What AI Use and Token Consumption for Legal Work Really Costs (Actual Numbers)

Takeaways:

• We measured token costs for legal tasks of varying difficulty across most available LLMs.
• Simple legal question: pennies. A straightforward case law query ran close to a dime on some OpenAI models and a fraction of a cent on open-weight models, across 1,595 responses from 57 models.
• Medium task (draft a client alert): $0.0405 on average across 200 responses — roughly double the simple task, with a wide spread from $0.27 (Fable 5) down to fractions of a cent (Mimo).
• Hard task (insurance coverage memo): $1.4274 average per completed response, median $0.2238. Fable 5.1 averaged $10.47. GLM 5.3 Flash did it for $0.004.

What Does Using AI for Legal Cost? We Tested Three Legal Tasks of Varying Complexity.

There has been much talk as of late about the real cost of artificial intelligence use not just for legal work, but in all industries. It is no secret that AI bills get expensive and companies are spending millions, if not billions on AI token usage. As far as the legal industry goes, at least two big name law firms recently announced they are bringing AI infrastructure in house rather than rely on other legal AI options. Some surmise that at least one reason is because AI pricing will inevitably move to consumption based.

A review of articles to date addressing the cost of token usage on legal work have mostly been estimations and, in some instances, speculation. One of the datapoints we collect via Certera during pairwise AI comparisons of legal related LLM output is what each prompt costs and overall average token cost for legal work. You can see these numbers in real time on our Legal AI Leaderboard.

So, what does it really cost, on average, to run a legal question or legal workflow though an AI model? If you look at our leaderboard, the short answer would appear to be pennies. However, looking at cost estimates in aggregate may or may not be helpful.

For example, as detailed below, a straightforward legal question costs nearly a dime with some OpenAI models, but a fraction of a cent to get an answer from open weight models. We track nearly 70 LLMs and obviously the average law firm or corporate legal department does not have that many AI subscriptions. In fact, if they are not using a legal based AI tool, law firms and in-house legal teams are most likely using bigger name models that are generally more expensive, so just saying an average legal prompt costs a few cents across all models is not that accurate nor helpful.

To get a more precise figure for legal AI pricing, we ran three task types through the many different LLMs we track. We tested what we will call a “simple” legal query, that is, a straightforward legal question with a short(ish) answer, a “medium” task, which is a task that is a little more labor intensive and requires tool use (i.e. like creating a document), and finally, we used a “hard” legal prompt that requires some deep thinking and is a multi-step legal workflow.

For the easy task, we kept it simple and used a case law related inquiry from the CaseHOLD dataset. For the “medium” task, we gave the model a United States Supreme Court opinion and asked the LLMs to create a law firm client alert about the holding. Finally, for the “hard” task we used the coverage task from the legal evaluation dataset we released earlier this year. That prompt is a full agentic legal workflow requiring the review of an insurance policy, relevant case law and then drafting a coverage analysis.

A caveat about this analysis: our goal was solely to determine what it actually costs to run legal tasks varying in complexity. There is probably little one could do to lessen the cost of the “simple” and “medium” tasks we tested. However, for the more complex insurance coverage workflow, assuming that it is a task a legal team uses often, there are techniques that can be used to lessen token costs by reusing certain information rather than having AI start from scratch every time. For really complex agentic tasks, different models can be used for different parts of the workflow. However, that type of analysis is beyond the scope and objective of this research, and our aim was truly just to determine true AI usage costs for a stratum of legal tasks. Also of note is that some of the very new models were not run on all three tasks because they were released during our research.

The “Simple” Task - a Question About Case Law

For the “simple” task in our test, we analyzed 1595 AI responses to legal questions from the CaseHOLD dataset. CaseHOLD questions are multiple choice questions derived from legal citations in court rulings. The prompt itself is a citation of legal authority in a legal opinion, but the actual holding cited is excluded from the prompt, so the LLM must correctly identify the missing holding language.

As set out in the table below, there is a wide cost range among the models tested, with OpenAI and Anthropic models being the most expensive and models like Minimax and Ministral least expensive.

However, maybe somewhat expected is fact that the accuracy rate for GPT 5.4, the most expensive, was 97% while the cheapest, Ministral 14b, was right only around 27% of the time. The overall failure rate on these tasks across all models, meaning they incorrectly identified the holding, was just over 20%.

Average Cost per Model Response — Simple Legal Tasks

Combined dataset: 1,595 model responses across 57 models. Sorted by average cost, highest to lowest.

The “Medium” Task - Draft a Law Firm Client Alert

For the medium task, we wanted the prompt to invoke a tool call, meaning we wanted to interact with an external tool via API call requiring more work than just answering a question, like drafting a document. To do so, we used one of the ready-made Certera legal workflow automations that creates a law firm client alert. For this exercise we had the LLMs review a recent United States Supreme Court opinion and create a client alert based on the holding in .docx format.

Interestingly, the average cost for the medium task was basically double that of the simple task at roughly $0.04 per task. However, a closer look reveals a decent sized delta. Anthropic models were again among the most expensive costing anywhere from $0.13 (Opus 5) to nearly $0.27 (Fable 5) to generate the client alert. Google Gemini Flash 3.5 was not far behind at roughly $0.14, but the models at the low end were much less expensive with the Mimo models coming in at fractions of a penny.

Doing a qualitative analysis for this task is a little more difficult than for the case holding task because output from both models followed instructions and produced a solid client update. Fable 5 covered a few more details than the Mimo output and Fable also had a higher wordcount. At the end of the day, both would have served as solid first drafts and any decision about quality would come down to author preference.

Average Cost per Model Response — Moderately Difficult Legal Tasks

200 model responses across 49 models. Average cost per model response: $0.0405. Sorted by average cost per response.

The Difficult Task - Using AI to Draft An Insurance Coverage Opinion

For the more involved and difficult task, we used a multi-step workflow in which the AI models draft an insurance coverage opinion. The task is taken from the legal evaluation dataset we released earlier this year and presents AI models with a factual scenario and instructs it to create a coverage opinion.

To do so, the AI models must review an insurance policy to identify relevant policy provisions, review at least five different court opinions and then draft a coverage memorandum. We ran the pairwise comparison 150 times, so there should have been a total of 300 individual coverage memo outputs, but in the end only 219 tasks completed, with many of the models timing out during the last step which was document generation.

Not surprisingly, this task was multiples more expensive than the other less difficult tasks. Anthropic’s Fable 5.1 clocked in at the most expensive at an average cost of $10.47 for the coverage task and on the other end of the spectrum, GLM 5.3 flash cost only $.004 to complete the task.

Of most interest is the quality of the memos produced from the most expensive model, Fable 5.1 compared to the least expensive, GLM 5.3 Flash. We used both human grading and LLM as a judge for this task to determine whether or not the models produced a passable output. Under both human grading and LLM as the judge, Fable only edged out GLM output by just over 5 points. The human grader gave them both much lower grades than the LLMs, giving Fable 74 out of 100 and GLM 5.3 a 69. Word counts in both memos were comparable and again, just like for the moderately difficult task, the winner at the end of the day would come down to author preference.

Average Cost per Model Response — Difficult, Multi-Step Legal Tasks

219 completed responses (of 300 total) across 49 models. Average cost per completed response: $1.4274 (median $0.2238). Sorted by average cost per response.

This is obviously not the final word about what it really costs to use AI for legal work or where it will end up, but is merely an empirical analysis of actual costs incurred to run these tasks. As noted above, it is not a far stretch to assume that pricing for legal AI will become consumption based. If that happens, it would seem that multiple models will be used to perform different tasks (many apps already work this way).

Regardless of how pricing shakes out, understanding the true token cost of legal AI workflows remains important. Rather than relying on a single one-size-fits-all LLM, law firms and corporate legal departments may need to strategically orchestrate multi-model strategies—reserving premium, top-tier models for complex, high-stakes agentic tasks like insurance coverage analyses, while routing routine queries and document drafting to cost-effective alternatives. By aligning model selection with task complexity, legal organizations can optimize performance, manage expenses, and unlock the full potential of AI technology responsibly.