Skip to main content
PRYSM is a universal AI routing engine. Point your existing OpenAI code at PRYSM and every prompt is automatically routed to the best-value model for the job — across 21 models from 9 providers (OpenAI, Anthropic, Google, DeepSeek, xAI, Mistral, Moonshot, Alibaba, Perplexity) — behind one API, one key, and one bill.

Drop-in for OpenAI

Change one line — the base URL. Every OpenAI SDK call works unchanged.

Smart routing

Intent classification picks the right model for each prompt: quality, balanced, or agility.

BRAIN.md config

A declarative, version-controlled routing file — the .cursorrules of model routing.

Cost guardrails

AgentGuard caps per-request spend and blocks models you don’t want agents touching.

PrysmProof receipts

Every response carries a tamper-evident SHA-256 receipt of what ran and why.

MCP server

Give Claude Desktop, Cursor, or Windsurf cost-aware routing as a native tool.

​
Why PRYSM

The model you reach for by default is rarely the best one for the task in front of you. A frontier model on a one-line classification is wasted money; a budget model on a nuanced contract is wasted quality. PRYSM makes that decision per request — so you get the right model every time without hand-tuning, while spend stays predictable.

Better outputs

Each prompt goes to the model that’s actually best at it — code, writing, math, translation, reasoning, and more.

Lower cost

Cheap models handle the easy 80%; premium models are reserved for the hard 20%. Typical savings are large versus an all-premium baseline.

Zero lock-in

OpenAI-compatible. No new SDK to learn, no rewrite, no juggling nine provider keys.

​
The drop-in pattern

PRYSM speaks the OpenAI API. Set model: "auto" and let PRYSM choose: 
from openai import OpenAI

client = OpenAI(
    api_key="prysm_sk_your_key",
    base_url="https://api.prysm1.com/v1",   # the only change
)

resp = client.chat.completions.create(
    model="auto",                          # let PRYSM pick the best model
    messages=[{"role": "user", "content": "Write a Python quicksort"}],
)
print(resp.choices[0].message.content)
Prefer a native client? The first-party SDKs (prysm1 for Python and @prysmai/sdk for Node) subclass the OpenAI client and add helpers for routing previews, usage, savings, and BRAIN.md auto-discovery.

​
What’s in every response

PRYSM returns a standard OpenAI payload plus a top-level prysm block describing the decision — which model ran, why, what it cost, latency, and a verifiable proof: 
{
  "id": "prysm-a1b2c3d4",
  "object": "chat.completion",
  "model": "deepseek-v3.2",
  "choices": [{ "message": { "role": "assistant", "content": "..." } }],
  "usage": { "prompt_tokens": 18, "completion_tokens": 240, "total_tokens": 258 },
  "prysm": {
    "routing": {
      "mode": "balanced",
      "model_display": "DeepSeek V3.2",
      "provider": "deepseek",
      "reason": "Code: 95% cheaper than GPT-5.2",
      "tier": "budget",
      "signals_detected": { "code": true }
    },
    "cost": { "input_usd": 0.0000050, "output_usd": 0.0001008, "total_usd": 0.0001058 },
    "latency_ms": 740,
    "proof": { "proof_hash": "sha256:a1b2c3d4e5f6a7b8", "verifiable": true }
  }
}
OpenAI-compatible clients ignore the extra prysm field. The PRYSM SDKs expose it cleanly through extension().

​
Next steps

Quickstart

Make your first routed request in under five minutes.

Authentication

API keys, the base URL, and environment variables.

How routing works

Modes, signals, and how PRYSM picks a model.

API reference

Every endpoint, parameter, and response field.