Code review vendors think they’re ready for AI. Models are getting cheaper. PRs are small. Maybe 100 lines, maybe 10 a day. That’s the planning assumption.
The reality is already past it. Leading AI-assisted developers push 500-line PRs, 30 a day.
Agent-native workflows deliver 2,000-line PRs from a single terminal. Orchestrated swarms deliver 20,000-line packages: feature code, scaffolding, and tests, ready to merge.
Industrialized swarms run continuously. Over a million lines from a single seat. Every day.
The vendors planned for a 2x increase. They’re facing 100x. 1,000+ PRs per day from a single seat is already happening.
Tooling and policies will catch up for early adopters in about six months. Broad market penetration is two-plus years out.
The race to unlimited
Code review tools didn’t stumble into unlimited pricing. They were pushed.
Early tools had rate limits. Per user, per time period, capped by tier. Same patterns CI providers and cloud infrastructure still use.
Then a few vendors raised venture capital and competed on “unlimited reviews.” The message was simple: why pay for a tool with limits when this one has none?
Everyone else matched. Not because unlimited made financial sense, but because the alternative was losing deals. Rate limits disappeared from pricing pages across the category.
At human speed, unlimited was affordable. One vendor’s state-of-industry report measures the median PR at 76 lines. A few of those per week cost pennies in tokens.
That median is already outdated. The gap between what vendors budgeted for and what’s arriving is the whole problem.
The trap
Every vendor in the category knows rate limits work. They used to have them.
The problem isn’t engineering. It’s game theory.
The first vendor to reimpose limits loses deals to the ones still promising unlimited. The VC-funded competitors can burn cash longer. So everyone stays unlimited and hopes volume stays manageable.
That hope had a shelf life. Agent-driven development just expired it.
The rate limits that existed before weren’t calibrated for this either. 50 PRs per month was generous for human speed. It’s a single morning at agent-native speed.
Bringing limits back means rebuilding them for a volume profile that didn’t exist when they were first designed.
Open source is where it shows first
Vendors subsidize open source for three reasons, each with different economics.
| Motivation | Gating logic | Cost tolerance | Failure mode |
|---|---|---|---|
| Community | Serve maintainers who need help | Highest intent, lowest budget | Cuts off the 30-star project with its first outside contributor |
| Marketing | Brand visibility with influencers | Gates on stars because impressions are the point | Sybil-attackable. Burner accounts cost $4/month |
| Workflow capture | Embed in core maintainer habits | Highest budget, most deliberate | Misses the next project that matters in a year |
Anthropic’s free tier starts at 5,000 stars. That’s workflow capture. They’re buying switching costs, not impressions.
Star-gating serves the marketing motivation. It doesn’t serve the other two.
Free-for-open-source tiers have no revenue cushion regardless of motivation. A single high-volume contributor burns thousands in tokens per day. “Unlimited” and “free” and “AI-heavy” is a triple bet against volume staying low.
Contributor counts are unreliable for a different reason. Dependabot, Renovate, and every other automated dependency tool show up as contributors.
A solo developer’s repo can show four or five “contributors” that are all bots. Any gating by contributor count is measuring a number that includes machines.
Per-repo limits miss the point. Spread agent workflows across 20 repos and each stays below the threshold. The unit that matters is the user, not the repository.
Gating the free tier delays the real problem. A pro user generating 100 PRs per day is worse unit economics than a free user doing the same thing.
Three levers out
The vendor who moves first takes the short-term competitive hit. The vendor who moves last takes the financial one. Three things make the transition survivable.
Rate limits calibrated for agent speed. Not the old limits. Not 50 PRs per month. Not 500.
A single developer running native workflows produces in an afternoon what a team used to produce in a month. Calibrate for that or the limits are decorative.
Volume tiers for high-throughput users. Large bucket allocations, overage pricing, committed-spend plans. Same patterns cloud infrastructure already uses, giving power users a path forward without going back to “unlimited.”
Token efficiency. The vendors with the lowest cost per review have the most room to offer volume. At scale, the difference between a $0.03 review and a $0.30 review is the difference between a sustainable business and an emergency.
Code review is a lagging indicator. Every other per-seat tool category kept their rate limits. Code review stripped them in a VC-funded race to the bottom.
Agent-driven development ends that race. The vendors who compete on signal quality instead of “unlimited” will own this market.
The ones still burning cash on unlimited promises will discover that burn rate scales with their customers’ agent speed, not their headcount.
That adjustment is coming whether vendors plan for it or not. The only question is whether it arrives as a strategy or as an invoice.