O

Olava

The model is the product. Onit AI Labs trained it narrowly on contract work, so what a customer buys is inference: extraction, drafting, review, summaries and classification produced by olava-3.0 or olava-2.0. The API, the command line tool and the downloadable weights are three routes to that model rather than an application with AI added on, and nothing would remain to sell without it. Legal primitives such as DOCX and PDF operations, tracked changes, comments and document assembly exist to carry the model's output into documents a lawyer can open and edit.

The Olava 2.0 model card reports results for each contract skill against two frontier models. On structured extraction it reports a micro F1 of 0.84 and a macro F1 of 0.81, against 0.82 and 0.81 for Claude Opus 4.8 and 0.79 and 0.78 for GPT-5.5. Drafting accuracy is 90.0% against 90.4% and 86.4%. Playbook review accuracy, measured as criteria passed over defined criteria on 160 items, is 88.4% against 88.3% and 79.9%. The research behind the card is cited by arXiv number, including A Few Good Clauses on extraction.

The card does not describe the contracts tested or name failure modes, and it gives no figures for olava-3.0, the current default. Nothing describes how an extracted value is tied back to its place in the source contract. The terms state that output may be incomplete or inaccurate and is not guaranteed free of hallucinations. Olava does not retrieve case law or statutes of its own, so there is no authority to cite.

Controls sit in the command line tool, where the model works as an agent; API calls simply return answers. Tool actions pass through permission prompts, and headless runs and evaluation suites record whether an approval was answered automatically. A flag that skips permissions exists for trusted automation only. Connecting an MCP server gives the model every tool that server exposes, so configured servers start disconnected. The review skill produces structured issues for human reviewers rather than final decisions.

The trial terms make the customer solely responsible for reviewing output, and they bar uses that make, or are a substantial factor in, decisions with a material legal effect on an individual. No confidence threshold stops the model, and no route is described for what happens after a wrong answer.

No customer, deployment or production result is named. Olava describes itself as not yet generally available, and access runs through a request form or a self serve console under trial terms. The cost figure on the model card, 78% lower cost per document under batched extraction, is a benchmark result rather than a customer outcome. Onit's earlier papers, Better Call GPT and Better Bill GPT, compared frontier models with lawyers and invoice reviewers; they are research that led to the model, not deployments of it.

Each hosted request is processed, answered and discarded. Prompts, answers, chat history and working files are not stored, and the model download keeps all content inside the customer's environment. Account systems hold only names, emails, team membership, hashed API keys, usage counts and payment details.

Section 8 of the Olava trial terms protects confidential information and promises advance notice of any legally required disclosure where the law allows it. Section 12.1 bars third party model providers from training on customer input for the benefit of others. The same terms let Onit use customer data in aggregated and anonymized form for product improvement, analysis and benchmarking, and use deidentified data to improve its AI services.

Olava's own description says content is not stored; the terms still reserve that right. Privilege and work product are not addressed, and nothing describes separation between customers or between matters inside one organization.

Section 4.12 of the trial terms states that Onit is not a law firm and gives no legal advice, recommendations, opinions or representation. Section 12.3 makes the customer solely responsible for reviewing and relying on output, including getting legal or other professional advice where appropriate.

The marketing pitches Olava as legal AI that applies legal judgment at scale, and the $9 Tinkerer plan opens it to anyone who signs up. Nothing limits use to lawyers or legal departments, no jurisdiction limit is stated, and the supervision and competence duties of a lawyer using it are not discussed.

Each contract skill was tested against Claude Opus 4.8 and GPT-5.5, and the model card sets out how. Extraction is measured by micro and macro F1, drafting by accuracy, and review by criteria passed over defined criteria on 160 items, with results for all three models. The research behind it is cited by arXiv number.

The trial terms prohibit uses that make, or are a substantial factor in, decisions with a material legal effect on an individual, or that decide access to health care or other services. No AI policy, governance framework or accountable owner is named, and nothing addresses uneven output across contract types, jurisdictions or languages.

Legal content stays out of the account systems. Hosted requests are discarded once answered, the command line tool keeps data and history on the user's computer, and the model download sends no content to Onit at all. What is stored is listed: names, emails, team membership, one way hashed API keys, usage counts, payment details held through Stripe and the history of accepted terms.

Six named subprocessors each have a stated role, data access and certification. Railway hosts the console and API, Supabase runs accounts and sign in, RunPod runs inference, Stripe handles payments, Resend sends email and PostHog counts anonymous usage. Teams have Owners and Members, removed members lose access at once with their keys revoked, and membership, cap and key changes are logged. The terms promise commercially reasonable safeguards but state no incident or breach notice practice.

Section 10.2 of the trial terms caps each party's total liability at the fees paid or payable under the order in the 12 months before the claim, and section 10.1 excludes indirect, consequential and punitive damages. The services and output are provided as is. Section 12.5 disclaims accuracy, fitness, noninfringement and error free operation for the AI features, and section 12.3 says Onit is not liable for decisions taken on output.

The only indemnity runs from the customer to Onit, and no warranty or insurance stands behind the output a customer receives.

Connections run through MCP. Microsoft 365, Google Workspace and CourtListener servers can be added in the command line tool and connected for headless runs. Setup covers the trust boundary each connection opens, since a connected server gives the model all of its tools. The API follows a chat completions shape and can be used inside third party command line tools. DOCX and PDF files can be read and written with tracked changes, comments and OCR, and packaged skills can be deployed as APIs or internal services.

Onit describes Olava as intelligence inside the legal operations platform its customers already use. No connection to Onit's own CLM or ELM products, a document management system or a Word add in is documented, and what each MCP server moves is left to the server.

Three deployment routes exist, each with a stated privacy boundary. The downloaded model is licensed and run on the customer's own infrastructure, and only account and license traffic reaches olava.ai. Hosted inference runs on RunPod behind an API hosted on Railway, answering each request and discarding it. The command line tool keeps files and history on the user's computer and sends prompts to hosted inference.

What reaches Onit in each mode is set out, and content is kept apart from account and billing data. No region is named for hosted inference or for the account database, and choosing where data is processed means running the model yourself.

No security attestation is stated for Olava or for Onit's operation of it. Its six subprocessors carry their own reports: SOC 2 Type II for Railway, RunPod and Resend, SOC 2 and ISO 27001 for Supabase, PCI DSS Level 1 for Stripe and SOC 2 with HIPAA for PostHog. Those cover the providers rather than Olava itself, and no trust center, penetration test summary or route to request a report is offered.

The models are Onit's own. olava-3.0, the default, and olava-2.0 are domain trained small language models with a 256k token context window, chosen by name in each API call against a fixed allow list. RunPod serves hosted inference, and the self hosted path runs the same model on the customer's hardware.

The base model each was trained from is not named, and the model card covers only Olava 2.0. Section 12.4 of the terms lets Onit improve, replace or deprecate models, and section 12.8 allows third party model providers or infrastructure, with no notice commitment for either. Pinning a model by name keeps a production call from shifting when the default changes.

Hosted use is priced by seat and token. Tinkerer costs $9 a month with 35 million tokens. Team costs $199 per seat a month with 230 million tokens per seat. Above the allowance, overage runs at $1.90 per million tokens, input and output counted together and billed a month in arrears.

Every account sets an overage cap between $1 and $10,000 a period, $500 by default, and service pauses at the cap. Allowances reset monthly without rollover, added seats are prorated, and a removed seat counts until the cycle ends. An account includes the API, model downloads and the command line tool, and no setup or implementation fee appears.

Olava is built for contract lifecycle work: extracting structure from existing agreements, drafting and amending language against playbooks, and reviewing agreements for risk, obligations and deviations. Repository extraction, due diligence and recurring contract audits are the named uses. Its stated users are legal teams, legal operations functions and the legal engineers and developers who build tools for them, and Onit places it inside legal operations.

The model is trained narrowly on contracts, so litigation, research and advice sit outside its stated focus. No list of unsupported document types, practice areas or jurisdictions is given, and multilingual support is claimed without naming the languages.