everything quiet · 99.97% 30d
Issue 04 · April 2026/23 models on hand

One endpoint,twenty-three minds,quietly chosen.

Prism is a soft-spoken router for production traffic. It picks the right model for each request — by length, intent, and quality — and gets out of the way. No re-prompting. No vendor lock-in. No drama.

8.3
ms route
23
models
47%
avg saved
live · last 12 requests
23 models on hand
summarise 4-page brief
haiku-4
412ms
$0.0009
compose contributor letter
opus-4.7
1840ms
$0.038
extract names from PDF
llama-4-8b
218ms
$0.0001
translate to Portuguese
sonnet-4.6
614ms
$0.012
classify intent
haiku-4
96ms
$0.0002
8.3ms
route p50
$0.0094
avg cost
−47%
vs single
Plover StudioLinnea BakeryVerge EditionsField Notes Co.Marlow WorkwearHelvetia CartographicNorthwind LabsMercury OSSubstrateFoundry 47Plover StudioLinnea BakeryVerge EditionsField Notes Co.Marlow WorkwearHelvetia CartographicNorthwind LabsMercury OSSubstrateFoundry 47
II — Apparatus

Four quiet rooms in one studio.

Routing studio
length > 1k tokens→ opus-4.7
intent = "code"→ sonnet-4.6
intent = "classify"→ haiku-4
region = "EU"→ self-hosted
fallback→ haiku-4

A visual editor for the policy. Click a model, see which requests it would have served. Adjust without redeploying.

Replay bench
cost ↓qual ↑
Cache
62%
semantic hit rate · 24h
avg saved · $0.0084 / req
Catalogue
opus-4.7
anthropic
considered
sonnet-4.6
anthropic
balanced
haiku-4
anthropic
fast
gpt-5
openai
considered
gpt-5-mini
openai
balanced
llama-4-70b
self
local
mistral-large
mistral
balanced
command-r+
cohere
balanced
Endpoint
single endpoint
https://api.prism.dev/v1/chat
accepts OpenAI, Anthropic, and Bedrock request shapes — answer comes back in your dialect.
III — Practice

Three rooms, one endpoint.

Editorial, product, and research teams all reach the same Prism URL. The decision happens behind the curtain.

i.
Editorial

A long letter routed through a careful model. A grocery list through a quick one.

Prism reads the request, not the prompt. Long-form writing reaches the considered model. A two-line classification reaches the fast one. The cost difference is yours to keep.

0
% saved versus single-model defaults
ii.
Product

One endpoint. Twenty-three models. Your code never changes.

Swap providers without touching a line of application code. Prism speaks OpenAI, Anthropic, and Bedrock natively. Your migration is a config file.

0
ms median routing decision
iii.
Research

A quiet bench for evaluating models on your own work.

Replay a week of traffic across alternative models. Compare quality, cost, and latency on requests that actually mattered. Decide with evidence, not the launch post.

0
% of replays finished in under a minute
IV — Comparison

The single-model habit, gently broken.

Capability
Prism
Single-model setup
Model selection
Per request, learned
Hardcoded constant
Provider migration
Zero-code, config flag
Sprint-long rewrite
Quality measurement
Built-in eval bench
Spreadsheet of vibes
Streaming compat.
Provider-agnostic
Per-vendor branches
Caching
Semantic, content-aware
URL hash
a quiet word

Prism replaced four if-statements and a feature flag with one endpoint. Quality went up and so did our weekends.

Maja Nilsen
Engineering, Plover Studio
0M
requests routed daily
0%
median cost saved
0ms
route decision
0
models on hand
V — Subscriptions

Plain plans, no fine print.

Quiet
$0free for hobby

For weekend projects and personal sites.

  • 12k requests / mo
  • All public models
  • Single endpoint
  • Community notes
Begin quietly
Studio
$64per seat / month

For independent shops and small product teams.

  • 1.4M requests / mo
  • Routing studio
  • Replay bench
  • Semantic cache
  • Three environments
  • Priority email
Take a 14-day look
Atelier
annual handshake

For larger teams and custom routing.

  • Unlimited requests
  • Private model pool
  • BYOK + region pinning
  • SCIM + audit log
  • Named partner
  • 99.95% SLA
Have a conversation
VI — Marginalia

A few honest answers.

A small classifier trained on your traffic considers length, intent, and tags. It then picks the cheapest model that historically meets your quality threshold for that class of request. You can override per-route.

From the studio

all notes
Apr 26
router
New 'editorial' classifier favours considered models for prose over 240 tokens.
Apr 18
studio
Routing studio now visualises p95 latency by model in a single chart.
Apr 09
models
Added six small models from Mistral, Cohere, and a self-hosted Llama 4 endpoint.
Mar 30
cache
Semantic cache lookups dropped to 1.4 ms median.

Begin quietly.

One endpoint. Twelve thousand requests on the house. No card, no committee.