Claude Code Ships /ultrareview: Cloud Bug-Hunting Fleet

Anthropic shipped /ultrareview in Claude Code v2.1.86 last week, bundled with the Opus 4.7 launch. The slash command spins up a fleet of reviewer agents in a remote sandbox, pulls your branch or a GitHub PR, and returns a verified list of bugs before you merge. That's the product. The more interesting part is the pricing sheet: Pro and Max get three free runs that expire 5 May, Team and Enterprise get none, and every run after the trial bills $5 to $20 as extra usage. It's Anthropic's first developer SKU where the Max subscription isn't the ceiling. It's the floor.

What Actually Ships

The command is a cloud equivalent of the existing /review slash, wired to an Anthropic-managed sandbox rather than the user's local process. Invoke it with no arguments and Claude Code bundles the current branch's diff against the default branch, including uncommitted and staged changes, and uploads it to the remote runner. Pass a PR number and the sandbox clones directly from GitHub.

What makes it distinct from the dozens of AI code-review tools already on the market is the multi-agent topology. Anthropic's documentation describes a "fleet of reviewer agents" exploring the diff in parallel, with each reported finding independently reproduced before being returned to the user. That's a step up in architecture from the single-pass review model Claude Code previously shipped - and a step up in compute, which is what the $5 to $20 price floor is compensating for.

The "cloud sandbox" framing is load-bearing for the pricing argument. Every run provisions isolated compute, clones or receives the repo state, and fans out multiple reviewer agents against the diff. That's a different cost structure than a single request-response API call, and Anthropic's billing model reflects it. Source: unsplash.com

The Pricing Structure

Every tier converges on the same per-run price after the trial. The Max 20x subscriber paying $200 a month pays the same incremental $5 to $20 per review as the Pro subscriber paying $20. Anthropic's framing is that /ultrareview "bills against extra usage rather than your plan's included usage." The honest translation is that the plan you bought doesn't include this feature. It includes three demonstrations of it, and only if you're on Pro or Max.

For Team and Enterprise - the SKUs aimed at the market segment most likely to actually want multi-agent cloud review - there are no demonstrations. Extra usage must be explicitly enabled by an admin before the command will run. If it isn't enabled, Claude Code blocks the launch and routes the user to billing settings.

The $5 to $20 range is not disclosed with a scaling formula. Anthropic's language is "depending on the size of the change." Community reproductions published this week on Shigjeta.net and Mejba Ahmed's technical blog suggest the scaling is roughly linear in lines-changed up to around 2,000 LOC, at which point the sandbox appears to cap or decline. Neither source has published a formal cost curve.

Who Benefits

Three audiences come out ahead.

Anthropic first. This is the first Claude Code feature priced outside the subscription bundle, which matters for revenue architecture more than it does for any single review. It establishes that the $20/$100/$200 tiers aren't a ceiling for what a power user can spend. Simon Willison caught the earlier test where Anthropic briefly restricted Claude Code to $100+ tiers before walking it back - "the uncertainty is really bad," he wrote. The /ultrareview model is the less confrontational version of that same move: leave the ceiling where it is, add a usage-billed feature above it. A team running /ultrareview on every non-trivial PR across a fifty-engineer org clears five figures a month in extra usage without touching the seat count.

Solo engineers and small teams reviewing substantial PRs. For someone shipping a 500-line refactor they actually care about, $10 for a verified, independently-reproduced multi-agent review is cheap compared to the alternatives. GitHub Copilot's review feature is slower and catches less. Cursor's agent-review is similar in price but local. The commercial AI-first review stack - CodeRabbit, Greptile, Qodo - sits in a $15 to $30 per-seat-per-month band and doesn't offer the parallel-verification architecture Anthropic describes. On any single pre-merge review where correctness matters, /ultrareview clears the alternatives.

Anthropic's infrastructure narrative. The company spent six months building Claude Code on the web and the broader cloud-sandbox runtime. /ultrareview is the first user-facing feature that forces traffic through that runtime. The second is /ultraplan, the planning counterpart shipping alongside. Anthropic needs that runtime to have real workloads on it before the next round of enterprise conversations.

Who Pays

The ones paying are, predictably, the ones who can least afford uncertainty in their monthly engineering bill.

Team and Enterprise, from the first keystroke. No free runs. An admin must turn extra usage on. The pricing documentation's phrasing - "Claude Code blocks the launch and links you to the billing settings" - aims to make the procurement conversation trivial, but the procurement conversation still has to happen. At a fifty-engineer org, the first Monday after rollout is an Excel model week.

The economics shift when reviews become a billable line item. An engineering org that normalises /ultrareview across every PR is paying for a review product on top of a seat license - the "extra usage" framing lets Anthropic sell both without renegotiating either. Source: unsplash.com

Pro and Max subscribers past 5 May. The three free runs are described in the docs as "a one-time allotment per account." They do not refresh. They don't roll over. Once consumed, or once 5 May arrives, every invocation bills at the $5 to $20 rate. Anthropic is giving $15 to $60 of product away to price-anchor the feature - a coupon, not a plan benefit. The user who integrates /ultrareview into their muscle memory during the trial window is the user most likely to keep paying afterward. That isn't an accident.

Third-party cloud deployments. The feature is unavailable on Amazon Bedrock, Google Vertex AI, and Microsoft Foundry. For enterprises that standardised on one of those three runtimes for data-residency or procurement reasons, this is the second Claude Code feature in a quarter that requires the first-party stack to access. Task budgets and managed agents already drew the same line. /ultrareview reinforces it.

Zero Data Retention customers. Also blocked, explicitly. The sandbox runtime stores repo state and review output during execution, which is incompatible with ZDR guarantees. For the regulated-industry customers who usually enable ZDR - finance, healthcare, classified government work - this is a non-starter. The pattern of high-value Claude Code features being unavailable on ZDR is becoming the organising principle of Anthropic's compliance tier, not the exception.

What It Does Not Tell You

Two things the release framing is quiet on.

The 5-to-10-minute duration is the happy path. Anthropic's docs cite "roughly 5 to 10 minutes." The community reproductions published this week report runs extending to 15 to 20 minutes on larger diffs. That's not catastrophic - /ultrareview runs as a background task, surfacing via /tasks, and the user's session remains interactive. But it's long enough that a CI-gated workflow using /ultrareview as a pre-merge check is not a realistic pattern at current latency. The product is a human-invoked second opinion, not a block-the-merge automation.

The "extra usage" phrasing works the seam. Claude's plan pricing treats extra usage as an opt-in mechanism for the rare case when a user exceeds their included tokens. /ultrareview uses the same billing rail to sell an entirely separate product. That's a legitimate business choice - a usage-metered SKU is a fine way to price cloud compute - but the "extra usage" naming suggests a spillover it isn't. It's a line item. Anthropic's next thirty days of documentation will either clarify that or leave the ambiguity in place, and both outcomes are instructive.

The product is defensible, the engineering is a real step up over single-pass review, and the $5 to $20 price point is probably correct for the compute it consumes. What Anthropic has also done, quietly, is ship the first developer-tool SKU where the Max subscription is the starting line rather than the finish. That is the precedent to watch, not the review.

Sources:

• Find bugs with ultrareview - Claude Code Documentation
• Is Claude Code going to cost $100/month? Probably not - it's all very confusing - Simon Willison, 22 April 2026
• Claude Code Ultra Review: I Watched It Hunt Bugs - Mejba Ahmed technical blog
• Claude Code ultrareview: Deep cloud bug hunt before you merge - Shigjeta.net
• Claude Code Changelog: All Release Notes (2026) - claudefa.st
• What's New With Claude Opus 4.7: Benchmarks, /ultrareview, and the Mythos Safety Layer - Techsy
• Claude Opus 4.7 launch announcement - Anthropic