One key.Every model.

Prism gives engineering teams one OpenAI-compatible endpoint in front of every model, and a log that says why each request went where it went.

Get an API key See every model

api.prism.example

Acme AI production Requests

All policies 1,284 today

Time	Model	Task	Status	Latency	Cost
09:41:02	google/gemini-3.1-flash	summarize	Routed	340 ms	$0.0021
09:41:02	google/gemini-3.1-flash	summarize	Routed	318 ms	$0.0019
09:41:01	deepseek/deepseek-v4-pro	summarize	Routed	602 ms	$0.0074
09:41:00	google/gemini-3.1-flash	summarize	Routed	355 ms	$0.0022
09:40:59	anthropic/claude-sonnet-5	summarize	Pinned	540 ms	$0.0186
09:40:58	google/gemini-3.1-flash	summarize	Routed	331 ms	$0.0020

Why is this job on Flash and not Sonnet?

prism policy show production 2s

prism score --task summarize --window 30d 1s

Your policy asks for the cheapest model above quality 80. Flash scores 84 on your own summaries and costs $0.35/M against Sonnet's $3.00/M, so it takes the job. The five that went to Sonnet were pinned by the caller.

Acme AI production Usage

This month $4,322 of $6,000

Time	Model	Task	Status	Latency	Cost
09:38:11	anthropic/claude-fable-5-1	plan	Over budget	1,180 ms	$0.4210
09:38:04	openai/gpt-5.6	plan	Routed	742 ms	$0.1180
09:37:52	anthropic/claude-sonnet-5	draft	Routed	533 ms	$0.0402
09:37:41	qwen/qwen3.8-max	draft	Routed	705 ms	$0.0176
09:37:30	google/gemini-3.1-flash	classify	Routed	328 ms	$0.0009
09:37:12	meta/llama-4-maverick	classify	Routed	214 ms	$0.0006

Why did this request cost $0.42?

prism usage --request req_8fa21c 1s

prism budget show key_prod 1s

It sent 38K tokens of context to Fable 5.1 at $10/M in and $50/M out. The key is at 72% of its $6,000 monthly cap. A rule that summarizes context above 20K tokens first would have cost $0.06.

Acme AI production Fallbacks

Last 15 minutes 3 fallbacks

Time	Model	Task	Status	Latency	Cost
09:41:07	openai/gpt-5.6	chat	Timed out	8,000 ms	—
09:41:07	anthropic/claude-sonnet-5	chat	Fallback	561 ms	$0.0140
09:41:03	openai/gpt-5.6	chat	529	210 ms	—
09:41:03	anthropic/claude-sonnet-5	chat	Fallback	548 ms	$0.0139
09:40:58	openai/gpt-5.6	chat	Routed	690 ms	$0.0121
09:40:54	openai/gpt-5.6	chat	Routed	668 ms	$0.0118

Is OpenAI degrading, or is it us?

prism health openai --region us-east 2s

prism fallbacks --since 15m 1s

First-token p95 on OpenAI us-east went from 700 ms to 8 s at 09:41. Your policy skipped the region after two failures and Sonnet took the next 214 requests. Nothing reached a user as an error.

Acme AI production Requests

Over 400K tokens 18 requests

Time	Model	Task	Status	Latency	Cost
09:35:12	google/gemini-3.1-pro	extract	Routed	4,120 ms	$2.3100
09:31:44	google/gemini-3.1-pro	extract	Routed	3,980 ms	$2.2400
09:28:03	anthropic/claude-fable-5-1	extract	Too long	—	—
09:27:58	google/gemini-3.1-pro	extract	Fallback	4,301 ms	$2.4000
09:19:22	google/gemini-3.1-pro	extract	Routed	3,870 ms	$2.1900
09:11:07	google/gemini-3.1-pro	extract	Routed	4,044 ms	$2.2700

Which models can take a 900-page contract?

prism models --min-context 1M 1s

prism policy explain long-documents 1s

Two: Gemini 3.1 Pro at 2M tokens and Fable 5.1 at 1M. The contract is 1.4M tokens, so Pro is the only one it fits in, and the rule sends it there before anything else is weighed.

Acme AI production Requests

Pinned models 312 requests

Time	Model	Task	Status	Latency	Cost
09:40:02	anthropic/claude-sonnet-5	code	Pinned	528 ms	$0.0310
09:39:58	anthropic/claude-sonnet-5	code	Pinned	544 ms	$0.0298
09:39:51	anthropic/claude-sonnet-5	code	Pinned	502 ms	$0.0281
09:39:44	deepseek/deepseek-v4-pro	code	Routed	618 ms	$0.0102
09:39:31	anthropic/claude-sonnet-5	code	Pinned	559 ms	$0.0305
09:39:18	deepseek/deepseek-v4-pro	code	Routed	604 ms	$0.0099

What would routing have picked instead?

prism simulate --pinned-off --since 1h 2s

prism diff --quality --cost 1s

DeepSeek V4 Pro on 246 of the 312, at 86 against Sonnet's 92 on your code traffic. That is $58 a day cheaper and six points worse. Fallbacks stay on either way.

Acme AI production req_8fa21c

Single request 4 decisions

Time	Model	Task	Status	Latency	Cost
09:38:11.204	router	policy match	production	4 ms	—
09:38:11.208	router	model chosen	best_quality	31 ms	—
09:38:11.239	anthropic	request sent	us-east	1,180 ms	$0.4210
09:38:12.419	anthropic	response	1,932 tokens	—	—
09:38:12.421	router	logged	req_8fa21c	2 ms	—

Show me every decision for this request.

prism trace req_8fa21c 1s

prism policy explain --request req_8fa21c 1s

Four lines, 38 ms of routing overhead, one provider call. The policy matched on task=plan, best quality under a $5.00 input cap chose Fable 5.1, and nothing fell back.

Eleven providers. One invoice, at their list price.

OpenAI
Anthropic
Google
Meta
Mistral
DeepSeek
xAI
Alibaba

Route by policy, not by hand.

A policy is a short file of rules: what to prefer, when to fall back, where the budget stops. Commit it beside your code and every request obeys it.

Cheapest above a quality floor, fastest within a margin, best under a price cap — or a pinned model.
Fallbacks fire inside the same call, even mid-stream, and the log records both legs.

Read the policy reference

production.policy.yml

live

when task is summarize

then prefer cheapest above quality 75
when p95 latency over 800 ms

then fall back to the next provider
when context over 400K tokens

then route to gemini-3.1-pro
when key spend over $200 today

then downgrade to fast tier and alert

Pushed 4 minutes ago · 2,318 requests routed since

Spend by model

September, to date

$4,322

−31% vs. August at the same point

$1,842

Sonnet 5
$1,210

GPT-5.6
$412

Gemini Flash
$388

DeepSeek V4
$296

Llama 4
$174

Others

See every token you spend.

Spend, tokens and latency by model, key and policy — live, and exportable as a row per request. The bill stops being a surprise.

Per-key budgets with alerts at 50, 80 and 100 percent, and a soft-stop that downgrades instead of failing.
Quality scores from a judge you pick, on your own traffic, so the router learns what good looks like for you.

Usage and limits

Every model behind one endpoint.

Prism is OpenAI-compatible. Change the base URL and the model id, keep everything else, and the whole catalog is one request away.

Read the quickstart

route.py

from openai import OpenAI

client = OpenAI(
    base_url="https://api.prism.example/v1",  # was api.openai.com
    api_key="pk_live_…",
)

reply = client.chat.completions.create(
    model="prism/auto",  # the policy picks the model
    messages=[{"role": "user", "content": prompt}],
)

print(reply.headers["x-prism-route"])
# → google/gemini-3.1-flash · cheapest_above(80) · $0.0021

140+
Models behind one endpoint: 11
Providers, one invoice: 99.99%
Routed uptime, trailing 90 days: 38 ms
Median routing overhead

Built for the workloads that actually run.

Agents: Give an agent one key and let each step pick its own model — planning on a frontier model, tool calls on a fast one.; model: "prism/auto"
Batch pipelines: Route a million overnight classifications to whatever is cheapest above your quality floor, with the bill capped in advance.; policy: "cheapest-above-80"
Consumer chat: Keep first-token time under 300 ms at peak by spilling over to the next provider before anyone notices.; fallback: ["meta", "google"]

We cut our model bill by a third in the first week and nobody on the team noticed a quality change. The log is what sold me: I can see why every request went where it went.

Priya Anand

Head of Platform, Loamworks

From the blog

Benchmarks, routing notes and what we learned running traffic for other teams.

All posts

Point your client at Prism. Keep everything else.

Free for a million routed tokens a month. Provider prices at list, on one invoice.

Get an API key See pricing