TypeLLM

The workflow model.

TypeLLM makes the workflow the unit of inference.

Chat modelDecision modelWorkflow model
PurposeGenerateDecideExecute
UnitMessageDecisionWorkflow
ExampleGPTJevTypeLLM

Switch from chat + decision model to TypeLLM

Safer
No out-of-schema hallucinations.
Simpler
No orchestration code to build or maintain.
Cheaper
The context is read and billed once, not once per step.
Faster
No round trips from multiple API calls, and independent fields run in parallel.
53%cheaper
1.9×faster
Measured on the support ticket workflow.
Support ticket workflowAPI callsInput tokensTime (s)
TypeLLM11491.38
Chat + decision model43202.60
Get started →
The ticket workflow as a graph: category and urgent from the ticket, then a refund only for billing and a reply from both. On the right, the same workflow as the JSON input of one API call.Workflow in one API calldepends_on, when and parallel fields.
A shiba inu looking over a bed, its head in a box.dog_head
BoxesFind something in an image.A grid of 512 squares lighting up: 26, then 255, then all 512 choices.512 choicesUp to 512 classes in one choice.A car show clip with a box on every vehicle in each frame, one car rising into the air.Arrays of boxesEvery match, one box each.The output types rolling past one by one after "type":.More output typesEvery answer comes back typed.
Four choices with their probabilities: billing 0.72, technical 0.18, account 0.07, other 0.03.Better probabilitiesA probability for each choice."thinking": a switch sliding from false to true to auto.Thinking modeTurn on thinking for hard decisions.
ModelType-safe outputAccuracy
Jev86.58%
GPT-5.6 Luna (none)89.18%
GPT-6 Astra (low)100.00%
TypeLLM (no thinking)84.42%
TypeLLM (thinking)98.70%
BenchmarkJevBench, 231 public tasks.