Credits by model: what the same task costs on every AI model

How to read this table

Each row prices the same work on a different model. The work is measured, not invented: from 337 completed production tasks over 90 days (to 2026-10-02), ranked by how much they asked of the model. The median task used 78,098 input tokens (53,823 from the prompt cache) and 2,450 output tokens; a small task (25th percentile) 13,020 input tokens (0 from the prompt cache) and 2,078 output tokens; a large one (90th percentile) 87,357 input tokens (0 from the prompt cache) and 17,422 output tokens.

The credits are today's: each model's live per-token price from the AI Gateway, synced daily, converted exactly the way your bill is — so a cheaper model shows fewer credits for the same task. Dollars are at the Starter price of $0.001 per credit; Scale and Agency pay less per credit.

These are model-token credits only. Tools a task calls — web search, the Python sandbox, a hosted browser, image or video generation — cost the same on every model and are listed in credits by task type.

The same task on every model (cheapest first)

What this means for choosing models

• The default, GPT-6 Luna, costs 29 credits for the median task — one of the cheapest models on the list, which is why it is the default for new workspaces.
• Frontier models cost one to two orders of magnitude more for the same task: GPT-5.5 Pro takes 8,872 credits where GLM-4.7 Flash takes 21.
• The orchestrator writes the plan; the specialists execute it. Plan quality matters most, so a stronger orchestrator with cheap specialists is the usual split — set each agent's model on the Agents page.
• The plan pages budget 1,350–1,900 credits per task. That figure includes tools and covers most tasks on mid-priced models; on the default model most tasks use far less, and on a frontier model a large task can use more.

Frequently asked questions

Which AI model is cheapest in Agent Planners?

GLM-4.7 Flash — 21 credits for the median measured task in model tokens. The default, GPT-6 Luna, is among the cheapest too.

How much more does a premium model cost per task?

Up to about 422× the cheapest model for the same task: GPT-5.5 Pro takes 8,872 credits for the median task. Mid-tier models such as Claude Sonnet and Gemini Pro sit in between.

Are these numbers measured or estimated?

The amount of work per task is measured from 337 production tasks; the price per token is each model's live price, synced daily. The table is the product of the two, recomputed whenever the page refreshes.

Do tools cost more on an expensive model?

No. Web search, the sandbox, the hosted browser and image / video generation have their own per-call price, the same on every model. Only model tokens change with the model.

How do I see what my own tasks cost?

The Billing page shows your last 30 days of credits, broken down by model and by tool.