What is DecisionNode?
DecisionNode is a hosted "System One" decision model from Bynn Intelligence, Inc., a Delaware corporation based in San Francisco. Like Jev it writes no text: you send a state (text, JSON, up to 16 images or video as frames) and named questions to POST /v1/decide, and it answers every question in one pass. It has six answer types: choice, score, truth (the probability that a statement is true), number, point and box, the last two giving coordinates in an image. A second endpoint, /v1/systemone, accepts the noul, choice, score and box format used by Jev-style clients and benchmark harnesses.
There are two hosted models on Bynn's own GPUs: DecisionNode, the more accurate one, and DecisionNode Flash, the faster and cheaper one, both currently at version 0.9.1-rc1 with a 64K-token context. Output tokens are free; input costs $0.042 and $0.035 per million tokens. Optional "boosters" add specialist readings of each image, such as age estimation, AI-image detection, face recognition and OCR, at a price per picture.
Two open-weight models, DecisionNode Edge (2.21B parameters, fine-tuned from Qwen3.5-2B-Base) and Edge Mini (0.85B, from Qwen3.5-0.8B-Base), were published on Hugging Face on 22 October 2026 under AGPL-3.0, with a commercial licence from Bynn for closed products. They take the same request as the API and run offline.
Is DecisionNode worth it?
DecisionNode is the broadest paid decision API we have reviewed, and the least proven. Beyond the choice, score and yes/no questions every rival answers, it reads numbers on a grid and returns points and boxes in images, and on Bynn's own board the full model averages 0.817 over 15 text suites against 0.803 for Jev 1.13, 0.786 for Cloudflare's Clef 27B and 0.746 for pplx-decider-27b. Bynn labels every one of those figures "Preliminary results, October 2026", measured by Bynn on development splits of 25,134 items. We found no independent test or press coverage; Bynn asked the JevBench project to run the models on 25 October, and the maintainer replied that it will make claims only after its own run.
The suite table is unusually candid. DecisionNode leads Jev by 16.8 points on prompt injection and 5.2 on rules and policies, but trails by 12.8 points on MMLU-Pro knowledge and 3.5 on new task families, and most other gaps sit inside its own 95% intervals. Flash falls to 0.572 on MMLU-Pro. On rendered scenes pplx-decider-27b scores 0.809 to DecisionNode's 0.762. Calibration, the property the product is sold on, is not yet shown: the benchmarks page says per-type calibration results will come with the full run, and point and box confidence is described as not calibrated yet.
Treat the speed claims with the same care. The homepage race against GPT-5.6 Luna, 25 ms to 3.78 seconds for sixteen checks, is labelled "Mock timings": it is built from medians and list prices, not a recorded run. Bynn's measured figure is 19 ms end to end for a short Flash request, against 1,300 ms for the chat model, but its own JevBench submission reports medians of 0.19 s for Flash and 0.35 s for DecisionNode from Frankfurt with a fresh connection per request, so real latency depends heavily on the network.
On price it is a match for Jev rather than a bargain: $0.042 per million input tokens is Jev's price, and pplx-decider-v1.2 charges $0.02 with open weights under Apache 2.0. The data terms deserve attention before the first call. Requests and answers are kept for 30 days and, unless a workspace opts out, may train Bynn's models and build datasets and models Bynn "may license or sell" after identifiers are removed; zero retention is free on every plan but has to be requested by email. The Edge models are a real option for offline work, at 74.2% on the same board for Edge, but AGPL-3.0 obliges anyone serving a modified Edge over a network to publish its source.
Where it earns its place
A robotics engineer on a 15-person warehouse automation team
Find the next parcel to pick in a belt-camera frame and flag damaged ones
Point and box questions return pixel coordinates with a present probability, which Bynn says is the field to gate on because place confidence is not calibrated yet. On Bynn's 300 held-out photos DecisionNode places a box at IoU 0.5 or better 92.7% of the time and Edge 85.0%; Bynn's terms forbid using it as the sole control of a machine.
A trust and safety lead at a marketplace screening 2 million listings a month
Score every new listing against written policy before it goes live
At about 500 input tokens a listing that is 1 billion tokens: $35 a month on Flash, $17.50 as batch jobs, or $42 on DecisionNode, with output free. Low-confidence Flash answers can be re-checked on DecisionNode; turn off model training and request zero retention first, since listings hold seller data.
A platform engineer at a 30-person insurtech already calling Jev
Compare a second decision vendor without rewriting the integration
The /v1/systemone endpoint takes noul, choice, score and box questions in the System One format, so an existing harness can point at a new base URL. Both models list $0.042 or less per million input tokens, so the test is accuracy on your own claims, not price; claims decisions about people need human review under Bynn's terms.
Before you commit
- Setup
- Sign in with an email link or Google, prepay a balance and send one curl to POST /v1/decide; SDKs exist for Python, TypeScript, Go, Java, .NET, PHP, Ruby and Rust. Edge installs with pip install decisionnode-edge (Python 3.10 or later) and serves the same endpoint on localhost.
- Integrations
- An MCP server exposes the question types as agent tools, and Bynn publishes a ROS 2 node and an MQTT bridge. Sessions stream decisions over a WebSocket for loops above 10 a second, and batch jobs run within 24 hours.
- Data and privacy
- Requests and answers are deleted after 30 days and are used to train and improve Bynn's models unless the workspace opts out; zero retention is available on every plan by writing to privacy@bynn.com. Bynn holds a SOC 2 report and ISO 27001:2022 certification, signs HIPAA business associate agreements on request, and offers a DPA with Article 28 terms.
- Cost at scale
- 1 billion input tokens a month costs $35 on Flash or $42 on DecisionNode, and half or 70% of that as batch jobs; images run at $0.09 per 1M tokens on DecisionNode, about 1,026 tokens for a 1,024 x 1,024 picture. Edge runs on a 6 GB GPU for inputs up to about 8,000 tokens, taking 9.5 ms on an RTX PRO 6000 or 0.46 s on a Ryzen 7 9800X3D CPU.
- Limits
- 64K tokens and 16 images per request; default rate limits of 40 requests and 100,000 input tokens a second per workspace. A DecisionNode request with no ready capacity is answered by Flash, at Flash's price and named in the response, unless you send "fallback": false; Edge was evaluated only up to 8,192 tokens and lacks video and boosters.
What's good
- Six answer types, including number, point and box
- 0.817 vs Jev's 0.803 on Bynn's 15 suites (self-reported, preliminary)
- Output free; Flash at $0.035 and batch Flash at $0.0175 per 1M input
- Open Edge models at 4.43 GB (BF16) and 0.8 GB (8-bit Mini)
What's not
- Trains on requests by default; opt-out is per workspace
- Calibration results not yet published; no independent tests
- 12.8 points behind Jev on MMLU-Pro knowledge
- Edge is AGPL-3.0; closed products need a paid licence
Best for
Teams that need typed decisions about images and video, including coordinates, behind one API, and are willing to switch off training and run their own evaluation first.
DecisionNode alternatives
DecisionNode: common questions
- Is DecisionNode worth it?
- DecisionNode answers more kinds of question than Jev at the same price, but every number behind it is Bynn's own and preliminary, and your requests train its models unless you opt out. We rate it 3.7 out of 5.
- How much does DecisionNode cost?
- DecisionNode $0.042 and Flash $0.035 per 1M input tokens; images and video $0.09 and $0.08 per 1M; output free; batch 30% and 50% off. Free daily tier of 50 Flash and 20 DecisionNode requests. Edge weights free under AGPL-3.0; commercial licence by request.
- What are the downsides of DecisionNode?
- Trains on requests by default; opt-out is per workspace; calibration results not yet published; no independent tests; 12.8 points behind Jev on MMLU-Pro knowledge; edge is AGPL-3.0; closed products need a paid licence.
- Who is DecisionNode best for?
- Teams that need typed decisions about images and video, including coordinates, behind one API, and are willing to switch off training and run their own evaluation first.
- How much does DecisionNode cost for a team or in production?
- 1 billion input tokens a month costs $35 on Flash or $42 on DecisionNode, and half or 70% of that as batch jobs; images run at $0.09 per 1M tokens on DecisionNode, about 1,026 tokens for a 1,024 x 1,024 picture. Edge runs on a 6 GB GPU for inputs up to about 8,000 tokens, taking 9.5 ms on an RTX PRO 6000 or 0.46 s on a Ryzen 7 9800X3D CPU.
- Is DecisionNode safe for company data?
- Requests and answers are deleted after 30 days and are used to train and improve Bynn's models unless the workspace opts out; zero retention is available on every plan by writing to privacy@bynn.com. Bynn holds a SOC 2 report and ISO 27001:2022 certification, signs HIPAA business associate agreements on request, and offers a DPA with Article 28 terms.
- What are the best DecisionNode alternatives?
- The closest alternatives we have reviewed are Jev, Perplexity Decider (pplx-decider), Clef, d3-turbo (vLLM Semantic Router), LFM3 and ScaleDown Constrained Decoding Models (Classify, Extract, Summarize).
Answers checked Oct 28, 2026.