LegalOn Says Model Routing Cut Its Codex Bill by 65% Without Slowing Delivery

When coding assistants move from pilot to default tooling, the bill stops being a rounding error. OpenAI has published a customer story describing how LegalOn Technologies, a legal AI company, reduced its estimated daily Codex spending by roughly 65% by routing work across three different models, restricting a high-speed mode by default, and setting budget caps that differ by business unit. The company says development speed held steady — though the story offers no measurement methodology to back that up.

What LegalOn changed

According to OpenAI's account, LegalOn initially gave developers unlimited access to GPT-5.5 in Fast mode and pushed Codex into design, implementation and everyday work. That produced the adoption the company wanted and a budget problem it did not: unconstrained use of the highest-capability model at the highest speed setting was on track to blow through the annual plan. Blanket restrictions were the obvious lever, and also the obvious risk — they would have clawed back the productivity gains that justified the spending in the first place.

The response was an internal group the company calls the AI-powered Development CoE, or AID CoE, which tested models, monitored usage and produced selection guidance that managers passed to their teams. The stated principle is start light and escalate: begin with the cheapest adequate model and move up only as task complexity demands.

The resulting assignments, as described by OpenAI, are:

• GPT-6 Luna, the lightest model: code implementation against clear requirements, everyday automations, simpler analysis and delegated tasks, used mainly as a subagent.
• GPT-6.1 Sol: standard design and analysis, relatively simple software design, routine numerical work and document creation, and tasks that need shorter completion times than Luna.
• GPT-6 Astra, the most capable: advanced judgment calls including architecture design, complex analysis and document creation, and acting as the orchestrator coordinating other agents.

Two details are worth flagging. First, Sol is positioned both as a step up in capability from Luna and as faster for some tasks, which suggests the routing decision is not a clean single-axis ladder of price versus quality. Second, Astra is described as the orchestrator — meaning the expensive model still sits at the top of multi-agent workflows, coordinating cheaper ones. That is a common pattern, and it means savings come from shifting the bulk of token volume downward, not from retiring the frontier model.

The governance layer

Model guidance alone rarely changes behavior. LegalOn paired it with administrator settings that impose monthly usage limits on departments and individuals, with AID CoE adjusting those limits as business needs shift. The company also turned off Fast mode by default and made it available on request.

OpenAI says that change raised internal concerns about development speed, and that teams compensated with practices such as running tasks in parallel. That is a plausible workaround — latency on a single request matters less if an engineer has three requests in flight — but it is also a workflow change that shifts cognitive load onto developers. The story does not say whether LegalOn measured the effect on cycle time, review quality or developer satisfaction.

The most interesting piece is the differentiated budgeting. Rather than applying a uniform efficiency target, the company asked its established LegalOn business to improve cost efficiency by up to roughly 20%, while new businesses in the launch phase received generous budgets to encourage heavy AI use. Senior Engineering Manager Yuta Tokitake is quoted saying new businesses prioritized business speed over cost efficiency, with the goal of raising output to drive growth. Treating AI spend as a portfolio allocation rather than a flat overhead line is a reasonable framing, and one most finance teams would recognize from cloud budgeting.

Reading the numbers carefully

OpenAI's headline says LegalOn "halves" its Codex costs; the results panel says estimated daily costs fell approximately 65% compared with GPT-5.5, and the mature business was asked to find up to about 20% greater efficiency. These are different measures of different things, and the story does not reconcile them in one place.

Several caveats follow. The 65% figure is explicitly an estimate of daily costs, not an audited spend reduction over a defined period. No absolute dollar or yen amounts are disclosed, nor is the number of developers, the volume of requests, or the time window over which the comparison was made. Crucially, the baseline is GPT-5.5 — so an unknown share of the savings may come from newer models simply being cheaper or more token-efficient per unit of work, rather than from LegalOn's routing policy. Price changes between model generations are not addressed in the source.

The claim that development speed was maintained is also unquantified. There is no cycle-time metric, no deployment frequency figure and no description of how the company compared before and after. For engineering leaders trying to replicate this, that is the gap that matters most: the cost side of this trade is easy to measure, and the productivity side is not.

The ROI question LegalOn says it has not solved

To its credit, the company appears to know this. Tokitake is quoted saying it is close to a given that AI accelerates system development, and that the real question is whether faster development translates into customer value. The company says the link between the two remained unclear.

In response, LegalOn is building its own metric that treats a feature release as the unit of measurement, linking the customer value a release delivers to the AI costs invested in producing it. Tokitake notes that per-task costs are trackable but the total cost of a complete piece of work often stays a black box. The pipeline to calculate this is still being built, according to OpenAI, so there are no results — and the story does not explain how "customer value" will be defined or quantified, which is where this kind of metric usually breaks down.

Alongside that, the company plans a knowledge base to capture which model combinations work for design, implementation and review, and says AI productivity gains are prompting it to rethink how work splits between people and AI, including workforce plans and an increased emphasis on AI skills in hiring. Tokitake adds a warning that excessive rules and budget restrictions can undermine organizational momentum, and that the goal is a flexible operating model balancing risk control against team speed.

Why this matters beyond one company

Vendor customer stories are marketing, and this one is no exception — it is published by the model provider whose products are being bought. But the operational pattern is generalizable and the trade-offs are stated unusually plainly: unlimited frontier-model access is unaffordable at scale, blanket caps destroy the value, and the middle path requires an internal function that tests models, writes routing guidance and tunes limits continuously. That is headcount and process, not a setting. Organizations evaluating similar programs should price in the cost of running their own version of AID CoE before counting the savings.