The lineup · 4 tiers

Four tiers.One key.

Switch models by changing one string in the request body. Same host, same key, same stream — a different mind answers.

Claude Opus 5.5

default
claude-opus-5-5

1Mcontext

1,000,000 tokens

Whole-repo work. Long refactors that touch forty files, agentic runs that have to hold the plan in their head from the first tool call to the last.

visionthinkingeffort

Claude Fable 5.1

deepest
claude-fable-5-1

1Mcontext

1,000,000 tokens

Thinks the longest of the four. Reach for it when being wrong is expensive — a migration plan, a race condition nobody can reproduce, a schema you only get to pick once.

visionthinkingeffort

Claude Sonnet 5

everyday
claude-sonnet-5

1Mcontext

1,000,000 tokens

The middle you will actually live in. Reviews, edits, test scaffolding, the glue between two services, the hundred small asks that make up a working day.

visionthinkingeffort

Claude Haiku 4.5

volume
claude-haiku-4-5-20251001

200Kcontext

200,000 tokens

Built for the tight loop. Classification, extraction, routing, labelling — the calls you make ten thousand times and never read one by one.

visionthinkingeffort
the switchboardmodule

The tier is a string in the body.

Everything above the model line is held — same host, same key, same headers, same stream. Change those few characters and a different mind answers on the connection you already had open.

switchboardcycling

Cycling every few seconds. Click any channel to hold.

endpointPOST api.ultracode.in/v1/messagesheld
keysk-uc-••••••••held
modelclaude-opus-5-5changed
context1,000,000 tokens

Whole-repo work. Long refactors that touch forty files, agentic runs that have to hold the plan in their head from the first tool call to the last.

one keyone base URLfour mindsthe Anthropic schema, unmodified
already seatedmodule

Four capabilities you will not have to plumb.

These run on the server side of the connection. There is no flag to set, no second vendor to sign up with, no tool schema to hand-write — and they are on every plan, including the trial.

POST /v1/messagesnothing added
"content": [
{ "type": "text", "text": "what changed in this?" },
{ "type": "image", … }, accepted on the same call
{ "type": "document", … } accepted on the same call
]
no tool schema · no second vendor · no flagon every plan, including the trial

Live web search

The model can look something up mid-turn. No tool definition, no scraper of your own, no second vendor in the loop.

on by default

Image understanding

Drop a screenshot, a whiteboard photo or an architecture diagram into the same content array you already send. It reads it.

on by default

PDF and document reading

Hand over the spec, the RFC or the 200-page vendor contract and ask about page 140. The file goes in the request.

on by default

Extended thinking

Ask for more reasoning on the turns that deserve it. Available on all four tiers, on the same key and the same route.

on by default
pick per requestmodule

Pick per request.

The key does not commit to a tier. Send Opus for the hard problem, Haiku for the next ten thousand calls, Sonnet for everything in between.