Cursor 3's PR Review Tab and the Bugbot Pricing Pivot

What changed in Cursor 3 this week

Anysphere's Cursor 3 release headline grabbed attention for one feature — Build in Parallel, the async subagent orchestrator that runs independent parts of a plan simultaneously. We covered that in depth over at Cursor 3: Parallel Agents Are the New IDE Surface, and it is genuinely the more spectacular bit of the release. But for any builder who has to write a procurement memo this quarter, the quieter changes matter more.

The three that should reshape your workflow planning are: a brand new PR Review tab inside the IDE, pinnable skill pills for the actions you actually repeat, and a context usage breakdown that finally tells you which subagent is hogging your tokens. Sitting alongside those is a billing change that flew under most radars — Bugbot has dropped seat fees entirely and moved to pure usage-based pricing. For teams in Bengaluru working to a tight per-developer envelope, and for UK enterprises whose finance teams measure tools in pounds-per-seat-per-year, the procurement maths just shifted.

This piece is the one to read if you already understand the parallel-agent angle and now need to plan budgets, redraw your code-review process, and decide whether Cursor stays your daily driver or moves to a split routing setup with Claude Code's /autopilot.

PR Review inside the IDE: what actually changes

The new PR Review tab is the feature that closes a loop Cursor had left half-open for two years. Until this release, the agent could draft a branch, open a PR and wait — and then the human round-trip happened in GitHub or GitLab in a browser tab. Reviewer comments, inline threads, top-level review notes: all of that lived outside the editor. The agent that wrote the code had no idea what reviewers had said unless you pasted the feedback back in.

Cursor 3 pulls the PR thread into the IDE alongside the working tree. You see the inline review comments next to the diff. The top-level reviewer note appears in the same surface as the agent's plan. Crucially, the agent can read those threads as context for its next iteration. The write-review-amend loop now happens without a context switch, and that single change is what will reshape solo-dev and small-team workflows over the next quarter.

What it means for solo developers

For one-person shops — and there are a lot in this audience, from Pune indie hackers to London freelancers — PR Review in the IDE removes the last reason to use GitHub's web UI day to day. You can self-review your own PRs in Cursor, run Bugbot against the diff for an automated pass, and merge. The browser tab opens once a week instead of forty times a day. The friction reduction is real, and for context-heavy work where every screen switch was a thought-state loss, this is the single biggest productivity win of the release.

What it means for teams

For teams, the picture is more interesting. The agent can now respond to reviewer comments — and that raises an immediate governance question. Should agents reply to senior engineers in PR threads? Should they amend code in response to a reviewer note without human approval? Should they leave their own counter-comments on a thread? Anysphere has, sensibly, left the policy choices to the team, but the defaults are permissive.

UK enterprises with formal code-review charters — banks, NHS digital, public-sector contractors — will need to update those charters this quarter. Indian agencies running large-team workflows for overseas clients face a similar question, particularly where audit trails on who-wrote-what matter for IP attribution. The technical capability is genuinely useful; the operating model around it is not yet settled.

The Bugbot pricing pivot — and why it matters more than you think

This is the bit Anysphere did not put in the headline. Bugbot — the automated pull-request reviewer that has been a quiet workhorse since 2025 — has dropped its seat fee entirely. There is no $40-per-developer-per-month line item any more. Teams are billed purely from their on-demand usage pool. Individuals are billed from their plan's included frontier-model allowance.

This mirrors almost exactly what GitHub did with Copilot, as we covered in GitHub Copilot Goes Usage-Based June 1: Budget Now. The signal here is no longer industry-specific — it is industry-wide. Every major AI coding tool is converging on usage billing because the underlying inference economics demand it. Flat seat pricing made sense when the marginal cost of an extra query was effectively zero. With agents that can drop $5 of inference on a single bug-hunt, it does not.

What pure usage billing costs in practice

For a Pro user at $20 per month, the included frontier-model usage covers a useful amount of Bugbot activity. Anysphere's documentation gives rule-of-thumb figures: a single Bugbot pass over a typical PR runs between $0.15 and $0.80 depending on diff size and model choice. That puts you somewhere between 25 and 130 Bugbot runs inside the included envelope. For a solo developer or a freelancer, that is probably enough.

For a team of ten engineers each opening four PRs a day, the maths is different. At an average $0.40 per Bugbot pass, that is $80 a day, $1,600 a month — and that is before any other agent activity. Teams plan at $40 per user per month, so a ten-person team has $400 of seat budget. The on-demand spend will land on top.

The five Cursor plans, side by side

Cursor 3 keeps the same five-tier plan structure that landed earlier in 2026, but the per-tier value calculation has shifted now that Bugbot folds into included usage. Here is the comparison most procurement teams ask for.

Plan	Monthly price	Frontier-model included	Best for	Bugbot impact
Hobby	$0	Limited — best-effort, throttled	Trying Cursor; weekend projects	Effectively unavailable at any volume
Pro	$20/mo	$20 of frontier model usage	Solo devs; light freelance work	25–130 PRs per month covered
Pro+	$60/mo	$60 of frontier model usage	Daily-driver builders; small studios	75–400 PRs per month covered
Ultra	$200/mo	$200 of frontier model usage	Heavy parallel-agent workloads	250–1,300 PRs per month covered
Teams	$40/user/mo	Pooled across the workspace	Companies wanting central billing and admin controls	On-demand spend pooled centrally

Two things to note from this table. First, the per-dollar value of frontier usage is identical across Pro, Pro+ and Ultra — you pay $1 of plan for $1 of inference. That is unusual for a SaaS pricing ladder, where you typically expect bulk-tier discounts. Anysphere is signalling that the tier choice is about headroom, not unit cost. Second, the Teams plan at $40 per user pools usage across the whole workspace, which is materially better for teams whose engineers have uneven workloads. A senior architect who runs heavy parallel agents two days a week can borrow from a designer-engineer who barely uses the agent at all.

Skill pills and the context usage breakdown

Two smaller features deserve attention because they affect how you operate the tool day to day. Skill pills let you pin your most-used skills as quick-action buttons next to the chat composer. If you find yourself typing "run the test suite", "open a PR with a conventional commit message" or "scaffold a new endpoint matching our pattern" five times a day, that is now a one-click pill. The pill is workspace-scoped, so your repo's conventions stay encoded in the UI rather than in muscle memory.

The context usage breakdown is the feature that experienced builders will quietly love. It shows you, per agent, exactly where your context window is going — how much is rules, how much is skills, how much is MCP server responses, how much is each subagent's working memory. Until now, diagnosing why an agent had run out of context was guesswork. Now it is a panel. Combined with the Pro plan's $20 ceiling, this is the feature that lets you actually engineer for cost rather than discover overrun.

Enterprise admin controls, finally fit for procurement

Cursor 3 also lands the enterprise admin features that UK and Indian procurement teams have been asking for. Model access can be restricted — useful when your data-residency policy bans US-hosted frontier models. Spend limits are now softer, with usage alerts instead of hard cutoffs, so an engineer doesn't get bricked mid-deploy. Admins can block specific providers or even specific model configurations more granularly than the old all-or-nothing toggle allowed.

For UK public-sector contractors who have to satisfy data-residency under the forthcoming UK AI Code of Practice, this matters. For Indian agencies serving regulated overseas clients — particularly finance and healthcare — the granular provider blocks are the difference between "we can use Cursor" and "we cannot use Cursor". Procurement folk should read the admin docs alongside the release notes.

Cursor versus the field: where it wins, where it still loses

This release sharpens the competitive picture rather than blurring it. Cursor now clearly owns the in-IDE PR review loop in a way no rival does. Anyone who tried our earlier comparison — Cursor Composer 2 vs Claude Code — will recognise that the review-loop gap was Cursor's biggest weak spot, and it has now been closed.

Where Cursor still loses is the CLI workflow. Claude Code remains the better tool for long-running batch refactors, scripted automation, and anything you want to run without an IDE open. The new /autopilot default makes it more aggressive than Cursor's interactive agents, and for engineers who live in the terminal that is the right shape. GitHub Copilot, meanwhile, retains the integration advantage on GitHub Enterprise Server installations and remains the procurement default for companies already paying for Copilot.

The pragmatic answer for most well-resourced teams in May 2026 is no longer "pick one". It is split routing: Claude Code for batch and scripted work, Cursor for the review and merge loop, Copilot for organisations whose ELA already covers it. The acquisition speculation in SpaceX's $60B Cursor option is a reminder that the strategic landscape can shift again at any time, and Cursor 3's quiet billing pivot is precisely the kind of move you make when you are sharpening for sale or for the next funding round.

What to do this week

1. Audit Bugbot usage. Pull the last 30 days of Bugbot invocations from your team workspace. Identify the largest five PRs by diff size — they will be your top-five cost drivers under the new billing.
2. Set diff-size gates. Add a workspace rule that only triggers Bugbot on PRs above 50 lines, or only on PRs tagged for-review. This single change can cut Bugbot spend by 40-60% for a typical team.
3. Pin your workflow skills. Identify the five things every engineer on your team types daily, and pin them as skill pills. Make this a team-shared workspace setting so conventions stay consistent.
4. Instrument context usage. Use the new breakdown panel for a week. Identify the worst-offending subagent or MCP server and tighten its prompt or scope before you upgrade plan tier.
5. Update your code-review charter. Decide — explicitly — whether agents may respond to human reviewer comments in PR threads. Write the policy down. Add it to your team handbook.

Cursor 3 is the release that turns Cursor from "a really good IDE for AI-assisted coding" into "the surface where the whole write-review-merge loop happens". The parallel agents got the headline. The PR Review tab and the Bugbot billing pivot will reshape your workflow and your invoices. Plan accordingly.

Primary sources: cursor.com/docs/models-and-pricing, releasebot.io/updates/cursor May 2026 release notes, and vantage.sh pricing analysis.