# RightNow decision model

RightNow is our first decision model.

It doesn't write text. You send it a state and your questions, and for each question it returns a decision with a probability for every option, a confidence, and a calibration version. Output is not charged, and refused calls are not charged.

RightNow AI is a YC-backed GPU research lab. This is the first checkpoint of our post-trained models, and more checkpoints will follow.



## The step before thinking

**Reflex Training.** Decision models usually read the whole input, then answer. RightNow decides on whatever has arrived so far: stream in text, JSON, images, voice, PDF and video, and it returns a probability for every option at any moment. It is trained to hold back until the evidence is there, then commit. That is the step before thinking.

**Reflex Architecture.** RightNow keeps a probability for every option and updates it as each part of the input arrives.

## One model, eight inputs

RightNow takes text, JSON, robot state, codebases, PDF, images, voice and video (preview).

Inputs: text, JSON, robot state, codebase, PDF, images, voice, video (preview).

Supported formats are PNG and JPEG for images; WAV audio at 16 kHz; and PDF for documents.

Each call takes:
- up to 1,024 questions
- up to 38,000 state tokens
- up to 16 attachments, including up to 4 images, 1 audio file and 1 video

Text questions take 2 to 10 options or 2 to 4 levels. Media questions take 2 to 26 options or 2 to 10 levels.

## Streaming decisions

You don't have to wait for the full input. Send changes as they happen: append text, update the state, or add attachments. RightNow returns versioned decisions. Every new decision carries its version, and the probability for every option updates as each part of the input arrives.

## An example

State: The blender I ordered arrived with a cracked jar.

Question (boolean): Did the item arrive damaged?

Decision: true

| Option | Probability |
| --- | --- |
| true | 0.999501 |
| false | 0.000499 |

Recorded example, RightNow, Oct 26, 2026.

## Speed

| Input | Decision latency | Including the input |
| --- | --- | --- |
| Text | 30 to 33 ms |  |
| Session | 34 to 36 ms |  |
| Image | 41 ms | 178 ms |
| Audio | 34 ms | 96 ms |
| Video (preview) | 41 ms | 168 ms |
| Document | 46 ms | 506 ms |

Medians. Record MODEL-SPEED-ROW-1027, Oct 27, 2026.

| Input | Decision latency | Including the input |
| --- | --- | --- |
| Text | 30 to 33 ms (p95 32 to 34 ms) |  |
| Single token | 7.2 ms |  |
| Session | 34 to 36 ms |  |
| Image | 41 ms | 178 ms |
| Audio | 34 ms | 96 ms |
| Video (preview) | 41 ms | 168 ms |
| Document | 46 ms | 506 ms |

Medians. Record MODEL-SPEED-ROW-1027, Oct 27, 2026.

## Benchmarks

### Artificial Analysis Decision Index

RightNow scores 63.20, measured 2026-10-28. This was measured by RightNow AI on the public held-out half. It is not an Artificial Analysis evaluation. The published rows come from Artificial Analysis's own evaluation, so the two are different evaluations. The chart shows positions and values only.

| Model | Index | USD per million tasks |
| --- | --- | --- |
| d1-ultra | 65 | 32 |
| SubQ Snap 1.1 | 65 | 30 |
| Elephant 9B | 65 | 8.4 |
| Drex Max | 64 | 23 |
| RightNow | 63.20 |  |
| Drex 2.0 | 58 | 41 |
| Jev 1.14 by TypeSafe | 57 | 42 |
| Decider v1.2 | 56 | 15 |
| GLiDE | 54 | 119 |
| Microsoft-Decision-1 | 54 | 31 |
| Mercury Decide | 53 | 15 |
| Celeris-1 Decision | 52 | 32 |
| d1 | 50 | 15 |
| Decider Flash v1.2 | 50 | 8.2 |
| GPT-6 Luna (Decisions) | 48 | 79 |

RightNow: measured by RightNow AI on the public held-out half on 2026-10-28, on the model version served that day. Not an Artificial Analysis evaluation. Published rows: Artificial Analysis (artificialanalysis.ai), downloaded Oct 30, 2026.

### Amazon Polarity

| Metric | Value |
| --- | --- |
| Accuracy | 97.33% |
| Cost per case | USD 0.0005726 |
| Time per case | 0.3371 s |

Amazon Polarity. Run BENCH-FRONTIER-1026A, Oct 26, 2026.

### Invoice Processing and Security Incidents

| Model | Agreement | USD per workflow | Seconds per workflow |
| --- | --- | --- | --- |
| RightNow | Pending | Pending | Pending |
| Claude Haiku 4.5 | 42.9% | $0.0558 | 30.8 s |
| Claude Opus 5 | 78.4% | $0.4856 | 92.1 s |
| Claude Sonnet 5 | 72.9% | $0.3616 | 241.3 s |
| DeepSeek V4 Flash | 69.8% | $0.0133 | 84.1 s |
| DeepSeek V4 Pro | 72.7% | $0.0830 | 137.4 s |
| GPT-5.6 Luna | 67.8% | $0.0081 | 21.4 s |
| GPT-5.6 Sol | 79.1% | $0.2152 | 34.3 s |
| GPT-5.6 Terra | 74.7% | $0.0778 | 17.3 s |
| Jev | 61.8% | $0.0011 | 0.5 s |

TypeSafe's published workflow boards, snapshot Sept 28, 2026. RightNow: Pending.

| Model | Agreement | USD per workflow | Seconds per workflow |
| --- | --- | --- | --- |
| RightNow | Pending | Pending | Pending |
| Claude Haiku 4.5 | 58.8% | $0.0047 | 3.4 s |
| Claude Opus 5 | 66.2% | $0.0574 | 15.1 s |
| Claude Sonnet 5 | 60.8% | $0.0271 | 18.9 s |
| DeepSeek V4 Flash | 37.9% | $0.0032 | 37.4 s |
| DeepSeek V4 Pro | 41.7% | $0.0234 | 60.0 s |
| GPT-5.6 Luna | 52.1% | $0.0013 | 7.0 s |
| GPT-5.6 Sol | 62.5% | $0.0295 | 8.5 s |
| GPT-5.6 Terra | 51.2% | $0.0119 | 5.7 s |
| Jev | 61.7% | $0.0001 | 0.3 s |

TypeSafe's published workflow boards, snapshot Sept 28, 2026. RightNow: Pending.

## RightNow and Jev by TypeSafe

| Capability | RightNow | Jev by TypeSafe |
| --- | --- | --- |
| Inputs | text, JSON, robot state, codebase, PDF, images, voice, video (preview) | Text, JSON; images opt in. |
| Voice | WAV audio | Not supported |
| Video (preview) | MP4 video (preview) | Not supported |
| Files | PNG, JPEG, WAV, PDF, MP4 | PNG, JPEG (opt in) |
| Standard context | 38,000 tokens | 64k per-request / 32k per-question |
| Questions per call | Multiple questions per call | Multiple questions per call |
| Probabilities | Per-option probabilities | Typed probabilities |
| Streaming | Streaming |  |
| versioned decisions |  |  |

Jev by TypeSafe column from TypeSafe docs, read Oct 28, 2026.

## Pricing

| Item | Price |
| --- | --- |
| Text input | 0.19 per million tokens |
| Cached text input | 0.019 per million tokens |
| Output | Not charged |
| Answered question | 0.0001 per answered question |
| Call fee | 0.00001 per answered call |
| Image | 0.003 per image |
| Audio | 0.0002 per started second |
| PDF | 0.002 per page |
| Video (preview) | 0.0008 per started second, 0.008 minimum per clip |

USD. Rate card sz-card-2026-11-11-h, as of Nov 11, 2026.

The call fee is never charged on refusals, zero-answer calls, or free, promotional or replay calls.

Two reference costs:
- 356 uncached tokens with one answered question costs USD 0.00018
- 38,000 tokens with one question costs USD 0.0073

New accounts get USD 1 in credit at sign-up. It's one per eligible verified email, subject to eligibility checks. Top-ups start at USD 5.

## Start building

- API: `https://api.rightnowai.co/v1`
- Docs: https://www.rightnowai.co/docs/decision-model
- SDKs: `@rightnowai/system-zero` on npm and `rightnow-system-zero` on PyPI

Updated 2026-11-13: the model's name is now RightNow decision model.