OrcaCyber Zero 1.0 and the $150,000 reserve: OrcaRouter joins the Omacom Foundation as a Distinguished Corporate Patron

OrcaRouter has joined the Omacom Foundation as a Distinguished Corporate Patron, with a donation of $150,000 in tokens to the Foundation's Strategic Token Reserves. That reserve is directed at security work, the Foundation's security and kernel teams first, and not at general inference. The first model those teams will draw on is OrcaCyber Zero 1.0, our coding model post-trained for vulnerability analysis, reproduction, and security research, which is in closed beta today. That is the announcement, and the rest of this page is the reasoning behind it.

The reasoning started with traffic. In the 30 days before this announcement, our models passed 1.5 million downloads on Hugging Face, and the shape of what moved through the router changed with it.

Cheapest-model-for-this-summary stopped being the only question arriving; "can anything here read this crash, this patch, or this report and tell me what it actually is" became a common one. A router is the wrong place to answer that question and a good place to notice it. So rather than discounting general inference, which is already cheap and getting cheaper, we are paying for the part of the work where a wrong answer costs the most.

What OrcaRouter joined, and where the tokens go

Distinguished Corporate Patron is the Omacom Foundation's corporate tier. The patronage is a single donation rather than a subscription, and the donation is $150,000 in tokens paid into the Strategic Token Reserves, one pool that the Foundation's own teams draw on for their own work.

The destination matters more than the amount. A reserve is not a coupon, and none of this ships as a free tier or a discount code. The tokens are spent when a Foundation engineer runs a real task through OrcaRouter, the same routing layer every other user touches. We would rather fund the compute behind a specific engineering question (does this bug reproduce, does this patch hold) than underwrite an open-ended pool of chat.

Why the donation is directed at security work and not general inference

General inference is the solved half of this market. A competent general model is available cheaply, routing to it is close to a commodity, and an extra $150,000 of general capacity buys a team very little it could not buy already. The marginal token there is worth roughly what it costs.

Security work behaves differently in three ways that matter to whoever is paying the bill.

• The tasks are long. Reading a report, building a reproducer, and checking a fix are not single-turn jobs, and they consume tokens in a way that summarisation does not.

• The models that do them well are specialised rather than general.

• The work has a checkable answer: a vulnerability either reproduces or it does not.

That last property is what makes the spend defensible. We are not funding opinions about security. We are funding attempts that end in something a kernel team can look at and accept or reject.

What was OrcaCyber Zero 1.0 trained to do?

Three jobs, all of them security jobs: vulnerability analysis, vulnerability reproduction, and security research.

Analysis is the reading task: given a report, a diff, or a component, work out what the flaw is and where it lives. Reproduction is the build task, and it is what separates a security model from a general coding model with a stern system prompt attached: produce an input that actually triggers the condition, not a description of one. Security research is the open-ended part, where nobody hands you a tidy ticket and the target has to be found first.

The measured result we have covers the reproduction job. On CyberGym Level 1, a benchmark drawn from 188 open source projects, OrcaCyber Zero 1.0 reproduces 1,478 of 1,507 real-world vulnerabilities on the first attempt, a 98.07% pass rate. First attempt is the number to read twice. The measurement offers no second try, and neither does a release window. What the model is and where it sits in the catalogue is on the OrcaCyber Zero 1.0 page.

Where the perimeter actually is: the calls a model makes

Zero 1.0 reads and reproduces flaws in code. It says nothing about the layer most teams are now building around it. The moment a model is handed tools, the interesting failures stop being text and become calls: which tool it picks, what arguments it passes, whether the result it gets back is allowed to change what it does next. A model that reads C well does not protect a pipeline in which an agent can be talked into calling the wrong endpoint with the wrong arguments.

That is the second half of the security work here, and it is why the word "security" in our catalogue does not begin and end with vulnerability research. Teams running agents against internal systems have to treat the call layer as an attack surface, which is the problem an agent firewall for tool and MCP calls exists to solve. Kernel work and security work see both halves of this, and the reserve covers both.

What does the closed beta mean for a team that wants access?

Access is granted per use case, not switched on for everybody. OrcaCyber Zero 1.0 is in closed beta, and the scope is drawn narrowly on purpose: trusted security research, vulnerability analysis and reproduction, and red-team workflows. Closed is the operative word, and access is scoped to that work rather than to general model traffic.

In practice, the request that moves fastest is the one that arrives with a target attached: a component you maintain, a class of vulnerability you chase, an engagement you are running this quarter. If you want in, say what you intend to point the model at. General chat requests do not queue ahead of a security team holding a real test case, and they should not. The beta widens as the evaluation holds up; we are not opening it faster than we can judge what comes back.

How the reserve gets spent

Tokens in a reserve are not assigned to a single model, and the mechanics are worth a paragraph because they explain the shape of the donation. Requests arrive at the router, and adaptive routing decides where each one goes. In practice, the split looks like this.

• Patch review, crash triage, a reproduction attempt — Where it goes The security-tuned route: OrcaCyber Zero 1.0 vs Why It is long, it is specialised, and the answer is checkable

• Release notes, changelog summaries, test names, a script's docstring — Where it goes A general model such as GLM 5.3 Flash vs Why It costs less and is better suited to the job

This is why the donation is tokens rather than a free tier of one model. The Foundation's teams get the same routing decisions as everyone else, with the security route available when a task deserves it and not charged against them when it does not. A reserve spent this way lasts longer than a reserve spent on one model, and it leaves the choice of model with the person who knows the task.

The takeaway

Security work is the right place for this money, and the reason is not sentiment. General inference is a commodity with a falling price; vulnerability reproduction is a task with a verifiable answer and a high cost per attempt, and it is where a donated token does measurable work. A kernel team does not care how a model sounds. It cares whether the reproducer runs.

If you already route through us, nothing about your traffic changes this week. What changes is on our side: the model we consider the most consequential thing we have built now has a first demanding user, one with a hard problem and no patience for a plausible-sounding wrong answer. That is a better test than any launch post, and we would rather be measured by it.

Sourcing note: vendor-reported figures are the $150,000 token donation to the Omacom Foundation's Strategic Token Reserves, the scope that reserve is directed at, and the closed-beta status and access conditions of OrcaCyber Zero 1.0. Our own measurement is the CyberGym Level 1 result of 1,478 of 1,507 first-attempt reproductions, a 98.07% pass rate, on a benchmark drawn from 188 open source projects, which we ran ourselves. Platform-reported is the 1.5 million downloads of our models on Hugging Face across the 30 days before this announcement, taken from the platform's own counters. Nothing here quotes a third party for a number, and no outside party is credited with a finding. Published 2026-09-19, figures checked against our own records the same day.