Skip to content
CODOROby Osokoro
Models and cost control — product preview

Where the money goes, and what stops it.

Two ways to pay for tokens, one place to see them. Featured models appear first and every model available through the live Vercel AI Gateway catalogue is searchable. On runs started through Codoro, the ledger is the same and caps are enforced before the call, not after the invoice.

This page describes the launch model, not a public checkout. Until a verified release manifest and staging billing proof are published, there is no public download or paid activation.

Extension boundary

In VS Code-compatible hosts, the stable language-model API exposes neither provider prices nor a usage total, and an extension cannot intercept the host's own agent mode. Codoro therefore governs the runs it starts, estimates them from its price table, and re-tokenises returned text as evidence. Calls started by the host's agent remain outside this ledger.

Option A

Your own API keys

Use your own Anthropic, OpenAI, DeepSeek, Google, xAI, Moonshot, Z.ai, Alibaba or Mistral API key. The provider bills you and Codoro takes 0%. Keys stay in the macOS keychain and are sent only to the vendor that issued them.

Option B

Codoro credits

Hosted usage debits the selected model's published retail rate from one dollar-valued Codoro balance. A $28 top-up adds $28 Codoro credits; top-ups do not expire and nothing auto-charges when they run out.

Featured provider list rates

USD per million tokens · before the hosted markup
ModelInOutRouted to by default
Claude Fable 5$10.00$50.00Architecture, hard debugging
Claude Opus 5$5.00$25.00Architecture, hard debugging
Claude Sonnet 5$2.00$10.00Everyday multi-file edits
Claude Haiku 4.5$1.00$5.00Renames, tests, commit messages
GPT-5.6 Sol$2.00$10.00Complex professional work
GPT-5.6 Terra$2.00$12.00Balanced coding work
GPT-5.6 Luna$0.20$1.20Cost-sensitive high-volume work
Gemini 3.1 Pro Preview$2.00$12.00Complex professional work
Gemini 3.7 Flash$0.75$3.75Balanced coding work
Grok 4.6$2.00$6.00Architecture, hard debugging
Grok 4.3$1.25$2.50Bulk refactors on a budget
Kimi K2.7 Code$0.95$4.00Everyday multi-file edits
GLM 5.2$1.40$4.40Balanced coding work
DeepSeek V4 Pro$0.66$1.98Bulk refactors on a budget
Codoro Fast$0.22$0.66High-volume mechanical work

This featured snapshot comes from Codoro's versioned Gateway catalogue. The in-app selector shows the selected model's final retail rate after its published markup; check that live rate before relying on an estimate.

Routing rules you write, in the repository

Automatic routing is a default, not a black box. The rules live in a file you can read, review and revert like anything else in the project.

# .codoro/routing.toml
[default]
model = "claude-sonnet-5"
cap_per_run_usd = 2.00

[rule.cheap_edits]
when = "files_changed < 3 and no_tests_touched"
model = "claude-haiku-4-5"

[rule.deep]
when = "task.label == 'architecture'"
model = "claude-opus-4-8"
cap_per_run_usd = 8.00

[rule.private]
when = "path matches 'infra/secrets/**'"
model = "codoro-fast"

Reserve, then settle

The estimate is held against your balance before the request, and reconciled against real usage when it returns.

Pause, never downgrade silently

At the cap the run stops and offers a cheaper model with the cost of continuing shown next to it.

Pooled team budgets

One monthly pool, a per-seat ceiling, and a ledger by developer, repository and model.

Export the ledger

CSV or the API, per run, with developer, repository, model, token usage and settled cost.

Voice and design, on the same ledger

Talking to the agent is not a second meter, and neither is moving something on the canvas. A spoken turn and a design edit are reserved, settled and capped exactly as a typed request is, on the balance this page has already described.

Recognized text uses the text ledger

On a target with supported speech recognition, the resulting text enters the same governed request path as typed text. Microphone capture and macOS system narration stay local and are not represented as a separate hosted-audio charge.

Unsupported recognition stays hidden

The desktop always exposes permission and microphone level state, but it renders speech-to-text controls only when its current host supplies a recognition engine. Local diff narration remains available through system speech.

A canvas edit is the same arrangement. Apply to code shows its estimate before it runs, reserves it, and settles against real usage after — on the same balance and the same per-run cap, whether the request was typed, spoken or dragged.

One account, one invoice

Sign in with the Osokoro account you already have. Codoro is billed on its own plan and appears as a separate line on the same invoice, so there is one card on file and one place to see what you spent.