Skip to content

To keep going please Log in.

or

By submitting my information, I agree to the Privacy Policy and Terms of Service.

To keep going please Log in.

or

By submitting my information, I agree to the Privacy Policy and Terms of Service.

To keep going please Log in.

or

By submitting my information, I agree to the Privacy Policy and Terms of Service.

To keep going please Log in.

or

By submitting my information, I agree to the Privacy Policy and Terms of Service.

Tetsuya Wakita on Why Enterprise Agent Bills Rise as AI Gets Cheaper

Diana Paluteder

On paper, enterprise AI keeps getting cheaper. Silicon Data’s index of token prices peaked at about $2.07 per million tokens in late May. By late October it had fallen below $1, a drop of more than half. OpenAI cut prices on its GPT-5.6 models by up to 80 percent in July, and Chinese open-weight models have pushed the cost of a single task down to a few cents. Companies rolling out AI agents often find their bills going up anyway. We asked Tetsuya Wakita why. He led the adoption of AI agents across the engineering organization of a publicly listed technology company. He watched the same paradox play out inside that company, and the way he resolved it explains most of what follows.

Tetsuya Wakita is a senior engineering leader with a proven track record of building, scaling, and leading high-impact engineering organizations and technology platforms. Until October 2026, he held the role of Vice President of Engineering at AnyMind Group, a Tokyo-listed technology company with engineering teams in Japan, Thailand, Vietnam, and India. He joined FourM, a Japanese ad-tech startup, in 2016 as its only engineer and came to AnyMind when it acquired the company.

There he designed and built AnyManager, a platform that helps publishers earn advertising revenue. It serves more than 10 billion ad impressions a month across more than 1,000 publisher sites in 15 countries. He became Vice President of Engineering in January 2025. In that role, he put AI coding tools in front of every engineer and tracked their use as an engineering metric, and he rebuilt the company’s internal AI agent platform from a prototype into a production system. He also wrote Zod AOT, an open-source compiler for Zod, one of the most widely used validation libraries in TypeScript.


You started as the only engineer at a small ad-tech startup. How did that lead to running engineering at a listed company?

I joined FourM in 2016 as its only engineer, so I built the whole ad platform myself. AnyMind acquired the company the following year, and I designed and built AnyManager there for a much larger network of publishers. After that I moved from tech lead to head of engineering, and in January 2025 to VP. When I took that job there were about 80 engineers in four countries.

You’ve spent the past few years leading AI adoption inside a large engineering organization. What did that involve day to day?

Mostly, getting AI coding tools into every engineer’s hands in a way we could measure. Buying licenses is easy. Knowing whether anyone is using them well is not. So we instrumented the tools, set adoption targets, and treated it like any other engineering metric instead of a nice-to-have.

At AnyMind you also introduced metrics for evaluating engineers. How far do you trust them?

They are useful for seeing the general trend, and that is all I ask of them. You cannot trust them completely, and they need constant upkeep. Say you count pull requests. People can push that number up without doing better work, for example by splitting one change into five. So no single number decides an evaluation. We started from established frameworks like DORA and kept adjusting as we saw how people reacted.

Token prices keep falling. Why do the bills go up?

Because usage grows faster than prices fall. Once people see what AI can do, everyone starts using it, and that part is good. But not many of them understand how it works underneath. When they write a skill, the only question they ask is whether it works. They test it on one model, usually the most capable one, and it ends up working only on that model. Then every call goes to the expensive model, including the ones a cheap model could handle. When Claude Opus 5 replaced Opus 4.8 this summer, plenty of teams found that skills they had not touched started behaving differently. Over the long run you have to design a skill so that it still works after you switch models, and almost nobody thinks about that at the start.

You also built Zod AOT, a compiler for one of the most widely used validation libraries in TypeScript. What problem does it solve, and how does it connect to the way you think about agent skills?

Zod checks the input to a large share of the TypeScript world, more than a hundred million downloads a week, and the official SDK for the Model Context Protocol, which is how agents talk to tools, is built on it. Zod is fast because the first time it sees a schema it generates specialized code for that shape at runtime. Edge platforms like Cloudflare Workers forbid runtime code generation for security reasons, so on the infrastructure where a growing share of agent traffic runs, Zod has to switch its fast path off. Zod AOT does the same specialization at build time instead. Nothing is generated while the server runs, so the compiled validators work anywhere, there is no cold-start cost, and in benchmarks they check complex nested objects up to 64 times faster than Zod v4’s own parser. Existing code stays as it is. It’s a build step.

Agent skills have the same problem in a different costume. A skill is the knowledge and steps an agent needs for one task, written by the person who knows the domain, and today most of them are loose prose. That is a big part of why they behave differently from one model to the next. A schema is a contract: it says what goes in and what comes out, so it can be tested. Write a skill’s inputs and outputs the same way and CI can check that it does the same thing on a cheap model and an expensive one.

How do you keep costs under control as agents spread across a company of about 2,000 people?

Separate what’s deterministic from what needs a model to reason about. The deterministic part, anything you could write as a clear set of steps, goes into code: Python, or whatever fits. Only the part that needs judgment gets written as a skill definition in something like markdown. That split is what lets a skill run on a cheap, commodity model instead of a top-tier one on every call.

Then match the model to the job. A lot of what an agent does is a small decision: which team gets a ticket, or whether a step is finished. None of that needs a model that writes prose. Decision models such as TypeSafe’s Jev and OpenAI’s Decisions API return a choice from answers you define in advance. Jev costs about four cents per million input tokens and charges nothing for output, and OpenAI says its Decisions API answers about ten times faster than its regular API. Send the small decisions there and keep the expensive model for the work that needs it.

Where do you see this heading as the model market keeps commoditizing?

The gap between frontier and commodity models will keep narrowing for most everyday tasks. Companies that have already separated their deterministic logic from their model-dependent reasoning will be able to swap in cheaper models as they catch up, without rewriting anything. Companies that built everything around one expensive model at the center are the ones who’ll end up locked in and paying for it.


Read more interviews here.

Best Crypto Exchange for Intermediate Traders and Investors

  • Invest in cryptocurrencies and 3,000+ other assets including stocks and precious metals.

  • 0% commission on stocks - buy in bulk or just a fraction from as little as $10. Other fees apply. For more information, visit etoro.com/trading/fees.

  • Copy top-performing traders in real time, automatically.

  • eToro USA is registered with FINRA for securities trading.

30+ million Users worldwide
Securities trading offered by eToro USA Securities, Inc. (“the BD”), member of FINRA and SIPC. Cryptocurrency offered by eToro USA LLC (“the MSB”) (NMLS: 1769299) and is not FDIC or SIPC insured. Investing involves risk, and content is provided for educational purposes only, does not imply a recommendation, and is not a guarantee of future performance. Finbold.com is not an affiliate and may be compensated if you access certain products or services offered by the MSB and/or the BD
Finbold Career

Join Finbold's newsroom, become a Sales Executive today!

Apply now to join Finbold as a crypto/finance news writer!

Latest posts

Finance Digest

By subscribing you agree with Finbold T&C’s & Privacy Policy

Related posts

Home

IMPORTANT NOTICE

Finbold is a news and information website. This Site may contain sponsored content, advertisements, and third-party materials, for which Finbold expressly disclaims any liability.

RISK WARNING: Cryptocurrencies are high-risk investments and you should not expect to be protected if something goes wrong. Don’t invest unless you’re prepared to lose all the money you invest. (Click here to learn more about cryptocurrency risks.)

By accessing this Site, you acknowledge that you understand these risks and that Finbold bears no responsibility for any losses, damages, or consequences resulting from your use of the Site or reliance on its content. Click here to learn more.