Changes coming to Oceanir

Oceanir is becoming three products under one roof. The first visible part of that change is Manta, our flagship model. Orca 1.6 will follow it. The application is changing alongside both: a fast estimate, an exact match, and a verification will no longer look like the same product wearing different labels.

This is a larger change than a model release. It affects who receives which capability, how an analysis is assembled, and where Oceanir draws a line around use. We want to explain that before the new experience reaches more accounts.

Introducing Manta

Manta is Oceanir’s flagship model for investigations where a location needs more support than a pin on a map. It is designed for evidence-heavy work: examining the visual signals in an image, narrowing a search area, testing candidate places, and returning the most defensible level of precision the evidence supports.

That last part matters. An image may support a country, a city, a neighborhood, or an exact site. Those are different conclusions. Open-world benchmarks make the stakes plain: the same task that rewards a planet-scale model can also score a guess thousands of kilometers from the truth.[1] Manta is being built to stop at the strongest conclusion it can defend rather than turn every uncertain search into a confident address. An unresolved result is part of the model’s job, not a failure hidden behind vague wording.

Manta also combines capabilities that have appeared separately in our research work. Broad geolocation establishes where to search. Street-level matching helps resolve a place more precisely. A verification stage checks whether the proposed location is actually supported by the submitted image. Users experience this as one analysis. We do not plan to expose the internal system as a sequence of model names or technical checkpoints.

Who is getting Manta?

Enterprise organizations first, then selected pilots and research partners, then wider access as we learn. The staged release lets us study failure cases inside real review workflows — claim files, newsroom decisions, property reviews — and decide which capabilities are ready for broader availability.

Is Oceanir going private?

No. Oceanir will continue to serve individual users. Orca remains the broadly available model, and the core experience of submitting an image and receiving a location analysis is not becoming Enterprise-only. Improvements to result clarity, uncertainty, and the application workflow will reach regular accounts as the interface rolls out.

Some Manta capabilities stay controlled at first because they are more powerful, more expensive to operate, and more useful inside a defined review process. Controlled access is a stage, not a destination.

Orca 1.6 follows Manta

After Manta’s initial release, Orca 1.6 will become the next commercial release in the Orca line. Orca is being made more focused: fast to run, straightforward to integrate, and useful across the wide range of images people bring to Oceanir every day. It remains the model most users and API integrations will meet first.

Orca and Manta are no longer two sizes of the same architecture. Orca is optimized for accessible, repeatable geolocation. Manta is optimized for deeper analysis and verification where the quality of the evidence record matters as much as the final coordinate. Each model gets better at its own job: Orca does not become a smaller Manta, and Manta does not become a slower Orca.

The application is changing with the models

The current application can make very different analyses feel too similar. A fast city-level estimate and a forensic site verification can arrive in the same results panel with comparable visual weight. That makes precision easy to misread and hides the difference between a plausible location and a verified one.

The redesigned experience separates those jobs into three modes. Geo Estimate is for the leading area: it returns the best supported place at the available level of precision, shows the uncertainty around it, and offers a clear next step when more work is needed. Street Match resolves the exact site: it matches the submitted image against street-level imagery to retrieve the specific place, when the evidence supports it. Property Verify is for verification work: it checks whether a claimed address or parcel is actually supported by the image, records the reasoning behind the conclusion, and can return unresolved when the evidence does not justify a verified answer.

We are also changing how an analysis is made visible. Instead of front-loading every clue, score, coordinate, and action, the map and result panel will reveal information in the order a user needs it. The submitted image remains connected to the result. Approximate areas are shown as areas. Exact results are shown as exact results. Verification becomes a deliberate next step rather than an unrelated button attached to every answer.

Why we changed how analyses are made

A confident wrong address is worse than a useful city-level result. The former can redirect an investigation, enter a report, or be repeated as fact. The latter gives a reviewer a bounded place to continue working. Verification practice has understood this for years: careful geolocation work is deliberately slow, multi-source, and willing to end in “cannot confirm.”[3] Our earlier interface rewarded specificity even when the underlying evidence was broad. The new experience is meant to reward defensibility.

This also makes the product easier to understand. Users should know whether Oceanir has estimated an area, located a site, or verified a claim. They should not need to interpret an internal depth code or a confidence number to discover what kind of result they received. Clear language and distinct workflows are part of model quality. A defensible conclusion has always been slower than a plausible one;[4] the interface should not pretend otherwise.

Visual geolocation is not synonymous with surveillance

As geolocation improves, an obvious question follows: can this be used to spy on people? The capability is real. A research model has already beaten one of the world’s best professional players at reading a place from pixels alone,[2] and that skill scales in both directions. Oceanir builds place intelligence, not person intelligence. Manta and Orca are trained to understand visual evidence about locations. They are not trained to identify a person, recognize a face, follow an individual across images, or infer who somebody is.

A location model is still a powerful tool. Our responsibility is to decide what we build around it, which deployments we accept, and what uses we refuse. Oceanir does not build person-search or persistent tracking workflows, and we will not position Manta as a system for monitoring individuals.

What safety means for Manta and Orca

The controls match the capability. Manta’s early deployments are reviewed: the intended workflow, the organization accountable for the result, and whether a human remains in the decision process. Orca’s broader availability carries a different emphasis — honest uncertainty, sensible rate limits, and an interface that does not turn an approximate result into a false promise.

Benchmarks and technical reports

We will keep model claims tied to published evaluation rather than treating an announcement as proof. The current Manta report includes open-world geolocation results and the Street-Match report covers the exact-place retrieval benchmarks used to evaluate that capability. Protocol notes and known limitations are included with the tables.

What to take from this

Questions we expect

What happens next

Manta enters its first controlled release, followed by Orca 1.6. The three modes — Geo Estimate, Street Match, Property Verify — arrive alongside those releases rather than in one final redesign. Full release notes live with each announcement.