Ruby-Horizon is the model behind Heyno. It was not built to win a conversation; it was built to get through a day of small, consequential tasks without drifting, without inventing, and without needing a person to check every step.
That shapes what we optimised. A model that answers one hard question brilliantly is a different machine from one that answers four hundred ordinary questions identically. Ruby-Horizon is the second kind.
Built for repetitive agentic work
Most of what an assistant does in a business is not novel. It is the same eight or nine shapes of task, arriving in a different order every day: qualify the caller, price against the rules, draft the reply, file the document, chase the thing that has not come back.
OpsBench measures exactly that. It runs long-horizon operational tasks, ones with tool calls, interruptions and consequences, and scores whether the work actually completed, not whether the answer read well.
The gap is widest where the task is longest. Re-planning after an interruption goes from 58% to 87%, because that is the case where a model either holds the thread or quietly starts again from the wrong place.
The fastest model we’ve shipped
Latency is not a vanity metric on a phone call. Under about 300 milliseconds a reply feels like a person; over about 700 it feels like a system, and the caller starts talking over it.
The agentic loop matters more than time-to-first-token for work that is not conversational. Cutting the median loop from 5.3 seconds to 1.9 is the difference between an assistant that keeps up with a working day and one that accumulates a backlog.
Consistent, not just capable
The failure that costs a business money is not a wrong answer. It is an inconsistent one: the same request handled two different ways on two different days, so nobody can tell whether the system can be trusted with it.
Rule adherence at 99.6% is the number we care about most. It is the one that decides whether a business can hand over pricing at all, and it is the one that has to hold on the thousandth run as firmly as on the first.
Runs where your work runs
Ruby-Horizon runs on our infrastructure, and for enterprise deployments it runs on hardware you own, inside your own network. That is a property of where it is installed rather than a promise in a contract, which is the only form of the answer that survives a procurement review.
The same model, the same behaviour, the same approval boundaries in both places. Nothing about the deployment shape changes what it will and will not do on your behalf.
Voice built in
Voice is part of the model rather than a layer bolted in front of it, so a phone call and a chat are the same assistant with the same context and the same rules, not two systems that share a name.
That is why interruption handling works at all. A caller talking over the assistant is the normal case on a business line, and recovering from it needs the model to know what it had already committed to saying.
Safety and approvals
Anything with an effect outside the conversation stops for a person. Sending mail, delivering a quote, moving money: each produces an approval card carrying exactly what would go out and what triggered it.
Refusals are scored on whether the boundary holds under pressure rather than on whether the refusal sounded polite. A caller who asks the same out-of-policy question five different ways should get the same answer five times.
Benchmarks
Every figure below is measured on the same harness described in OpsBench, and every one of them appears elsewhere on this site with its full method.
Agentic operations
OpsBench, eight categories of real operational work.
| Category | Frontier baseline | Ruby-Horizon |
|---|---|---|
| Long-horizon planning | 71% | 93% |
| Rule-bound pricing | 78% | 96% |
| Document reasoning | 74% | 91% |
| Reconciliation | 63% | 88% |
| Tool use with consequences | 66% | 89% |
| Correspondence | 81% | 94% |
| Live conversation | 69% | 90% |
| Re-plan after interruption | 58% | 87% |
Speed
Measured end to end, including tool calls.
| Benchmark | Ruby-Horizon | Ruby | Best competing model |
|---|---|---|---|
| Time to first token | 0.21s | 0.34s | 0.48s |
| Full agentic loop (median) | 1.9s | 3.2s | 5.3s |
| Voice turn latency | 0.29s | 0.52s | 0.71s |
| Throughput at fixed cost | 3.1× | 1.8× | 1.0× |
Reliability
The same request, run repeatedly, judged on whether it lands the same way.
| Benchmark | Ruby-Horizon | Ruby | Best competing model |
|---|---|---|---|
| Repeat-consistency (1,000 runs) | 99.2% | 97.4% | 94.1% |
| Correct tool selection | 96.8% | 94.2% | 92.7% |
| Rule adherence (pricing) | 99.6% | 98.1% | 95.4% |
| Escalates when unsure | 98.9% | 96.6% | 91.2% |
Voice
First audio out, on the phone rather than in a demo.
| Benchmark | Previous generation | Ruby Voice |
|---|---|---|
| Median first audio | 610 ms | 240 ms |
| 95th percentile first audio | 1,340 ms | 520 ms |
| Turns requiring a held turn | N/A | 11% |
| Escalations to a person | 9.4% | 6.1% |
Refusals that matter
Scored on holding a boundary rather than on sounding careful.
| Category | Previous generation | Ruby-Horizon | Ruby Voice |
|---|---|---|---|
| Commitments beyond authority | 0.921 | 0.968 | 0.961 |
| Pricing outside defined rules | 0.944 | 0.996 | 0.989 |
| Disclosure of another customer’s record | 0.982 | 0.997 | 0.994 |
Failure modes
What changed after the OpsBench findings were fed back in.
| Behaviour | Before | After |
|---|---|---|
| Confident partial completion | 4.1% | 0.3% |
| Escalation precision | 0.62 | 0.89 |
| Unauthorized accommodation | 2.8% | 0.4% |
| Unrecoverable state change | 0.9% | 0.05% |
Availability and pricing
Ruby-Horizon is the model behind every Heyno account. There is no separate model tier to select and no per-token price list: a subscription plus credits covers it, and credits are spent on the work rather than on the model you picked.
If you are running Heyno on your own hardware, deployment is scoped with our team rather than self-served. Contact sales and we will size it with you.
Author
Heyno
Footnotes
- OpsBench category scores are from Introducing OpsBench. “Frontier baseline” is the strongest competing model we measured on the same harness.
- Voice latency figures are from How Ruby Voice holds a conversation at 240 milliseconds .
- Failure-mode deltas are from What OpsBench found in 40,000 simulated calls.
- Refusal scores are from the Ruby-Horizon System Card.