​
Pricing comparison

Evidence for buyer-cost claims in the Litepaper. Last reviewed: 23 September 2026.
Savings depend on the model. Malibu is not uniformly cheaper. This page compares a dated snapshot of the Malibu rate card with public OpenRouter list prices for the same model family.
Buyers have no payment method today. Usage runs against a daily token quota and is metered at these rates. The rates show what Malibu lists, not what you are billed today. See Get an API key → Quota and billing.

​
Live Malibu pricing

Rates are credits per million tokens, and usd_per_million_credits converts them to USD (it is 1.0 in this snapshot, so 1,000,000 credits = $1.00). For the full formula, see Metering & billing. For economic regimes, see Economics.

​
Dated snapshot: 23 September 2026

Malibu card: generated_at 2026-09-22T10:18:09Z, version bab7acfb…. OpenRouter: public GET https://openrouter.ai/api/v1/models on 23 Sep 2026. Each listing’s pricing.prompt / pricing.completion × 1e6 gives USD per 1M tokens. The blend columns price 1M prompt + 0.5M completion tokens. ● = served by at least one provider on 23 Sep 2026 (/v1/models listed mlx-community/Llama-3.2-3B-Instruct-4bit and mlx-community/Qwen3.6-27B-4bit). Every other row is priced but was not being served that day. Malibu rate-card keys: meta-llama/llama-3.2-3b-instruct, qwen3.6-27b, qwen3-8b, meta-llama/llama-3.1-8b-instruct, openai/gpt-oss-120b, qwen3.8-27b, qwen3.5-35b-a3b, google-gemma-4-26b-a4b-it, z-ai/glm-4.5-air, qwen3.6-35b-a3b, qwen3-32b, nemotron-3-nano-30b-a3b, qwen3.5-27b, qwen3-coder-30b-a3b-instruct, qwen3-30b-a3b-instruct-2507, openai/gpt-oss-20b, qwen2.5-coder-32b-instruct. OpenRouter IDs use the same family names under their vendor namespaces (meta-llama/, qwen/, openai/, google/, z-ai/, nvidia/).
A cheaper rate-card row does not mean the model can take requests today. Check GET /v1/models (with a key) or GET /v1/stats/models (without one). Quantization, latency, and receipt semantics also differ from OpenRouter, even when the family name matches. OpenRouter’s /models price is one listed price per model. Individual OpenRouter endpoints can be cheaper or more expensive.

​
How to rerun the comparison

  1. Download https://api.malibu.tech/v1/rate-card and record generated_at and version.
  2. Download OpenRouter (or another aggregator’s) list prices and record the date.
  3. Convert Malibu credits: usd_per_mtok = credits_per_mtok / 1e6 * usd_per_million_credits.
  4. Compare the same token mix. If your agent replays context with a sticky conversation tag, include prompt_cache_hit_rate_per_mtok.
  5. Leave out closed flagship models that have no OSS equivalent.

​
Default row (no model match)

A request for a model with no matching rate-card row bills at the default row: 0.50/0.50 / 1.00 per 1M prompt/completion tokens. That is usually more expensive than aggregator OSS list prices. Use a model with its own priced row.

​
When Malibu is not cheaper

  • Rows no provider is serving. A price without supply saves nothing.
  • Several 20B–35B rows. In this snapshot, gpt-oss-20b, Qwen3 30B A3B Instruct, Qwen3 Coder 30B, Qwen3.5 27B, and Nemotron 3 Nano are above OpenRouter list, and Qwen3 32B is even.
  • Latency and availability. Requests route to consumer Macs, not dedicated GPU clusters.
  • Confidential compute. Providers read prompts in plaintext, with no TEE. This page compares price only.