Reference

AI model pricing & credits ranking

TL;DR

Every agent can run on a different model, and each model costs a different number of credits. This ranking groups the selectable models into three cost classes — $ Cheap, $$ Mid, $$$ Premium — so you can pick the cheapest model that still does the job. The exact charge is billed live from the AI Gateway's real per-token price; classes are the relative guide.

How model pricing works here

These prices are the AI model providers' own per-token rates — not a markup we add. A premium model like GPT-5.5 Pro or Claude Opus simply costs the provider (OpenAI / Anthropic) ~$120 per million output tokens at the source; a cheap model like Gemini Flash costs ~$4. Agent Planners passes that through: we meter the exact tokens each run uses and bill the provider's real cost. You only ever pay premium-model prices on the agents you deliberately set to a premium model — keep them on a $ model and you pay $ prices.

Usage is metered in credits. A credit is a fixed slice of real model + tool cost, so an expensive model (GPT-5.5 Pro, Claude Opus) costs proportionally more credits than a cheap one (Gemini Flash, DeepSeek) for the same work — billed live from the AI Gateway's per-token price.

Output tokens dominate the cost, and cheaper models can be 20–30× less per token than premium ones. Because you set the model per agent, the simplest way to cut spend is to keep your specialist agents on a $ model and reserve premium models for the orchestrator (the planner) where quality matters most.

Want it in dollars? Credits convert at the $0.001/credit list price (100,000 credits = $100). So a $ model billed at ~4 credits per 1,000 output tokens is about $4 per 1,000,000 output tokens, versus ~$120/1M for a premium model. The ranking table below shows both credits and this ≈ $ reference per model.

Cost classes

ClassApprox. output cost≈ Price / 1M outputBest for
$ Cheap~4–8 credits / 1K output$6.00Lowest credit cost — ideal for specialist agents.
$$ Mid~12–60 credits / 1K output$36.00Balanced cost and capability.
$$$ Premium~120 credits / 1K output$120.00Highest cost — reserve for the hardest work.
The ≈ Price / 1M output is the AI provider's own token price (e.g. ~$120/1M for a frontier model at the source) — passed through, not marked up by Agent Planners. Credit ranges are approximate (the metering fallback); the exact charge comes from the AI Gateway's live per-token price. The $ column converts credits at the $0.001/credit list price (100,000 credits = $100).

Model cost ranking (cheapest first)

ModelProviderClassCredits / 1K (in · out)≈ $ / 1M (in · out)Reasoning
DeepSeek V4 FlashDeepSeek$ Cheap1 · 4$1.00 · $4.00
Gemini 2.5 Flash (cheap)Google$ Cheap1 · 4$1.00 · $4.00
Gemini 3.1 Flash LiteGoogle$ Cheap1 · 4$1.00 · $4.00
Gemini 3.5 Flash (default)Google$ Cheap1 · 4$1.00 · $4.00
Qwen3.5 Flash (cheap)Alibaba (Qwen)$ Cheap2 · 8$2.00 · $8.00
Claude Haiku 4.5Anthropic$ Cheap2 · 8$2.00 · $8.00
GPT-5 miniOpenAI$ Cheap2 · 8$2.00 · $8.00Yes
GPT-5.4 miniOpenAI$ Cheap2 · 8$2.00 · $8.00Yes
Grok 4.1 Fast (reasoning)xAI$ Cheap2 · 8$2.00 · $8.00Yes
Llama 3.3 70B (cheap)Meta (Llama)$ Cheap5 · 20$5.00 · $20.00
GLM-4.7 Flash (cheap)Zhipu (GLM)$ Cheap5 · 20$5.00 · $20.00
DeepSeek R1 (reasoning)DeepSeek$$ Mid1 · 4$1.00 · $4.00Yes
DeepSeek V3.2 (cheap)DeepSeek$$ Mid1 · 4$1.00 · $4.00
DeepSeek V3.2 Thinking (reasoning)DeepSeek$$ Mid1 · 4$1.00 · $4.00Yes
DeepSeek V4 ProDeepSeek$$ Mid1 · 4$1.00 · $4.00
Qwen3 Max Thinking (reasoning)Alibaba (Qwen)$$ Mid2 · 8$2.00 · $8.00Yes
Qwen3.7 MaxAlibaba (Qwen)$$ Mid2 · 8$2.00 · $8.00
Qwen3.7 PlusAlibaba (Qwen)$$ Mid2 · 8$2.00 · $8.00
Gemini 2.5 ProGoogle$$ Mid3 · 12$3.00 · $12.00
Llama 4 MaverickMeta (Llama)$$ Mid5 · 20$5.00 · $20.00
MiniMax M2.7MiniMax$$ Mid5 · 20$5.00 · $20.00
MiniMax M3MiniMax$$ Mid5 · 20$5.00 · $20.00
Magistral Medium (reasoning)Mistral$$ Mid5 · 20$5.00 · $20.00Yes
Mistral Large 3Mistral$$ Mid5 · 20$5.00 · $20.00
Mistral Medium 3.5Mistral$$ Mid5 · 20$5.00 · $20.00
Kimi K2 Thinking (reasoning)Moonshot (Kimi)$$ Mid5 · 20$5.00 · $20.00Yes
Kimi K2.6Moonshot (Kimi)$$ Mid5 · 20$5.00 · $20.00
Grok 4.3xAI$$ Mid5 · 20$5.00 · $20.00
GLM-5.1Zhipu (GLM)$$ Mid5 · 20$5.00 · $20.00
GLM-5.2Zhipu (GLM)$$ Mid5 · 20$5.00 · $20.00
Claude Sonnet 4.6Anthropic$$ Mid12 · 60$12.00 · $60.00Yes
Claude Sonnet 5Anthropic$$ Mid12 · 60$12.00 · $60.00
Claude Opus 4.6Anthropic$$ Mid30 · 120$30.00 · $120.00Yes
GPT-5OpenAI$$ Mid30 · 120$30.00 · $120.00Yes
GPT-5.2 (cheaper)OpenAI$$ Mid30 · 120$30.00 · $120.00Yes
GPT-5.4OpenAI$$ Mid30 · 120$30.00 · $120.00Yes
GPT-5.5OpenAI$$ Mid30 · 120$30.00 · $120.00Yes
Gemini 3 Pro (preview)Google$$$ Premium3 · 12$3.00 · $12.00
Gemini 3.1 Pro (preview)Google$$$ Premium3 · 12$3.00 · $12.00
Claude Opus 4.7Anthropic$$$ Premium30 · 120$30.00 · $120.00Yes
Claude Opus 4.8Anthropic$$$ Premium30 · 120$30.00 · $120.00Yes
GPT-5.5 Pro (expensive)OpenAI$$$ Premium30 · 120$30.00 · $120.00Yes
Credits / 1K = how many credits 1,000 input·output tokens cost (the metering rate from rates.ts); ≈ $ / 1M converts that to dollars per million tokens at the $0.001/credit list price. Output tokens dominate spend. Exact charges are billed live from the AI Gateway's per-token price and can differ. Model list as of June 2026 (42 models). Switch any agent's model on the Agents page; reasoning models ignore the temperature setting.

How to choose

  • Specialist agents (Analytics, Optimization, Creative, Research, Safety, Report) do well on a $ Cheap model — that's the default.
  • Keep the orchestrator on a stronger model: plan quality drives the whole run.
  • Reasoning models cost the most per token — reserve them for genuinely hard analysis.
  • Running low on credits? Drop your specialists to a $ model to stretch the balance — see credits & billing and best practices.

Frequently asked questions

Which AI model is cheapest?
The $ Cheap class (e.g. Gemini Flash, DeepSeek, Llama 3.3 70B, GLM Flash) costs the fewest credits — typically ~4–8 credits per 1K output tokens, versus ~120 for premium models like GPT-5.5 Pro or Claude Opus.
Does a cheaper model hurt quality?
For executing the planner's steps (the specialist agents), cheap fast models are reliable — that's why they're the default. Plan quality matters most for the orchestrator, so keep that on a stronger model.
How are credits charged per model?
Per token: input + output tokens × the model's live AI Gateway price, converted to credits. A typical task runs ~400–800 credits; premium models cost much more per token than cheap ones.
How much is that in dollars?
Credits convert at the $0.001/credit list price (100,000 credits = $100). A $ Cheap model (~4 credits per 1K output tokens) is about $4 per 1M output tokens; a premium model (~120 credits per 1K) is about $120 per 1M. The ranking table shows the ≈ $ / 1M reference for every model. Exact charges are billed live from the AI Gateway's per-token price and can differ.
Is Agent Planners marking up the model cost?
No. The $120/1M for a premium model is OpenAI's / Anthropic's own published per-token price at the source — we pass it through and meter the exact tokens your run uses. The big numbers in the Premium row are how much frontier models genuinely cost the provider, not an extra fee we add. You only pay them on agents you set to a premium model; the default cheap models cost ~$4/1M.
Can I set a different model per agent?
Yes — every agent's model is switchable on the Agents page through the built-in AI Gateway, so you can mix a premium orchestrator with cheap specialists.
Pick the right model for each agent

Set a cheap model for specialists and a strong one for the orchestrator — start free, no card.