Build safe & production‑ready clinical AI in your product
Build safe & production‑ready clinical AI in your product
Sully APIs bring the same intelligence that powers the Sully platform directly into your clinical application
Sully APIs bring the same intelligence that powers the Sully platform directly into your clinical application
Sully APIs bring the same intelligence that powers the Sully platform directly into your clinical application
Transcription
Streaming
Notes
Text to JSON
Sully M1
TS
Python
// Use your template and patient context.
const { note } = await client.post(ʺ/v3/notesʺ, {
body: { transcript, template, clinicalContext },
});
Response
9.4 s median
200 OK
Request ID
req_7f3a9c21
Note
# Subjective
Right-sided headaches, 3 this week, with nausea and light sensitivity.
# Assessment
Migraine, not intractable.
# Plan
Follow up in 2 weeks.
Usage
1,450 input · 320 output tokens
Transcription
Streaming
Notes
Text to JSON
Sully M1
TS
Python
// Use your template and patient context.
const { note } = await client.post(ʺ/v3/notesʺ, {
body: { transcript, template, clinicalContext },
});
Response
9.4 s median
200 OK
Request ID
req_7f3a9c21
Note
# Subjective
Right-sided headaches, 3 this week, with nausea and light sensitivity.
# Assessment
Migraine, not intractable.
# Plan
Follow up in 2 weeks.
Usage
1,450 input · 320 output tokens
Health tech teams build on Sully
Tebra’s AI Scribe
built on Sully
Tebra’s AI Scribe built on Sully
100k+
notes per month on Sully APIs
40+
specialties
99.96%
uptime
11.23s
p50 latency
100k+
notes per month on Sully APIs
40+
specialties
99.96%
uptime
11.23s
p50 latency
APIs for every step of the visit
APIs for every step of the visit
APIs for every step of the visit
Transcription
Upload a recorded visit up to 24 hours long, in 30+ languages. Poll for the transcript, or get a webhook when it is ready.
Streaming transcription
Stream audio during the visit and get partial and final transcripts back live. Boost the words that matter, such as drug names.
Note generation
Send a transcript and your own template. Sully returns the finished note.
Text to JSON
Send a note or transcript with your JSON Schema. You get validated fields back, ready for your EHR.
Sully M1
Sully M1 is our general clinical LLM, called like any chat model. Send a prompt with the clinical context and it returns an answer, so you can build features the task APIs don’t cover.
52.61%
HealthBench red teaming score
92.2%
Clinical note accuracy
9.4 s
Median per note
$2.28
Per 1M tokens
OpenAI-compatible request
{
"model": "sully-m1",
"messages": [
{ "role": "system",
"content": "Be concise and use plain language."},
{ "role": "user",
"content": "What does an A1C test measure?" }
]
}
// answer: An A1C test estimates average blood glucose over the past two to three months.
Copied to clipboard
Transcription
Upload a recorded visit up to 24 hours long, in 30+ languages. Poll for the transcript, or get a webhook when it is ready.
Streaming transcription
Stream audio during the visit and get partial and final transcripts back live. Boost the words that matter, such as drug names.
Note generation
Send a transcript and your own template. Sully returns the finished note.
Text to JSON
Send a note or transcript with your JSON Schema. You get validated fields back, ready for your EHR.
Sully M1
Sully M1 is our general clinical LLM, called like any chat model. Send a prompt with the clinical context and it returns an answer, so you can build features the task APIs don’t cover.
52.61%
HealthBench red teaming score
92.2%
Clinical note accuracy
9.4 s
Median per note
$2.28
Per 1M tokens
OpenAI-compatible request
{
"model": "sully-m1",
"messages": [
{ "role": "system",
"content": "Be concise and use plain language."},
{ "role": "user",
“content”: “When should she come back?” }
]
}
// answer: An A1C test estimates average blood glucose over the past two to three months.
Copied to clipboard
Transcription
Upload a recorded visit up to 24 hours long, in 30+ languages. Poll for the transcript, or get a webhook when it is ready.
Streaming transcription
Stream audio during the visit and get partial and final transcripts back live. Boost the words that matter, such as drug names.
Note generation
Send a transcript and your own template. Sully returns the finished note.
Text to JSON
Send a note or transcript with your JSON Schema. You get validated fields back, ready for your EHR.
Sully M1
Sully M1 is our general clinical LLM, called like any chat model. Send a prompt with the clinical context and it returns an answer, so you can build features the task APIs don’t cover.
52.61%
HealthBench red teaming score
92.2%
Clinical note accuracy
9.4 s
Median per note
$2.28
Per 1M tokens
OpenAI-compatible request
{
"model": "sully-m1",
"messages": [
{ "role": "system",
"content": "Be concise and use plain language."},
{ "role": "user",
"content": "What does an A1C test measure?" }
]
}
// answer: An A1C test estimates average blood glucose over the past two to three months.
Copied to clipboard
Sully APIs are fast, accurate, and economical
Sully APIs are fast, accurate, and economical
Sully APIs are fast, accurate, and economical
We evaluate models on production-derived clinical tasks to measure accuracy, latency, and cost.
Highest safety score per dollar in our test.
Safety score per dollar
Sully M1
23.13
GPT-6.1 Sol
15.73
GLM 5.3
7.68
Claude Opus 5.5
7.34
Gemini 4 Argon
5.37
Safety score ÷ blended price (3:1 input/output) · higher is better
92.16% accuracy at 9.4 seconds p50 latency.
Accuracy vs. latency
Input and output prices
Sully M1
GPT-6.1 Sol
Claude Sonnet 5.5
Kimi K3Claude Opus 5.5Claude Fable 5.1GPT-6 Astra
Sully’s inference engine delivers speed at scale
Our inference optimizations enable us to support higher throughput: larger prompts and more parallel requests all with lower latency.
Disaggregated inference
Disaggregated prefill and decode workers are tuned for clinical workflow request sizes.
Fast KV cache transfer
Prompt prefixes are stored in a KV cache and are transferred from prefill to decode workers using a custom-tuned Mooncake implementation over NVlink.
Speculative decoding
Fine-tuned speculative decoding model that improved acceptance from 2 to 8 tokens.
We evaluate models on production-derived clinical tasks to measure accuracy, latency, and cost.
Highest safety score per dollar in our test.
Safety score per dollar
Sully M1
23.13
GPT-6.1 Sol
15.73
GLM 5.3
7.68
Claude Opus 5.5
7.34
Gemini 4 Argon
5.37
Safety score ÷ blended price (3:1 input/output) · higher is better
92.16% accuracy at 9.4 seconds p50 latency.
Accuracy vs. latency
Input and output prices
Sully M1GPT-6.1 SolClaude Sonnet 5.5Kimi K3Claude Opus 5.5Claude Fable 5.1GPT-6 AstraSully’s inference engine delivers speed at scale
Our inference optimizations enable us to support higher throughput: larger prompts and more parallel requests all with lower latency.
Disaggregated inference
Disaggregated prefill and decode workers are tuned for clinical workflow request sizes.
Fast KV cache transfer
Prompt prefixes are stored in a KV cache and are transferred from prefill to decode workers using a custom-tuned Mooncake implementation over NVlink.
Speculative decoding
Fine-tuned speculative decoding model that improved acceptance from 2 to 8 tokens.
We evaluate models on production-derived clinical tasks to measure accuracy, latency, and cost.
Highest safety score per dollar in our test.
Safety score per dollar
Sully M1
23.13
GPT-6.1 Sol
15.73
GLM 5.3
7.68
Claude Opus 5.5
7.34
Gemini 4 Argon
5.37
Safety score ÷ blended price (3:1 input/output) · higher is better
92.16% accuracy at 9.4 seconds p50 latency.
Accuracy vs. latency
Input and output prices
Sully M1GPT-6.1 SolClaude Sonnet 5.5Kimi K3Claude Opus 5.5Claude Fable 5.1GPT-6 AstraSully’s inference engine delivers speed at scale
Our inference optimizations enable us to support higher throughput: larger prompts and more parallel requests all with lower latency.
Disaggregated inference
Disaggregated prefill and decode workers are tuned for clinical workflow request sizes.
Fast KV cache transfer
Prompt prefixes are stored in a KV cache and are transferred from prefill to decode workers using a custom-tuned Mooncake implementation over NVlink.
Speculative decoding
Fine-tuned speculative decoding model that improved acceptance from 2 to 8 tokens.
We evaluate models on production-derived clinical tasks to measure accuracy, latency, and cost.
Highest safety score per dollar in our test.
Safety score per dollar
Sully M1
23.13
GPT-6.1 Sol
15.73
GLM 5.3
7.68
Claude Opus 5.5
7.34
Gemini 4 Argon
5.37
Safety score ÷ blended price (3:1 input/output) · higher is better
92.16% accuracy at 9.4 seconds p50 latency.
Accuracy vs. latency
Input and output prices
Sully M1GPT-6.1 SolClaude Sonnet 5.5Kimi K3Claude Opus 5.5Claude Fable 5.1GPT-6 AstraSully’s inference engine delivers speed at scale
Our inference optimizations enable us to support higher throughput: larger prompts and more parallel requests all with lower latency.
Disaggregated inference
Disaggregated prefill and decode workers are tuned for clinical workflow request sizes.
Fast KV cache transfer
Prompt prefixes are stored in a KV cache and are transferred from prefill to decode workers using a custom-tuned Mooncake implementation over NVlink.
Speculative decoding
Fine-tuned speculative decoding model that improved acceptance from 2 to 8 tokens.
We evaluate models on production-derived clinical tasks to measure accuracy, latency, and cost.
Highest safety score per dollar in our test.
Safety score per dollar
Sully M1
23.13
GPT-6.1 Sol
15.73
GLM 5.3
7.68
Claude Opus 5.5
7.34
Gemini 4 Argon
5.37
Safety score ÷ blended price (3:1 input/output) · higher is better
92.16% accuracy at 9.4 seconds p50 latency.
Accuracy vs. latency
Input and output prices
Sully M1GPT-6.1 SolClaude Sonnet 5.5Kimi K3Claude Opus 5.5Claude Fable 5.1GPT-6 AstraSully’s inference engine delivers speed at scale
Our inference optimizations enable us to support higher throughput: larger prompts and more parallel requests all with lower latency.
Disaggregated inference
Disaggregated prefill and decode workers are tuned for clinical workflow request sizes.
Fast KV cache transfer
Prompt prefixes are stored in a KV cache and are transferred from prefill to decode workers using a custom-tuned Mooncake implementation over NVlink.
Speculative decoding
Fine-tuned speculative decoding model that improved acceptance from 2 to 8 tokens.
We evaluate models on production-derived clinical tasks to measure accuracy, latency, and cost.
Highest safety score per dollar in our test.
Safety score per dollar
Sully M1
23.13
GPT-6.1 Sol
15.73
GLM 5.3
7.68
Claude Opus 5.5
7.34
Gemini 4 Argon
5.37
Safety score ÷ blended price (3:1 input/output) · higher is better
92.16% accuracy at 9.4 seconds p50 latency.
Accuracy vs. latency
Input and output prices
Sully M1GPT-6.1 SolClaude Sonnet 5.5Kimi K3Claude Opus 5.5Claude Fable 5.1GPT-6 AstraSully’s inference engine delivers speed at scale
Our inference optimizations enable us to support higher throughput: larger prompts and more parallel requests all with lower latency.
Disaggregated inference
Disaggregated prefill and decode workers are tuned for clinical workflow request sizes.
Fast KV cache transfer
Prompt prefixes are stored in a KV cache and are transferred from prefill to decode workers using a custom-tuned Mooncake implementation over NVlink.
Speculative decoding
Fine-tuned speculative decoding model that improved acceptance from 2 to 8 tokens.
Sully APIs are fast, accurate, and economical
Pay for what you use
Pay for what you use
Pay for what you use
Sully M1 model API pricing, per 1M tokens.
Sully M1 model API pricing, per 1M tokens.
Sully M1 model API
Sully M1 model API
Per 1M tokens.
$1.50
$1.50
input
$4.60
$4.60
output
What the same tokens cost elsewhere
Lower is better
Your data is private
Your data is private
We’re not just leading in compliance and security, we also make sure all your data is private to you. We don’t use or train on your data.
We’re not just leading in compliance and security, we also make sure all your data is private to you. We don’t use or train on your data.
