What Trellis Data released
On 10 November, the Canberra company Trellis Data announced Agentaus 1. It's a large language model with 220 billion parameters, which the company says makes it about 50% larger than its previous model. Parameters are the learned weights inside a model. The count gives a rough idea of a model's capacity and of the hardware needed to serve it. Trellis calls Agentaus 1 "near-frontier" and says it ranks 43rd in the world among foundation models. The announcement came in the same week Canberra set out plans for mandatory frontier AI rules. Press coverage followed in Startup Daily on 12 November, Technology Decisions on 13 November and PS News on 16 November.
Michael and Rachel Gately founded Trellis in 2016. According to Startup Daily, the company is backed by Main Sequence, the venture fund CSIRO founded. The first Agentaus platform launched in May 2026. PS News reports that the original Agentaus was built by adapting open-source frameworks, while Agentaus 1 is Trellis's own proprietary model.
How the platform is put together
Agentaus is a hosted service: chat, agents and an API built on top of the model. The Agentaus FAQ says inputs, outputs and files are stored and processed in Australian data centres on infrastructure Trellis runs itself. For security reasons the FAQ doesn't name the facility operators or their locations. The FAQ also says Trellis doesn't train models on customer prompts, outputs or files. That commitment is written into the user agreement, so customers have no opt-out setting to manage.
The pricing page describes two operating modes. Protected Mode keeps data in Australia. External Mode sends work to what the page calls "global engines", and the paid plans include monthly Claude and ChatGPT credit, though the page doesn't say how much. Builders who handle sensitive records should treat External Mode as a separate data flow with its own residency, and configure it on purpose.
The FAQ lists Trellis's security credentials. The company has held ISO/IEC 27001 certification since 2022. It self-assesses at Maturity Level 2 of the Essential Eight (the Australian Signals Directorate's baseline set of security controls), measured against the November 2023 version of that model, and plans a formal independent assessment within 12 months. Deployment on IRAP-assessed infrastructure, which has been checked under the Australian government's cloud security assessment scheme, is available on request.
The API and what it costs
The developer quickstart says the Agentaus API is compatible with OpenAI's chat completions API. Users generate keys from the dashboard and pay as they go with prepaid credits. According to the same page, tool calling works through cURL and the OpenCode coding agent, and support for the Hermes Agent and Pi frameworks is planned. API usage costs A$1.70 per million input tokens and A$6.70 per million output tokens. The docs we read don't give a base URL, model names, a context window size or rate limits, and builders will need all four before planning a production integration.
For people who don't need the API, the subscription tiers on the pricing page are:
- Free: A$0 a month, with a daily allowance of Protected Prompts
- Standard: A$24 a month, with five times the free allowance
- Professional: A$107 a month, with unlimited Protected Prompts and private agents
- Team: A$24 per seat a month for 2 to 19 seats
- Business: A$107 per seat a month, adding SSO, audit logs and spend limits
- Enterprise: custom pricing on a minimum one-year term, billed at 200 or more Business seats, with on-premises, customer-cloud or air-gapped deployment
What hasn't been verified
Trellis's announcement doesn't name the benchmark behind its world ranking. PS News reports a score of 84.4% on MMLU-Pro and says Trellis will publish detailed results soon. MMLU-Pro is a harder version of the MMLU knowledge benchmark that adds more reasoning questions and gives ten answer options per question instead of four. No outside party has published a reproduction of the 84.4% figure, and as far as we can tell no model card is available. Technology Decisions notes that Trellis says the model runs efficiently on smaller compute clusters, but the company hasn't published hardware figures to back that up.
The documentation and the launch coverage don't agree on one point. The Agentaus FAQ currently says all models running in Protected Mode are customised open-source models, and that Trellis is still building its own frontier model, with an announcement expected soon. That wording may simply predate the launch. Even so, anyone signing up should ask Trellis which model a given API key actually calls.
Michael Gately told Technology Decisions that Agentaus 1 and all of Trellis's products already meet the government's proposed AI safety and security standards. Startup Daily points out that the government is still preparing the legislation for those standards. Until the standards exist in law, Gately's statement is a company claim that nobody can check.
What it means for builders in New Zealand
Agentaus keeps data in Australia, and the FAQ makes no commitment about New Zealand. Builders whose clients require data to stay in New Zealand won't get that from Agentaus. For New Zealand personal information, the Privacy Commissioner's guidance on disclosing information outside New Zealand says an organisation that puts personal information in cloud servers will usually stay responsible for it, because section 11 of the Privacy Act treats the provider as the organisation's agent. That changes if the provider uses the information for its own purposes. In that case the Information Privacy Principle 12 checks on overseas recipients start to apply. Trellis's contractual promise not to train on customer data supports agent treatment. A New Zealand builder still carries the responsibility if a breach happens at the provider's end.
The OpenAI-compatible interface keeps a trial cheap. A team that already uses the OpenAI SDK can point a test client at the Agentaus endpoint once Trellis supplies the base URL and model name, run its own evaluation set, and compare answer quality and token cost against its current provider. Running your own tests matters more than usual here, because the published benchmark evidence is thin. Tool calling should be tested separately, since the docs only list support for two clients so far. Teams that put Agentaus 1 through real workloads can post their results on ASAI's projects page, where other members working on regional data residency can see them.
Sources
- Trellis Data joins global AI ranks with launch of near-frontier sovereign AI model: Agentaus 1 (Trellis Data)
- Trellis Data newsroom
- Agentaus pricing
- Agentaus developers: getting started
- Agentaus FAQ
- Canberra scaleup Trellis Data's sovereign AI model is now up and running for business and government (Startup Daily)
- Trellis Data launches near-foundation AI model (Technology Decisions)
- Australia's own AI powers up to rival global giants (PS News)
- Trellis Data launches near-frontier model Agentaus 1 (Australian Cyber Security Magazine)
- MMLU-Pro: A More Robust and Challenging Multi-Task Language Understanding Benchmark (arXiv)
- Disclosing personal information outside New Zealand (Office of the Privacy Commissioner)
- Principle 12: Disclosure outside New Zealand (Office of the Privacy Commissioner)
Build with people working on the same things
Join the Auckland Society for AI: NZ$599 a year with a ForgeBox and the members' projects area. Our Auckland meetups are open to everyone.