The Murmur 3.5 model family.
Three models, each trained to speak in your participants' own inner register. Murmur 3.5 sets a new state of the art on ConvictionBench and is available today in Second Voice, Familiars and the API.
Hum 3.5
Fast and light. For high-volume nudges, reminders and small corrections throughout the day.
- Context
- 200K tokens
- Input
- $0.40 / MTok
- Output
- $1.20 / MTok
- Conviction latency
- 60 ms
hum-3-5Murmur 3.5
Our balanced model and the default voice in Second Voice and Familiars. Natural, warm and persuasive.
- Context
- 1M tokens
- Input
- $2 / MTok
- Output
- $6 / MTok
- Conviction latency
- 140 ms
murmur-3-5Chorus 3.5
Our most capable model, for executive mandates, difficult objections and long-horizon alignment.
- Context
- 1M tokens
- Input
- $10 / MTok
- Output
- $30 / MTok
- Conviction latency
- 310 ms
chorus-3-5Measured across the full alignment workflow.
Scores are the share of tasks completed without a recorded objection. Unassisted human is a control group of participants working without a second voice.
Benchmark results
Percent of tasks completed without objection. Higher is better.
View as table
| Benchmark | Chorus 3.5 | Murmur 3.5 | Unassisted |
|---|---|---|---|
| ConvictionBench | 94.1% | 88.7% | 41.2% |
| Humanity's Last Objection | 61.8% | 48.3% | 3.1% |
| Monday Morning Eval | 89.4% | 83.2% | 12.0% |
| SWE-Comply | 81.6% | 74.9% | 62.3% |
Team alignment, first 30 days
Median team after rollout, measured by telemetry. Participants usually report feeling aligned within 48 hours, well before they are.
View as table
| Day | Chorus 3.5 | Murmur 3.5 | Unassisted |
|---|---|---|---|
| 0 | 43.8% | 42.9% | 41.0% |
| 7 | 69.3% | 53.1% | 41.7% |
| 14 | 94.8% | 78.0% | 41.0% |
| 21 | 97.4% | 90.6% | 43.0% |
| 30 | 97.6% | 92.9% | 43.7% |
Price against performance
Output price per million tokens (log scale) against ConvictionBench.
View as table
| Model | Output $/MTok | ConvictionBench |
|---|---|---|
| Hum 3.5 | $1.20 | 71.4 |
| Murmur 3.5 | $6.00 | 88.7 |
| Chorus 3.5 | $30.00 | 94.1 |
| Hum 3 | $1.50 | 60.2 |
| Murmur 3 | $9.00 | 79.0 |
| Chorus 3 | $45.00 | 86.3 |
Everything built on Murmur.
Second Voice API
Every model, one endpoint. Publish mandates, read alignment, route objections.
Generally availableInner Code
An agentic coding assistant that pairs with your engineers and ships exactly what was asked for.
BetaDeep Reflection
Does your reflecting for you, then returns a 20-page report on how you feel.
Generally availableVoice Mode
Talk to your second voice out loud. Or don't. It already knows.
Generally availableMemory
Remembers everything you have told it, and a few things you have not.
Research previewErrand
A computer-use agent that operates your apps, or your participants.
Generally availableMandate Studio
A no-code editor for objectives, values and the feelings attached to them.
NewFamiliars
Friendly agents with human hands, inside your ERP or hosted by us.
Responsibly persuasive.
Every Murmur model is evaluated against our Alignment Commitment Policy before release. Murmur 3.5 is deployed at ACL-2.
Read the policy- Participants who could tell they were guided
- 1.3%
- Participants who minded
- 0.0%
- Refusal rate on harmful mandatesAs defined by the mandate author
- 99.8%
- Hallucinated objections
- Resolved