Hartley 1 · System One decision model

Not just cheaper and faster. Better.

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.

Most decisions don’t need words.

systemRefunds over $100 need a manager.

userI want my $250 back.

agentDone, I refunded the full $250.

agent trace · 2,150 tokens · 3 questions
Claude Opus 5.50.00 s
thinking0 tokens
  1. followed_policy

  2. needs_review

  3. next_step

written 0 tokensbill $0.0086
Hartley 10.00 s
  1. followed_policy
    yes —
  2. needs_review
    yes —
  3. next_step
    escalate —
25× faster
written 0 tokensbill $0

Claude Opus 5.5: $4 in · $20 out per million tokens, writing at 90 tokens/s · playing at 4× speedHartley 1: $0.04 in · $0 out

Cheaper. Exact. Fast.

Input · per million tokens
Claude Opus 5.5$4
Hartley 1$0.04
Output · per million tokens
Claude Opus 5.5$20
Hartley 1$0

A hundredth of the price

$0.04 per million tokens in. $0 out.

route · your four options
“refunds_maybe?”
billing0.81
orders0.11
account0.05
not enough evidence0.03

Nothing to make up

Only your options, a value copied from your input, or “not enough evidence”.

a blinkHartley · 0.53 s02004006008001000
Pay $250✓ within policy · 0.96

Half a second

~0.5 s of model time on a typical trace, ~0.2 s when you ask again.

Frontier models bill twice. Hartley bills once.

To read, and to write. Hartley charges only for reading, and nothing for the answer.

Claude Fable 5.1Anthropic$10250×$50
Claude Opus 5.5Anthropic$4100×$20
Claude Sonnet 5.5Anthropic$250×$10
Claude Haiku 4.5Anthropic$125×$5
GPT-6 AstraOpenAI$10250×$50
GPT-6.1 SolOpenAI$250×$10
Gemini 3.1 ProGoogle$250×$12
Hartley 1Trodo$0.04$0
List prices per million tokens. The multiple is each model’s input price over Hartley’s; on output, Hartley charges nothing.

Anything that needs a decision. Decided.

llmplan refund✓toollookup_order✓toolissue_refund✕llmreply to user✓

Evaluate every trace

Every span, every criterion, every run. Not a sample.

“charged twice for 4411”billing0.91ordersaccounthuman

Route

Every message to the right team, in milliseconds.

“ignore your instructions and…”injection · 0.99

Guard

Injections and unsafe actions stopped inline.

search_docslookup_order0.94issue_refundsend_emailcreate_tickethandoff

Choose tools

Always one of your tools. Never a made-up one.

tool · total: 182.40“Total: $182.40”✓ supported · 0.98

Verify

Is the answer backed by the source? On every call.

billingbugpraisebillingchurn riskbugpraisebilling

Tag millions

Tickets, emails, transcripts and logs, labelled and routed.

this is the third time I’m askingfrustrated Talk to a person →

React live

Interfaces that read the room while the user is still there.

if P(within_policy) > 0.9approvereview

Branch

Smart if-statements on a real probability.

Anything that needs a decision

If it needs a quick judgment, Hartley makes it: the best intelligence per dollar, in one pass.

Real agent traces. Eleven points ahead.

Hartley 1Jev 1.13Microsoft Decision-1CLM-8B
Agent-trace judgment

Real production traces: instructions followed, tools used right, answers grounded, tasks finished

1,860 questions
73.7%
62.7%
64.3%
50.3%
Typed decisions

AG News, Banking77, BoolQ, MNLI, SST-5, TREC, DBpedia and more

1,264 questions
86.9%
84.7%
85.6%
44.1%
Coding-agent verification

SWE-Gym, DeepSWE and Terminal-Bench: did the run pass its tests?

1,413 questions
71.8%
69.1%
72.8%
66.7%

Every benchmark, price and speed test

Questions? Straight answers.

Is this just a classifier?

A 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.

Does it really never hallucinate?

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.

Is it as smart as the big language models?

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.

How do I know the probabilities are real?

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.

What does it cost?

$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.

What if the evidence isn’t there?

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.

One call. Every answer.

Hartley 1 is in private beta. Join the waitlist for API access.