Today we're introducing Iris-1, the first in our family of domain-specific world models, purpose-built to understand industrial environments.
Building on almost a decade of experience deploying vision AI solutions in logistics and manufacturing, we developed Iris-1 because state-of-the-art frontier models do not meet our clients' expectations for cost and accuracy in their environments.
Iris-1 uses a world model backbone derived from Meta's open V-JEPA 2.1, with a proprietary vision-language bridge that enables querying, filtering, and annotating events. To teach Iris-1 our world, we performed continued pretraining on a proprietary, anonymized video data set.
We compare Iris-1 against frontier models on a set of video-understanding evaluation scenarios. In one test focused on forklift incidents, Iris-1 correctly identified 90% of collisions, while the next-best alternative, Google Gemini 3.8 Flash, caught only 29% at 72 times the cost. Iris-1 is a major leap forward, giving AI agents the eyes they need to understand the worlds of industrial operations.
Why this matters
Industry and trade make up a third of the US economy, as measured by GDP, and almost none of it is visible to AI. In operations, agents need to see both the physical and digital world to unlock AI's power. Where systems of record capture transactions, vision sensors can understand reality, but to make this possible, AI needs to be able to understand the unique world of warehouses, factories, and beyond. Iris-1 is our first step to make this vision a reality.
While Iris-1 does not claim general-purpose vision, our goal is to offer domain-specific frontier intelligence cost-optimized to reliably perform the most important task at hand: identifying behaviors, events, and patterns from large amounts of video footage in industrial environments.
Above All Else: Balanced Accuracy
We tested Iris-1 against Google Gemini 3.8 Flash and Anthropic Claude Sonnet 5.5 on our held-out test set: videos captured by cameras on forklifts that may contain safety violations. To compare the models, we posed four safety questions to each model.
The general-purpose models ran at low, medium, and high thinking. The results were clear: Iris-1 achieved the highest balanced accuracy at the lowest cost per query.
Up and left is better. Iris-1 is one point: there aren't differentiated thinking levels.
| Iris-1 | Gemini 3.8 Flash | Claude Sonnet 5.5 | |
|---|---|---|---|
| Identified real collisions | 90% | 29% | 7% |
| Named what was hit | 74% | 17% | 2% |
| Explained a jolt that was not a collision | 62% | 12% | 6% |
| Open descriptions labeled facts correctly | 75% | 14% | 59% |
In the case of collisions, general-purpose models answer "no collision" to almost every video, so they miss almost all real collisions. Iris-1 identifies 90% of real collisions, names what was hit in 74%, and labels 75% of the facts in an open description correctly.
A World Model, Not A Language Model
It is important to distinguish between the underlying technology of our approach and other frontier large language models today. A language model predicts the next token; a world model learns a representation of the world from raw, noisy sensor input. As a result, Iris-1 works in representation space, not pixel space. Many experts argue that this offers a stronger foundation for building true intelligence for the real world.
We think the results on our dataset support this, at least in our domain.
Based on our real-world experience across hundreds of warehouses and factories, we find that the world model approach is a more effective strategy for industry applications. Iris-1 is more accurate, more cost-efficient, and extends more easily to new use cases. Detecting a new behavior is a small probe on top of the world model embedding, trained from a plain-English description and a few hundred labeled examples. With our agentic platform, AiOn, that process can already run automatically, so anyone on the floor can add a new query or filter by describing it in plain English.
It knows when it doesn't know
Every answer from Iris-1 carries a calibrated confidence. Each one is checked against the most similar events our team has labeled. When they disagree, Iris-1 says it is not sure instead of guessing. On the floor, a model that declines is worth more than one that is confidently wrong.
Same Clip, Distinct Answers
To highlight the power of Iris-1, three events from the test set are showcased below, with the corresponding results. Each model reviewed the clip and the question below. The results are uniquely different. We note that long answers are shortened for brevity, and cost is itemized per answer.
What happened: the camera truck is stationary. A second lift pulls up alongside and makes contact.
- Gemini 3.8 Flash
- Missed"No, there was no collision. Another forklift pulled up and stopped alongside the stationary forklift without making contact." $0.0045
- Claude Sonnet 5.5
- Missed"There's no clear evidence of a collision, but there is a close approach." (261 words) $0.047
- Iris-1
- Correct"Yes, the lift hit another forklift." $0.00003
What happened: the camera truck is stationary. Another forklift drives into it from the side.
- Gemini 3.8 Flash
- Wrong object"The forklift reverses and backs directly into a shrink-wrapped pallet/stack of cargo." $0.0052
- Claude Sonnet 5.5
- Missed"[The cab] stays steady in every frame. This looks like a close pass, or a near miss at most." (203 words) $0.047
- Iris-1
- Correct"Yes, the lift hit another forklift and hit product or a load." $0.00003
What happened: another lift bumps the stationary stand-up truck.
- Gemini 3.8 Flash
- Missed"A second forklift approached and maneuvered very close to the stationary forklift, but it came to a stop without making contact." $0.0050
- Claude Sonnet 5.5
- Correct"Probably a minor contact, or at least a very close call. The images don't settle it." (188 words) $0.048
- Iris-1
- Correct"Yes, the lift hit another forklift." $0.00003
Cost Is No Longer The Constraint
Understanding an operation means complete event coverage across all sites of the network. On a general-purpose model, that task is expensive and requires considerable compute. In contrast, Iris-1 is small and efficient; it can review every event we process on a handful of GPUs and answer multiple questions from a single pass through the encoder.
Iris-1 costs about $0.00003 per answer. Gemini 3.8 Flash costs 72 times as much, Claude Sonnet 5.5 1,556 times, both at their cheapest setting.
For an average warehouse with 100 lifts, applying these models to physical operations would demand significant investment to review every event. We estimate that running Claude Sonnet 5.5 on every event captured at a site with 100 lifts would cost the average operator $250,000 per year versus Iris-1 at just $160.
| One plant, 100 lifts, one question per event | Iris-1 | Gemini 3.8 Flash | Claude Sonnet 5.5 |
|---|---|---|---|
| Per answer | $0.00003 | $0.0019 | $0.042 |
| Per day | $0.44 | $32 | $685 |
| Per year | $160 | $11,500 | $250,000 |
Capability Available Today
Iris-1 is rolling out to all OneTrack clients today within the OneTrack operations platform. They can use this powerful new model in a few different ways.
First, every safety event now arrives with Iris-1's description of what happened. The home page shows what Iris-1 found across all sites: each risk factor, how often it occurs, which factors occur together, and which sites carry the most risk.
Second, users can simply search for the kinds of events they are looking for directly from within our application. This reduces time spent searching for footage of incidents or events of interest.
Third, Iris-1 is available as the "eyes" to our AI Agent, AiOn. Clients with access to AiOn can now ask their agent to perform complex vision analysis across events, such as scoring root causes, comparing behaviors, and searching for specific things like "operator dropping a pallet onto a center rider".
Near misses were up week over week, mostly in the dock aisles.
Want me to look into what is behind the near misses?
Iris-1 read all 214 near misses from the last 30 days. Most trace back to operators not looking where they travel, on 2nd shift in dock aisles 4–6. It was unsure about 12, so I sent those to your review queue instead of guessing.
Want me to save "not looking where it travels" as a live filter, so new events surface as they happen?
Looking Ahead
Iris-1 is our first entry into building and deploying these kinds of models for environments where domain-specific precision and cost-effectiveness matter. Building on almost a decade of deploying vision systems into warehouses and factories, we are already scaling the underlying approach beyond the four walls of the warehouse for other environments, such as factories and yards, all the way to farms, ports, and other places of action in the economy.
By connecting the digital world from systems of record with our domain-specific world models, we think we can unlock a bigger opportunity than any other general-purpose approach.
If you are operating a business that needs eyes for operations, we can't wait to hear from you: hello@onetrack.ai.
FAQ
I'm a client of OneTrack, do I have access to Iris-1?
Iris-1 is live for all OneTrack clients inside the OneTrack platform.
Is Iris-1 a large language model?
No. Iris-1 is a world model: it learns a representation of scenes and how they change, from video. It works alongside language models, which handle reasoning, code and conversation.
What is it trained on?
A proprietary dataset of anonymized real-world operational footage, recorded through OneTrack deployments across hundreds of sites.
Does it replace per-task vision detectors?
For forklift and material-handling operations, it covers what per-task detectors did, from one model. Other environments follow in future versions.
Can I run it on my own cameras?
Today Iris-1 runs inside the OneTrack platform. We are open to discussing broader access and applications with partners.
Method
- Test set: real forklift events from live operations, hand-labeled by OneTrack, selected for difficulty, held out from training. Four questions: collision, direction of travel, fall harness, hi-vis. A separate open-description task scores the labeled facts a free-text answer gets right.
- Gemini 3.8 Flash and Claude Sonnet 5.5 received the same question, word for word, with no instructions, examples or length limit, at low, medium and high thinking. Tables show each model's best level per row.
- API costs are billed tokens at each model's cheapest thinking level. Iris-1's cost is the power one consumer GPU draws, shared across the questions one pass answers; hardware, staff and labeling are not included. The plant example assumes 100 lifts at about 165 events per lift per day, one question per event, for one year.
- Public video benchmarks do not cover industrial operations, so the tasks are our own.
- "A third of the US economy": US Bureau of Economic Analysis, value added by industry as a percentage of GDP, Q2 2026. Manufacturing 9.7, construction 4.2, transportation and warehousing 3.2, mining 1.6, utilities 1.6, agriculture 0.8, wholesale 6.4, retail 5.7; together 33.2.