Claude Code Tries Claude App Rewrite: Pricing Impact

Boris Cherny, who leads Claude Code at Anthropic, has revealed an unusually long coding-agent experiment: ask Claude to reproduce the company’s Electron desktop app in Swift, let it inspect both versions in a Mac virtual machine, and keep comparing screenshots until the work is done.

The eye-catching detail is scale. During a Startup School 2026 interview, Cherny said the task was still running after about 14–15 days. Asked how many agents it had spawned, he said he was not sure and guessed “thousands, tens of thousands.” That is evidence that agents can persist and fan out on difficult software work—not evidence that the rewrite succeeded, that Anthropic will ship it, or that buyers should run uncapped agents for weeks.

For current costs, use the generated comparison above, our Anthropic pricing page, OpenAI pricing page, and token calculator. The source did not name the model behind the experiment, so no listed API rate can be used to calculate Anthropic’s actual spend.

What happened

Cherny said he started a Claude Tag session, Anthropic’s Claude-in-Slack product, as an experiment to see how the desktop app might feel if it were native. He connected the agent to a macOS runner on GitHub, gave it access to an empty Swift codebase, and asked it to run the existing Electron app in the virtual machine.

The core instruction was to screenshot the Electron app, inspect it pixel by pixel, compare it with the Swift version, and continue until finished. Reports have framed the story as Claude Code rewriting Claude, but the primary interview describes Claude Tag as the operating surface. Cherny did not disclose the underlying model, token volume, compute usage, human interventions, current progress, or a completion date. His agent-count estimate was explicitly a guess, not a measured usage report.

Verified detail	What it means	What it does not prove
Claude received a Mac runner and repository access	Tools and environment setup enabled the task	A prompt alone performed the rewrite
Screenshot comparison supplied a feedback loop	The agent had a way to inspect visual differences	Pixel matching guarantees native Mac design
The session persisted for roughly two weeks	Long-horizon agent work is possible	The run was continuously productive
Cherny guessed the workflow spawned thousands of agents	Dynamic workflows can fan out extensively	Anthropic reported an exact agent count or bill
No price or product launch was announced	Existing Claude budgets still apply	A native Claude app is shipping

Pricing impact

There is no Claude price change in this announcement. The financial signal is that elapsed time and agent count are poor proxies for delivered value. A long-lived, highly parallel workflow can consume model tokens, runner time, test resources, logs, storage, and reviewer attention while still being far from an acceptable result.

The live table and chart above compare current API routes for a fixed workload, while the inline calculator lets you set a bounded token assumption. They are planning tools only. Claude Tag is a Team and Enterprise surface with organization-level spend controls, and its billing cannot be reconstructed from public API rates without usage data. See our earlier Claude Tag pricing analysis for the product context.

Anthropic’s official model-pricing documentation still lists the same current public Claude ladder shown above. The canonical pricing dataset and public pricing API match those official rates byte for byte.

Labs status: public Claude included, private Tag run blocked

The AI Pricing Guru Labs leaderboard already includes Claude Fable 5, Opus 5, Sonnet 5, and Haiku 4.5 on a deterministic 49-task suite. In the latest accepted run, Sonnet 5 and Opus 5 each answered 48 of 49 tasks correctly with zero API errors; Fable 5 answered 46 of 49 with zero API errors.

Those public-model results do not validate a 15-day application rewrite. A launch-specific Labs run is explicitly blocked because Claude Tag is not a public model endpoint, the model snapshot is undisclosed, the source code and Mac runner are private, the task was unfinished, and Anthropic published no artifact or acceptance test to reproduce. We will not relabel a public Claude result as evidence for this private workflow.

What this means

Cherny’s broader lesson was verification: the important skill is making a difficult task check its own work along the way. The experiment does that with screenshots, but visual similarity is only one acceptance test. A desktop rewrite also needs functional tests, accessibility checks, performance targets, crash reporting, native interaction review, and human product judgment.

John Gruber’s Daring Fireball commentary highlights the design risk. Reproducing an Electron interface pixel for pixel in Swift may preserve non-native choices instead of using AppKit or SwiftUI well. In cost terms, an agent can efficiently optimize the wrong target unless the specification measures the product outcome—not merely visual similarity.

Who benefits—and who loses

Agent-platform teams benefit because this is a concrete example of persistent tool use: repository access, a macOS runner, visual inspection, and asynchronous work coordinated through Slack. Engineering leaders also get a useful evaluation pattern for migrations and UI parity work.

Teams with weak tests or unlimited agent loops lose. So do buyers who mistake duration for progress. The source provides no shipped artifact or quality score, making it impossible to compare the experiment’s cost with a human-led rewrite.

What developers should do now

1. Define completion with functional, visual, performance, and accessibility tests before starting an agent.
2. Set token, runner-hour, and wall-clock limits, with human review at fixed checkpoints.
3. Track accepted changes, test-pass rate, regressions, retries, and reviewer time—not only session duration.
4. Separate parity work from redesign. First decide which behavior must remain and which interface choices should become native.
5. Stop or re-scope when the same failure repeats; “do not stop” is not a production budget policy.

For a broader comparison, read our best AI for coding guide. If you need an independent quota for testing long-running coding workflows, compare the Z.ai coding plan on a bounded repository task before using it as a fallback.

Affiliate disclosure: we may earn a commission from the sponsored link above. It does not affect the analysis.

Bottom line

The Claude app rewrite experiment is a strong demonstration of agent persistence and a warning about open-ended objectives. Claude had tools and a visual feedback loop, yet the task remained unfinished after roughly two weeks. The practical win is not “agents can replace an app team”; it is that verification, budgets, and stopping rules determine whether long-running agents create value or merely continue running.

Sources: Root Access transcript of the Boris Cherny interview, Startup School 2026 video at the relevant segment, Daring Fireball commentary, the Hacker News discussion, Anthropic’s official model pricing documentation, Anthropic’s official pricing page, and AI Pricing Guru’s live pricing dataset.