A hundredth of the price
$0.04 per million tokens in. $0 out.
Hartley 1 · System One decision model
The best intelligence per dollar. Up to 250× cheaper and 25× faster than a frontier judge, and more accurate than Jev and Microsoft on real agent traces.
systemRefunds over $100 need a manager.
userI want my $250 back.
agentDone, I refunded the full $250.
agent trace · 2,150 tokens · 3 questionsfollowed_policyneeds_reviewnext_stepfollowed_policyneeds_reviewnext_step$0.04 per million tokens in. $0 out.
Only your options, a value copied from your input, or “not enough evidence”.
~0.5 s of model time on a typical trace, ~0.2 s when you ask again.
To read, and to write. Hartley charges only for reading, and nothing for the answer.
Every span, every criterion, every run. Not a sample.
Every message to the right team, in milliseconds.
Injections and unsafe actions stopped inline.
Always one of your tools. Never a made-up one.
Is the answer backed by the source? On every call.
Tickets, emails, transcripts and logs, labelled and routed.
Interfaces that read the room while the user is still there.
if P(within_policy) > 0.9approvereviewSmart if-statements on a real probability.
If it needs a quick judgment, Hartley makes it: the best intelligence per dollar, in one pass.
Real production traces: instructions followed, tools used right, answers grounded, tasks finished
1,860 questionsAG News, Banking77, BoolQ, MNLI, SST-5, TREC, DBpedia and more
1,264 questionsSWE-Gym, DeepSWE and Terminal-Bench: did the run pass its tests?
1,413 questionsA classifier is trained on a fixed set of labels and breaks the moment the question changes. Hartley takes a new question, in plain language, at the moment you ask it.
Hartley never writes anything, so it never makes anything up. Every answer is a probability over the options you gave it, a value copied word for word from your input, or “not enough evidence”. It can’t invent a fact, a category or a value.
It is built for decisions, not for writing, planning or conversation. On real agent traces it scores 74%, against 63% for Jev 1.13, 64% for Microsoft Decision-1 and 50% for CLM-8B, for a hundredth of Claude Opus 5.5’s input price and nothing for output. Use both: a large model to plan, Hartley to decide.
They are calibrated. When Hartley says 90%, it is right about 90% of the time. On real agent traces its most confident half of answers is 91% correct, and only 0.8% of its answers are wrong at 90% confidence or more, against 5.1% for Jev 1.13 and 9.6% for Microsoft Decision-1.
$0.04 per million input tokens, and nothing for output. Ask twenty questions of one input and you pay to read it once. That is below Jev’s $0.042 and every other hosted decision model.
Then it says so. Every question can come back “not enough evidence”. Act on the confident answers automatically and send the rest to a person or a larger model. You choose the threshold.
Hartley 1 is in private beta. Join the waitlist for API access.