Isaacus Review (2026): Melbourne's Kanon Legal AI Models, Tested for Australia
The gate: Five stars need 21+ and Australian Fit of at least 4/5.
Skip to the verdictThe verdict: Isaacus is the most interesting Australian AI company most lawyers will never directly use. The Melbourne-founded outfit builds Kanon, a family of legal-specialist AI models — for search, reranking, classification and document enrichment — that its own benchmarks rank top of the legal pile. But this is a developer API, not an app. You write code or get a developer to. There is no document upload box, no chat window, no "summarise this contract" button. Pricing is in US dollars, billed per million tokens, and the default cloud runs on US infrastructure. The genuine Australian story sits in the self-hosted path, where firms and government run the models inside their own AWS account. If you build legal software, or run a legal-tech or in-house engineering team, Isaacus is worth a serious look and scores well. If you are a solo practitioner or a small firm wanting a tool to use on Monday morning, this is not it. Score: 17/25 — Very Good.
What it is
Isaacus is a Melbourne-founded AI research company building foundation models for the legal domain. Its founder, Umar Butler, was previously Assistant Director of Data Science at Australia's Attorney-General's Department, and the company lists offices in Melbourne and Riyadh.
The product is a family of models branded Kanon, sold through an API:
- Kanon 2 Embedder — turns legal text into vectors so software can search by meaning, not keywords.
- Kanon 2 Reranker — re-sorts a set of search results by genuine relevance.
- Kanon 2 Enricher — converts unstructured documents into structured knowledge graphs.
- Kanon Universal Classifier — a zero-shot model that tags or categorises documents without training.
- Kanon Answer Extractor — pulls a specific answer out of a passage (extractive QA).
The plain-English point: Isaacus does not write your advice or chat with you. It is the engine that sits underneath legal search and document tools — the kind of thing a legal-tech vendor or a corporate IT team plugs in. Like any AI tool, it does not replace qualified legal advice; it is infrastructure for retrieval and classification, not a lawyer.
What it costs in Australia
Isaacus bills usage-based, in US dollars, per million tokens. There is no monthly subscription — you pay for what you process. The published rates (USD), with an indicative AUD conversion at ~1.40 and a per-million-tokens basis:
| Model | Price (USD / 1M tokens) | Approx AUD (~1.40) |
|---|---|---|
| Kanon 2 Embedder | US$0.35 | ~A$0.49 |
| Kanon 2 Reranker | US$0.35 | ~A$0.49 |
| Kanon Universal Classifier | US$1.00 | ~A$1.40 |
| Kanon Answer Extractor | US$1.50 | ~A$2.10 |
| Kanon 2 Enricher | US$3.50 | ~A$4.90 |
There is no fee for holding an account — you are charged only on usage. New accounts add a payment method to claim free credits (Isaacus doesn't publish the exact credit amount). Isaacus states "taxes may apply based on jurisdiction," so treat GST as added at checkout, not included — and confirm GST handling against your tax invoice, since a US-dollar API bill to an Australian business has its own GST quirks. Volume and alternative pricing are by negotiation. Self-hosted deployment is available via AWS Marketplace, billed through your AWS account on separate terms.
A note on scale: a million tokens is roughly 750,000 words. So the headline rates look tiny, but real legal-search workloads run to billions of tokens. Budget on volume, not the per-million figure.
How it works / features
You sign up on the Isaacus platform, add a payment method, and create an API key. Official SDKs exist for Python (pip install isaacus) and JavaScript/TypeScript (npm install isaacus). From there, a developer sends documents to an endpoint and gets back vectors, rankings, classifications or extracted answers.
There are three ways to run it:
- Online platform / API — Isaacus hosts the models; you call them over the internet.
- AWS Marketplace (self-hosted) — the models run inside your own AWS account, air-gapped with zero outbound network access.
- Sovereign / private deployment — bespoke arrangements, including on-premises and air-gapped environments, by contract.
The standout technical claim is on the Embedder. Isaacus says Kanon 2 Embedder topped the Massive Legal Embedding Benchmark (MLEB) as of 16 October 2025, beating OpenAI's Text Embedding 3 Large by 9% on that benchmark and running materially faster than rival models. Treat these as vendor-reported benchmark results — they are Isaacus's own published figures, not an independent test we ran.
What it does well
It is genuinely specialised. General-purpose embedding models from OpenAI and Google are trained on the whole internet. Isaacus trains from scratch on legal data. For legal search and retrieval, domain specialisation is a real advantage, and the benchmark claims — if they hold up independently — back that.
The pricing model is honest. No seat licences, no minimum monthly spend stated, no lock-in tier. You pay per token. For a workload you can estimate, that is easy to model and hard to over-pay on.
The self-hosted path is a serious compliance story. Running the models inside your own AWS account, air-gapped with no outbound network access and Isaacus having zero visibility into your usage, is exactly what a law firm or government department handling privileged material wants to hear.
It is Australian-founded with public-sector credibility. The founder came out of the Attorney-General's Department, and Isaacus describes itself as the first Australian-based sovereign foundation AI company in the public-service supply chain.
Where it falls short
It is not an app. This is the headline caveat. There is no interface a non-technical person can use. You need a developer, or you need to be one. For our core reader — a small-business owner or solo practitioner — that alone rules it out as a direct tool.
The default cloud is US infrastructure. Despite the Australian founding story, the online platform and API run on US-based Azure and AWS, with the website on Hetzner in Oregon. So the standard product does not process your data onshore. Onshore processing is the self-hosted path, run in your own account — not the default.
Pricing is USD. Every invoice carries currency risk for an Australian buyer, and GST treatment on a US-dollar API bill needs checking with your accountant.
Benchmarks are self-reported. The "world's most accurate" framing rests on Isaacus's own published results. Impressive, but not yet independently verified by us.
Australian fit
This is the most Australian tool in its category we have reviewed — and the nuance matters. Founding and posture: Melbourne-headquartered, founder ex-Attorney-General's Department, explicitly positioned as a sovereign Australian AI builder. Privacy: Isaacus's policy states it will not train models on your inputs and outputs without your express consent, and does not retain inputs and outputs under normal operations (it does keep account, billing and usage metadata). It references the Australian Privacy Principles and the Privacy Act 1988. Data residency: here is the catch — the default online platform runs on US infrastructure (Azure US, AWS US). Genuine onshore, air-gapped processing comes via the self-hosted AWS Marketplace deployment in your own account, where Isaacus cites AWS IRAP assessment against PROTECTED controls for the relevant SageMaker environment. Currency: USD, GST added per jurisdiction. Support: Australian time zone and local context, which most US legal-AI vendors cannot match.
So Australian fit is strong on intent, governance and the self-hosted compliance path — but the out-of-the-box cloud is not onshore. That earns a 4, not a 5.
Score
| Criterion | Score | Why |
|---|---|---|
| Value for Money | 4/5 | Transparent usage pricing, no minimums, free credits — but USD billing and currency risk for AU buyers. |
| Ease of Use | 1/5 | Developer-only. No interface, no upload box. A non-technical reader cannot use it at all. |
| Feature Depth | 4/5 | Deep, legal-specialised model family — embedding, reranking, classification, enrichment, extraction. |
| Australian Fit | 4/5 | Melbourne-founded, Privacy Act-aware, sovereign self-hosted path — but the default cloud runs on US infrastructure. |
| Reliability & Support | 4/5 | SDKs, AWS Marketplace distribution, air-gapped enterprise option, AU time-zone support. |
| Total | 17/25 |
⭐⭐⭐⭐ Very Good (16–20).
Australian Fit is 4/5, so the Australian Fit gate is cleared on that count — but the 17/25 raw total places this firmly in the four-star band regardless. See our methodology for how the gate works.
Bottom line
If you build legal software or run a legal-tech product, Isaacus is one of the most compelling building blocks available, and the fact that it is Australian-founded with a sovereign deployment path is a genuine edge over US rivals. Trial it.
If you run an in-house legal or engineering team at a firm or government department handling sensitive material, the self-hosted, air-gapped AWS deployment is the part to evaluate — it answers the data-sovereignty question the default cloud does not.
If you are a solo practitioner or small firm hoping for a tool you can open and use, this is the wrong product. You want an application built on top of models like these, not the models themselves. Wait for the legal-tech tools that embed Isaacus, or pick a finished app instead.
If you are simply Isaacus-curious as an Australian, it is worth knowing this exists — a homegrown legal AI foundation model competing on benchmarks against OpenAI and Google is a real Australian story.
Update history
- 9 Oct 2026 — Re-checked for the AUD-claim queue (#425), isaacus.com/pricing read on an Australian connection: every per-million-token rate in the table stands (Kanon 2 Embedder and Reranker US$0.35, Universal Classifier US$1.00, Answer Extractor US$1.50, Kanon 2 Enricher US$3.50), the page still says "all amounts in USD", and the US$100 free credit for new users is still offered (plus a US$5k startup programme). No figure changes;
aud_native: falsestands. Capture:Dossiers/captures/isaacus-pricing-2026-10-09.txt. - 21 Sep 2026 — Citation fix:
docs.isaacus.com/pricingreturned HTTP 404 — found in a full-corpus citation check. Repointed toisaacus.com/pricing, which carries the same pay-as-you-go, per-token, USD-only pricing this review already describes. No claim changed. Wayback: capture pending (archive.org unavailable 21 Sep 2026). - 23 Jun 2026 — First reviewed.
Frequently asked questions
Can I use Isaacus without a developer?
No. Isaacus is an API and a set of models, not an app. There is no interface, upload box or chat window. You need to write code, or have a developer integrate it.
Is Isaacus pricing in Australian dollars?
No. Pricing is in US dollars, billed per million tokens of usage. Taxes may apply by jurisdiction, so treat GST as added, and check currency and GST handling with your accountant.
Is my data processed in Australia?
Only if you self-host. The default online platform runs on US infrastructure (Azure and AWS in the US). Onshore, air-gapped processing comes via the self-hosted AWS Marketplace deployment in your own account.
Does Isaacus train on my legal documents?
Isaacus's privacy policy states it will not train its models on your inputs and outputs without your express consent, and does not retain inputs and outputs under normal operations.
Does Isaacus replace a lawyer?
No. It is retrieval and classification infrastructure for legal software. It does not give advice and does not replace qualified legal professional advice.
Read next
Sources
- https://isaacus.com/ (primary)
- https://isaacus.com/about (primary)
- https://isaacus.com/privacy (primary)
- https://isaacus.com/blog/introducing-kanon-2-embedder (primary)
- https://isaacus.com/blog/isaacus-joins-the-aws-marketplace (primary)
- https://isaacus.com/pricing (primary; pay-as-you-go per-token pricing, USD only, US$100 free credit for new users — repointed 21 Sep 2026,
docs.isaacus.com/pricingreturned HTTP 404; content unchanged, confirmed against the live page) - https://docs.isaacus.com/quickstart (primary)
Wayback Machine snapshots queued.
Price Watch · Free
Know before your AI bill changes
A short email when AI tools change what they charge Australians: the new price in AUD, what it means for you, and whether to switch. No hype, no spam, unsubscribe any time.