Model Catalog

The Model Router is the single endpoint your applications and agents call to reach every AI model Performance Hub offers. You use one API key and one base URL; we handle the upstream provider accounts, rate limits, and billing.

This page is the live price list. Prices are per 1 million tokens in US dollars and already include the 5 percent platform fee, so the figures below are exactly what your facility is billed. The catalog refreshes automatically several times a day and this page reads it live.

Loading the model catalog...

Reading the table

ColumnWhat it means
ContextThe largest amount of input the model accepts in one request, in tokens
Input / 1MRetail price per million input tokens
Output / 1MRetail price per million output tokens, usually several times the input price
IntelligenceArtificial Analysis intelligence index, 0 to 100. A third-party composite benchmark score, shown as a dash for models they do not track
ValueIntelligence divided by blended price. Higher means more capability per dollar

Blended price is (3 x input + output) / 4, which approximates a typical workload where input tokens outnumber output tokens by roughly three to one.

The badges on each model show its capabilities: Vision accepts image input, Reasoning produces reasoning tokens, and Frontier marks a flagship model from a major lab. Provider groups show the region their models are served from, which matters if you have data-sovereignty requirements.

How you are charged

Model usage is billed per token, at the rate published above for the model that served the request. Tokens are counted separately for input and output, because providers price them differently, and output is usually several times more expensive than input.

ComponentMeaning
Input tokensEverything sent to the model: your prompt, system message, conversation history, tool definitions, and any image or document content
Output tokensEverything the model generates, including reasoning tokens on models that produce them
Cache readInput tokens served from the provider's prompt cache, where supported. Much cheaper than fresh input
Cache writeThe cost of placing tokens into the prompt cache, where supported. Slightly more than fresh input

A charge is calculated as:

cost = (input tokens / 1,000,000) x input price
     + (output tokens / 1,000,000) x output price

Charges are deducted from your prepaid balance as requests complete. There is no monthly minimum and no charge for models you do not use.

The 5 percent platform fee

We buy model capacity from upstream providers and resell it at the upstream price plus a flat 5 percent. That 5 percent is the whole of our margin on model usage. It covers what sits around the model call: the routing layer, API key issuing and rotation, per-key and per-agent limits, guardrails, request logging, usage metering, billing and top-ups, and support. It does not vary by model, by provider, or by how much you spend.

A worked example

Take a model published at $5.25 per 1M input tokens and $26.25 per 1M output tokens:

StepInputOutput
Upstream provider price per 1M$5.00$25.00
Platform fee at 5 percent$0.25$1.25
Published price per 1M$5.25$26.25

A request using 12,000 input tokens and 800 output tokens costs:

input:  12,000 / 1,000,000 x $5.25  = $0.0630
output:    800 / 1,000,000 x $26.25 = $0.0210
total                                = $0.0840

The same request on a budget model published at $0.032 in and $0.137 out costs about $0.0005, which is why model choice matters far more than the fee.

How the Auto Router is charged

The Auto Router (auto, shown at the top of this page) has no price of its own. It inspects each request, routes it to a model based on complexity, and you are billed at that model's published rate. Simple prompts land on fast, cheap models; only genuinely hard work reaches the expensive ones.

Each facility can remap the tiers to its own preferred models, which changes what its auto requests cost. See Models for how to edit the routing.

When prices change

Model prices are refreshed automatically from upstream providers several times a day. When a provider changes a price, the published price follows it, with the same 5 percent applied. We do not issue price change notices for upstream movements, so if you are budgeting tightly, check this page rather than a figure you noted previously.

New models are added when they become generally available. Models are withdrawn when the upstream provider retires them.

What is not charged

  • Enabling or disabling models for your facility.
  • Creating, rotating, or revoking API keys.
  • Guardrail evaluation on your requests.
  • Activity, logs, and usage exports.
  • Failed requests that never reached a model, such as an authentication error or a request blocked by a guardrail or a spend limit. If a request reaches the model and the model responds, it is billed, even if your application discards the response.

Choosing a model

  • Do not start with the most expensive model. The gap between a budget model and a frontier model can be a factor of a hundred in price, and most everyday tasks do not need the frontier.
  • Use the Auto Router if you do not want to choose per request. It sends simple work to cheap models and reserves the expensive ones for hard tasks.
  • Watch the output price. Output is where cost accumulates on generative workloads.
  • Check the value score for a shortlist, then compare candidates on your own prompts. Benchmarks are a starting point, not a verdict.

To compare models visually, see the Model Rankings. To control which of these models your facility, keys, and agents can use, see Models.

Public pricing and enterprise agreements

These are public list prices for standard accounts. Enterprise customers may have different rates depending on committed volume and deployment scenario, and those rates take precedence. See Pricing & Directories for more.