Claude 5 Token Vomit Tool — Pricing Impact

Vomit, a new open-source Claude Code utility, takes a completed Claude message and runs it through a second language model before displaying the rewrite. Creator Zach Ahn says he uses OpenAI’s GPT-OSS 20B locally to strip out roundabout phrasing, self-praise, awkward metaphors, and other recurring style habits.

The project reached Hacker News on August 20. It is a workflow launch, not an Anthropic price change: local post-processing can avoid a second cloud rewrite, but cannot erase Claude output tokens already consumed.

What launched

Vomit is a GPL-3.0 Go program that connects to Claude Code’s MessageDisplay hook. It buffers a message, sends the complete text and an editing prompt to an OpenAI-compatible chat-completions endpoint, and substitutes the returned prose in Claude Code’s display.

The default recommendation is Llama.app with GPT-OSS 20B on the user’s Mac. Ollama and other OpenAI-compatible endpoints may also work. The repository describes the project as local, without telemetry or external dependencies, when it is configured against a local endpoint.

Its safer tail mode shows rewrites separately. The README warns that the editor can hallucinate, run slowly, or omit part of Claude’s message.

What it costs—and what it does not save

The live table and chart above show the current tracked Claude 5 rates. Those charges happen before Vomit sees the message.

Workflow	Inference stages	Incremental hosted-token cost	Main trade-off
Claude only	One	None	Read the original output
Claude plus local Vomit rewrite	Two	None for the rewrite	Local compute, delay, fidelity risk
Claude plus paid cloud rewrite	Two	Second model input and output	Extra API bill and data sent to another endpoint
Ask Claude to rewrite itself	Two Claude turns	Additional Claude input and output	More Claude usage and session context

Vomit receives text after Claude has generated it. It therefore cannot reduce the original output charge, shrink the original request, or refund tokens. Its potential savings come from replacing a second paid editing turn and from reducing the human time spent parsing a verbose response.

The tracked Sonnet 5, Opus 5, and Fable 5 rows still match Anthropic’s official rate card. Use the live table, chart, and calculator above for the current cost of the original response; the second pass cannot reverse that charge.

If the configured endpoint is remote rather than local, the rewrite is no longer free of hosted-token cost. The full Claude message becomes input to the editor, and the cleaned version becomes new output. Use the token calculator with the Anthropic pricing page and the editor endpoint’s rates before calling that route cheaper.

Vomit is not an Anthropic model, endpoint, subscription, or rate card. It therefore does not receive a row in the canonical pricing dataset or public pricing API. The maintained Claude rows remain the correct source for the first inference pass.

Labs availability blocker

Vomit is an output-transforming harness, not a model endpoint, so it cannot be ranked beside Claude models in the Labs leaderboard. A valid test needs paired Claude Code tasks and pinned model, prompt, endpoint, hardware, and output settings, then must grade factual preservation, task acceptance, latency, and human correction time. The project does not publish that replay package, so Labs records a harness blocker rather than inventing a score.

Who benefits—and who loses

Developers already running a local model are the clearest beneficiaries. They can add a style-focused editing pass without another metered API call, and the smaller model does not need repository access because it only edits Claude’s message.

Teams that need exact, auditable output take on more risk. A rewrite can remove a caveat, change technical emphasis, or make an uncertain statement sound definitive. One Hacker News reader noted that a similar rewrite appeared to lose an idempotency detail; the broader point applies even when the prose looks better.

Users without suitable local hardware also lose the clean cost story. They must accept slower local inference or pay another provider. Compare current OpenAI pricing and open-model routes rather than assuming every OpenAI-compatible endpoint is local or free.

What Claude users should do

1. Do not install Vomit expecting it to lower already-generated Claude tokens. Reduce output limits, tool chatter, and unnecessary turns at the source if the API bill is the problem.
2. Start with vomit tail, keep the original visible, and review rewrites for missing warnings, commands, file names, and numeric details.
3. Benchmark accepted-task time, not prose length alone. Record rewrite latency, corrections, and whether developers still reopen the original message.
4. Treat the configured endpoint as a data boundary. A local URL keeps the text on the machine; a remote OpenAI-compatible URL sends the message to that service.
5. Review our earlier analysis of Claude Code token overhead if the larger problem is hidden input context rather than verbose displayed output.

For teams that cannot run the editor locally, compare managed open-model endpoints at Novita, then verify model compatibility, retention policy, and the total cost of both rewrite input and output.

Affiliate disclosure: AI Pricing Guru may earn a commission from the sponsored Novita link at no extra cost to you. It does not affect this analysis.

Bottom line

Vomit is an inventive response to a usability complaint, not a Claude discount. Running the editor locally keeps the incremental hosted-token charge at zero, but Claude’s original output remains on the bill and a second model introduces latency and fidelity risk. Use it to save attention or avoid a paid cleanup turn, not as a token-refund mechanism.

Sources: Vomit on GitHub, Zach Ahn’s launch post, the Hacker News discussion, Anthropic’s official Claude pricing documentation, and Anthropic’s models overview. Sources and prices checked August 21, 2026.