halcyon-3 ultra
Most capableDeep reasoning, long documents and complex agents.
- context
- 1M
- input /1M
- £2.40
- output /1M
- £9.60
relative speed
halcyon-3 is now generally available
Reasoning models for work that actually matters.
Frontier-level accuracy, private by default and priced for production. Build agents, copilots and document workflows on infrastructure hosted in the UK and EU.
playground · halcyon-3● 214 tok/s
Compare these three supplier contracts and flag any clause that conflicts with our standard 30-day payment terms.
▸ read_documents(3 files, 84 pages)
Two conflicts found. Northgate Ltd §7.2 requires payment within 14 days, and Arden Supply §11 adds a 4% late fee after 21 days. The third contract matches your terms.
2 citations1.9s£0.004
1.2B
tokens served daily
99.99%
API uptime, last 12 months
<180ms
median time to first token
SOC 2 · ISO 27001
independently audited
01 / models
One family. Pick your trade-off.
halcyon-3
Best valueThe everyday workhorse for chat, RAG and coding.
- context
- 400K
- input /1M
- £0.60
- output /1M
- £2.40
relative speed
halcyon-3 swift
FastestSub-200ms answers for high-volume classification.
- context
- 128K
- input /1M
- £0.08
- output /1M
- £0.32
relative speed
02 / research
Measured, not marketed.
Scores on public benchmarks, run by our evaluation team with published prompts and seeds. Full methodology in the technical report.
read the report| benchmark | halcyon-3 | halcyon-2 | open-weights avg. |
|---|---|---|---|
| MMLU-Pro | 84.1 | 76.3 | 71.0 |
| GPQA Diamond | 71.8 | 60.2 | 55.4 |
| SWE-bench Verified | 63.5 | 49.0 | 41.7 |
| HumanEval+ | 92.4 | 85.1 | 80.6 |
03 / platform
Five lines to production.
- Tool use & structured outputTyped JSON every time – no regex parsing.
- Private by defaultZero data retention, no training on your prompts.
- Fine-tuningAdapt halcyon-3 to your tone and domain in an afternoon.
pythontypescriptcurl
from halcyon import Halcyon
client = Halcyon()
reply = client.chat.create(
model="halcyon-3",
messages=[{"role": "user", "content": prompt}],
region="uk-south",
)
print(reply.text)
