Andy Garcia · freelance dev
FRENES
Back to the blog
AI & Pricing

OpenAI Says LegalOn Cut Its Estimated Daily Codex Costs 65%. The Baseline Was GPT-5.5 in Fast Mode, Which GPT-6.1 Sol Undercuts by 80% or More at List Price.

October 19, 2026 8 min read

On Thursday, October 8, OpenAI published a customer story about LegalOn Technologies, a company that sells AI for legal work. The headline says LegalOn "halves Codex costs while maintaining development speed." The results card goes further: a 65 percent "reduction in estimated daily costs", and 20 percent off in its mature business areas.

The method is sensible. LegalOn stopped running its most capable model on everything, matched three OpenAI models to three kinds of task, turned Fast mode off by default and put budget caps on departments, groups and individuals.

I am CTO of a LegalTech built on Claude, and I orchestrate my own work with Claude Code, so "which model gets which job" is a question I answer too. I read this story as a peer, not as a prospect. The practices are worth copying. The numbers prove less than the headline suggests, and the gap is useful for anyone who has to report AI savings to a finance team.

What LegalOn did

The starting point, per the story: developers had unlimited access to the company's main model, GPT-5.5 in Fast mode. Left alone, that "would inevitably exceed the annual budget."

An internal group, the AI-powered Development CoE (AID CoE), tested models and wrote selection guidelines. Teams now start with a lightweight model and move up as tasks get harder. GPT-6 Luna handles implementation with clear requirements, mainly as a subagent. GPT-6.1 Sol handles standard design, data analysis and documents. GPT-6 Astra handles architecture and orchestration.

Three controls sit around that. Fast mode is restricted by default and allowed on individual request. Administrator settings set monthly usage limits for departments and individuals. And budgets follow the business: the established LegalOn business "was asked to improve cost efficiency by up to approximately 20%", while new businesses got generous budgets, because, as Senior Engineering Manager Yuta Tokitake explains, they put business speed ahead of cost efficiency.

None of that is exotic. It is what a finance team would ask for, run by engineers who know which tasks need the big model.

What the numbers say, and do not

The 65 percent is a reduction in "estimated daily costs". The page gives no amount, no period, no headcount, no token volume and no method. Estimated means it is not an invoice. Daily means it is not the annual budget the story opens with.

The title and the card give two different numbers: "halves" in the headline, 65 percent on the card. The page does not say which one is the bill.

The 20 percent appears on the card as a "cost reduction in mature business areas". In the body, it is what that business was asked to deliver, "up to approximately" 20 percent. The Japanese edition uses the same verb of request. The results list then counts "approximately 20% greater efficiency" as delivered, but no passage says how it was measured.

"Development speed maintained" has no metric at all. The story says teams ran tasks in parallel to make up for losing Fast mode. Plausible. Not measured.

The arithmetic the story leaves out

The summary line I would quote is this one: "A 65% reduction in estimated daily costs compared with GPT-5.5 by choosing among GPT-6 and 6.1 models." It compares two model generations, and the price list moved between them.

When OpenAI launched GPT-5.5 on April 23, it priced the model at $5 per million input tokens and $30 per million output tokens, and said Fast mode in Codex generated tokens 1.5x faster "for 2.5x the cost". At that multiplier, the old default ran at the equivalent of $12.50 per million input tokens and $75 per million output tokens, the figures OpenAI's API price list still shows for GPT-5.5 in Fast mode.

The new lineup, at Standard speed and list price, per million tokens:

  • GPT-6.1 Sol: $2 input, $10 output. That is 84 and about 87 percent less per token than GPT-5.5 in Fast mode.
  • GPT-6 Astra: $10 input, $50 output. A fifth and a third less than GPT-5.5 in Fast mode.
  • GPT-6 Luna: $0.10 input, $0.50 output. Under 1 percent of the old rate.

Two more facts frame it. Turning Fast mode off on the same model cuts the rate by 60 percent at 2.5x. And even at Standard speed on both sides, GPT-6.1 Sol is 60 percent cheaper than GPT-5.5 on input and about 67 percent cheaper on output. That bracket contains the 65.

For current models, the Codex docs I read on October 19 charge Fast mode at 2.5x on included plan usage and 2x on purchased credits. Even at 2x, GPT-6.1 Sol undercuts GPT-5.5 in Fast mode by 80 percent or more, and Astra costs no more per token.

So every model in LegalOn's new lineup, the most capable one included, costs the same or less per token than the old default. A 65 percent cut does not show that governance saved money on top of the price list. Had every token moved from GPT-5.5 in Fast mode to GPT-6.1 Sol, the per-token cost would have fallen by more than 65 percent. My reading: the difference comes from Astra's share, from more tokens (parallel tasks, generous budgets for new businesses), or from a starting mix that was not all Fast mode. That could be good news, more work for less money. The page does not give the numbers to tell.

Two caveats on my side. These are list prices per token, not LegalOn's bill. Codex plans bill in credits, and OpenAI's docs warn against using API prices to estimate included usage. For the GPT-6 models, the credit rates keep the same ratios (GPT-6.1 Sol at 50 and 250 credits per million input and output tokens, Astra at 250 and 1,250), and GPT-5.5 no longer appears on the credit table. And models differ in how many tokens they spend per task, which can move the result either way.

A baseline that has since been retired

Two dates matter. GPT-6.1 Sol shipped on September 29, at OpenAI's DevDay. The three-model lineup the story credits existed for nine days before the story went up on October 8, unless LegalOn had earlier access, which the page does not mention. A daily estimate is the kind of figure a window that short can produce. LegalOn is no ordinary customer, either: it announced a strategic partnership with OpenAI in July 2025, and it relayed the story, 65 percent included, on its own site on October 14.

Then, on October 14 at noon PDT, GPT-5.5 retired from Codex for users signed in with ChatGPT, which is how the Codex docs say Fast mode is offered. OpenAI's migration note tells paid plans to choose GPT-6.1 Sol and to replace gpt-5.5 in workspace defaults, saved model settings, custom agents and scripts. API-key sessions are not affected.

Every Codex workspace on that path had to leave the baseline six days after the story was published, governance program or not. For those users, the comparison can no longer be rerun in Codex.

The honest part of the story

The most useful paragraph comes near the end, and it is LegalOn's voice, not OpenAI's. Tokitake says the cost of individual tasks can be tracked, but the total cost of a feature release "often remains a black box." So LegalOn is building a pipeline that links each feature release to the AI cost spent on it and to the customer value it delivers. The story says plainly that the pipeline is not finished.

That convinces me more than the 65. It names the right unit. Tokens are an input; a shipped feature is an output. For BidForge, the tool I built that reads French public tender files for small businesses, the unit I would track is the tender file read, not the token. After the September 22 launches I argued that price per token is no longer the number to read. LegalOn reaches the same conclusion from the buyer's side.

What I would check before quoting a saving like this

Five things, all cheap.

Reprice the baseline. Price last quarter's usage at the current list. The gap between last quarter's bill and that number is what the vendor's price list gave you. Only the gap between that number and your current bill is what your governance did.

Make speed a logged exception. Fast mode, at 2x to 2.5x the Standard rate, is the largest multiplier in this comparison. LegalOn's request-based approach is right, and the requests are data: who needs speed, and for what.

Know what a cap does mid-task. Codex docs say that when a user hits the plan's usage limits during an active turn, the agent can keep working on that turn, subject to fair use limits. Check whether your own caps behave the same way before you size them.

Treat retirements as cost events. A retired model rewrites your baseline and your saved configurations on someone else's date. Keep model names in configuration, not in code.

Count cost per shipped unit. If you cannot say what a feature cost in AI spend, a per-token saving is an answer to the wrong question.

LegalOn's governance is a good template. Before anyone quotes its 65 percent, they should reprice the baseline themselves.

Sources

A project like this one?

I design and deploy products like this. Let's talk.

Let's talk