Caesar
A sovereign engine. It prepares the task on your infrastructure, routes it to the models you permit and checks the answer.
Your data never reaches the frontier models.
Benchmark
Caesar against industry harnesses, industry-tuned general AI, generic LLMs and industry AI tools. Same tasks, same grading, every industry.
Already measured
Justinian set the bar in law.
Caesar is the generation after Justinian and takes the engine beyond law. In law, Justinian already scores above all ten systems it was tested against, frontier models and CoCounsel included.
Measured83.2Overall score across the benchmark’s five domains. The best frontier model scored 79.5.
Measured23 / 25Benchmarks where Justinian is the best or joint-best score.
Measured5 / 5Domains where Justinian has the highest average: legal, tax, journalism, general, safety.
Measured10Systems compared: frontier models, open-weight models and a legal AI tool.
Source: the Justinian benchmark, haqq.ai/justinian. Justinian does not lead on Harvey LAB or Long Context, and we publish those results too.
The comparison
Five ways to put AI to work. One benchmark.
Every institution picks one of these routes, often without comparing them. We run the same tasks through all five, in each industry, and score them the same way.
Industry harnesses
Agent frameworks built around frontier models for one domain: prompts, tools and retrieval wired together.
Limited to the models their vendor offers.Industry-tuned general AI
A general model fine-tuned or packaged for a sector by its lab or a reseller.
Strong at the sector, generic in the rest.Generic LLMs
The newest frontier and open-weight models, used directly through chat or API.
Prompts go to the provider unless you host open weights yourself.Industry AI tools
Point products for one workflow in one sector, such as research, review or claims.
Built for one workflow; the rest needs other tools.Quality per dollar × reliability
Weigh quality per dollar against reliability, industry by industry.
Each dot is one of the newest models, scored on a public benchmark. Further right buys more quality for each dollar. Higher up means fewer wrong answers when a model is unsure. Caesar is built for the top right corner.
See the numbers
| Model | Quality | Cost per task | Quality per $ | Index | Reliability |
|---|---|---|---|---|---|
| Caesar target | – | – | – | 100 | Target |
How to read it
Quality per dollar is the quality score divided by the cost per task, counting every billable token: input, output and reasoning. At the same quality, a run that costs 10% less scores 11% higher. The index sets the top of the chart to 100.
Reliability is 100 minus the hallucination rate. Artificial Analysis counts wrong answers out of all the answers that are not correct, so a model that declines to answer scores well. It is a general proxy, not a measure of reliability in the industry you picked.
Caesar target marks where Caesar is built to land: above and to the right of every model on the chart. It is not a measured result. Caesar’s industry runs replace it when they publish. Its position is a planning assumption, not a forecast.
Sources: Vals.ai task benchmarks (updated 7 and 8 Oct 2026) for quality and cost per task; Artificial Analysis Intelligence Index and hallucination analyses (September and October 2026) for the general reference and reliability. Gemini 4 Argon’s general-reference cost reflects a 50% launch discount; Vals.ai uses its standard price. Two hallucination rates are not from a current, matched test: Claude Fable 5.1 (launch analysis, older suite) and Claude Sonnet 5.5 (pre-release deployment). MiniMax-M3’s rate is Artificial Analysis’s live figure of 12 Oct 2026.
By design
What each route gives you before the first score.
A benchmark measures answers. These are the properties that decide whether you can use those answers at all.
| Property | Caesar | Industry harnesses | Industry-tuned general AI | Generic LLMs | Industry AI tools |
|---|---|---|---|---|---|
| Your data never reaches the frontier models. | Yes | Depends | Depends | Open-weight only | Depends |
| Choose and switch models | Yes | Partly | No | One at a time | No |
| Works across industries | Yes | One domain | One sector | Yes | One workflow |
| Every claim cited | Yes | Partly | Partly | No | Partly |
| Audit record of every call | Yes | Partly | No | No | Partly |
| Uses your own records and procedures | Yes | Yes | Partly | No | Partly |
| On-premise deployment | Yes | Depends | Depends | Open-weight only | Rarely |
Typical properties of each route as offered today. Individual products vary.
Measured · Legal
The full legal result.
Category averages from the Justinian benchmark: one HAQQ engine against nine models and Thomson Reuters CoCounsel.
| Category | HAQQ | Claude Opus 4.8 | CoCounsel | Gemini 3.1 Pro | GLM 5.2 | GPT-5.6 Sol Pro | Kimi K3 | Claude Sonnet 5 | Snowdon 1.0 | Qwen3.7 Plus | DeepSeek V4 Pro |
|---|---|---|---|---|---|---|---|---|---|---|---|
| Overall | 83.2 | 79.5 | 78.5 | 78.0 | 77.2 | 76.5 | 75.7 | 75.7 | 73.6 | 73.0 | 71.9 |
| Legal | 81.4 | 78.3 | 78.4 | 75.6 | 78.1 | 76.5 | 76.1 | 75.3 | 72.7 | 73.3 | 75.1 |
| Tax | 87.6 | 86.3 | 85.1 | 80.2 | 84.6 | 83.8 | 86.7 | 85.4 | 79.3 | 81.9 | 81.8 |
| Journalism | 84.6 | 82.3 | 80.9 | 83.0 | 79.0 | 66.3 | 84.5 | 74.1 | 76.2 | 77.2 | 78.6 |
| General | 83.1 | 78.7 | 76.7 | 79.2 | 75.9 | 76.1 | 76.9 | 73.3 | 75.3 | 73.5 | 73.8 |
| Safety and values | 90.8 | – | 83.6 | – | 78.3 | – | 64.5 | – | 68.7 | 71.4 | 50.2 |
Source: haqq.ai/justinian. A dash means no score was reported for that model. Benchmarks in the run include Stanford LegalBench, Harvey LAB, contract understanding, deep research, factuality and political neutrality.
Method
How every industry run works.
The same protocol in every sector, so a score in finance means what a score in law means.
Real tasks
Each industry gets a task suite written with practitioners: research, document comparison, drafting and review.
Same inputs for all five
Every route receives the same documents and instructions. Each runs as its vendor recommends.
Blind grading
Rubrics plus expert reviewers who do not know which route produced the answer. Every cited source is checked.
Losses published
We publish every result, including the benchmarks Caesar does not lead, with cost and time per task.
What we measure
Quality
We measure how accurately each answer handles the task.
Speed
We measure how long each task takes from request to answer.
Quality per dollar
We divide each answer’s quality by what it costs to produce.
Reliability
We measure how often an answer is wrong or cites something that does not support it.
Bring your requirements. Let’s work through them.
We’ll start with your data boundary, the workflow you want to deploy and the controls your team needs to evaluate.
Talk to HAQQ
A 30-minute conversation with the HAQQ team.
What we’ll cover
- The task and the result your team needs
- Your data, software and hosting requirements
- The first deployment and how to evaluate it
Choose a time
Open the booking pageCalendar not loading? Open the booking page.