Costs and budgets.

A Code task can spend on more than the model answering your prompt. Web access, media generation, graph runs, review calls, and delegated work can have separate costs. Choose the model, scope the work, and inspect usage as the task progresses. This guide covers CLI 0.9.0; see Install and update for release availability.

#Read the session's usage

Inside Code, use:

/providers
/cost

/providers shows model rates when the backend supplies them, in credits per million tokens. An absent rate is unknown, not free. /cost reads the saved session log and reports recorded usage, including prior turns in a resumed session. A fork contains copied history; its report identifies that inherited usage. Do not add the parent and fork totals as if all usage were new charges.

Token counts, estimated tool costs, and reported credit charges answer different questions. The CLI usage report is useful for following the task; the account's transactions are the billing record. Pending or missing figures are not proof that nothing was charged.

#Set a limit before starting

In your terminal:

kaitoi code --max-credits 100 --max-steps 20

These are example bounds, not a price estimate. --max-credits creates a meter for this CLI invocation. Before another model step, Code checks the credits it has observed. At or above the limit it stops starting further model steps. The step already in progress finishes, so spending can pass the limit. This is not an account-wide ceiling or a reservation that guarantees the final total.

--max-steps bounds the agent's model/tool loop for a turn. It limits how long the agent can keep working, rather than setting a price per step. For durable Crew work, review the Crew run's own budget and continuation limits too.

With a 10-credit limit and 6 recorded credits, the next model step can start. If that step then reports another 6 credits, the meter reaches 12. The completed step is not reversed; the following model step is blocked. A limit is therefore a stop-before-next-step guard, not a guaranteed final charge. The illustration switches between these two states and does not execute a request.

If the last model step finishes without reporting credits, a capped invocation also stops: Code cannot enforce the budget with an unknown charge. Inspect the message and usage before deciding whether to start again. Removing the limit removes that protection; it does not resolve the missing accounting information.

#Know which work has its own cost

Work	What to account for
Hosted model conversation	Provider/model rates, input and output, and applicable cache usage.
Direct local-model inference	Zero Kaitoi inference credits; your hardware or server may have its own costs.
Managed web reads	Separate web charges can apply, including in Plan mode. A read is not necessarily free.
Media generation	Review the quote and concrete per-call approval. Saving the result locally can fail after generation has already charged.
Graph runs	Execution costs are separate from the conversation that proposed the graph.
Automatic review or compaction	Additional model work can contribute to usage; shortening context is not a free operation by definition.
Claude Code or Codex engines	Their installed CLI uses its own account. Kaitoi tools invoked by the workflow can still have separate charges.

Selecting a local main model does not move every tool or subagent to that model. Review model routing and engine selection when you combine services.

#Keep a task bounded

Give Code an acceptance condition and a stopping point: for example, inspect one bug, propose a fix, and run the relevant checks. Use Plan to investigate before granting edits, but remember that model and managed web usage can still cost credits. Request a media quote before generation, and read Crew's admission card before approving a longer run.

When a task stops on budget, inspect the work and errors already produced. A larger budget authorizes more work; it does not make an incomplete task complete or restore a failed output. Continue only after deciding what remains useful.

#Model and media rate examples

1 credit = $0.01. The following rate snapshot was checked on 23 October 2026. Use it to understand the units and compare example workloads, not as a promise of your next charge. Models, defaults, provider rates, and account markups can change. The account catalog shown by /providers or /models, the concrete media quote, and the final transaction take precedence.

#Hosted models

These example rates are in credits per million tokens. Cache reads refer to input already cached by the provider; cache writes and other applicable components can be billed separately.

Provider	Model	Input	Output	Cache read	Context window
Anthropic	Claude Sonnet 5	200	1,000	20	1M tokens
Anthropic	Claude Opus 5	500	2,500	50	1M tokens
Google	Gemini 3.8 Flash	75	375	7.5	1,048,576 tokens
OpenAI	GPT-5.6 Sol	500	3,000	50	1,050,000 tokens
OpenAI	GPT-5.6 Terra	200	1,200	20	1,050,000 tokens
OpenAI	GPT-5.6 Luna	20	120	2	1,050,000 tokens
AWS Bedrock	Claude Sonnet 5	220	1,100	22	1M tokens
AWS Bedrock	Claude Opus 5	550	2,750	55	1M tokens

At the recorded Gemini 3.8 Flash rates, 20,000 fresh input tokens and 1,000 output tokens cost 0.02 × 75 + 0.001 × 375 = 1.875 credits, approximately 1.9 credits, before any separately charged tools or other work.

OpenRouter models enabled under Settings → Integrations → OpenRouter show their rates in /models when Kaitoi has them. Direct local-model inference uses no Kaitoi inference credits; an external server or Ollama Cloud account may have its own charges.

#Web search and reading pages

Action	Recorded example price
A web search, up to 10 results	About 2 credits
Reading one web page	About 1 credit

#Media

Each example is for the settings shown. Inspect the selected model and quote before approving a call; output count, duration, quality, and provider can change the price.

Media	Model	Recorded rate	Example
Image	Nano Banana 2 Lite	$0.042 an image	4.2 credits for one image, up to 4 in a run
Image	Nano Banana 2	$0.08 an image	8 credits for one image
Image edit	Kontext Max	$0.08 an image	8 credits
Video, from a prompt or an image	MiniMax H3, 768p	$0.06 a second	30 credits for 5 seconds, 90 for 15
Speech	Gemini 3.8 Flash TTS	$0.000045 a character	4.5 credits for 1,000 characters
Speech	ElevenLabs Eleven v3	$0.0001 a character	10 credits for 1,000 characters
Music	Lyria 2	$0.0033 a second	9.9 credits for its 30-second clip
Sound effect	ElevenLabs Sound Effects	$0.0022 a second	1.1 credits for 5 seconds, up to 22 seconds
3D model	Tripo 3.1	$0.30 a model	30 credits from an image

A text-to-3D request can first generate a reference image. With the recorded Nano Banana 2 Lite and Tripo rates, that example totals 34.2 credits. The quote is an estimate; inspect the run's reported charge and account transactions after execution.