last artifact of an economics regime the whole vertical just left. Get the vendor to model your last 90 days of actual usage under their consumption pricing before signing anything new, the way procurement guidance from Gartner and others is now recommending across this category.

Worth locking in if you're the vendor's smallest, lightest user: heavy discounts on legacy seat contracts do still exist as a retention tactic, and a light-usage team can extract real value from being the exception a sales rep doesn't want to lose. For everyone else, the delta to track isn't the sticker price on the pricing page. It's the gap between what your team actually consumes and what the contract assumes you will. Signals, not narratives, is the only way to underwrite that gap correctly.