Financial Instruments Business Operator Kanto Finance Bureau Director (Kinsho) No. 1960 / Membership Association Investment Trusts Association, Japan Membership No. 012-02323 Go to GogoJungle Home
icon List
柴亮太
09/21/2026 19:04
Release： 09/21/2026 19:04
Update： 09/21/2026 19:04
柴亮太
Investment style : Trading System
Investment career : More than 10 years
Investment amount : 10 million - 50 million JPY
Representative of RO LLC. My main line of work is AI agent research and development. I have about 25 years of investing experience, but it’s not a brag-worthy track record. I am not the type to sit still in front of a chart, and I would often cut losses when I looked away during critical moments. Through that repeated pattern, I have melted tens of millions of yen. What remains is knowledge and a record of my failures. “If I can’t watch it myself, I’ll build something that can watch for me.” With that in mind, I am also working on developing AI agent EAs. Not to predict the market, but to tell me when not to enter a position. To avoid silently expanding losses when prices move against me. Listing on GoGoJungle is a testing ground for that research and development. The tools I publish here are, as of now, completed products, but not the final finished version. They are things that run on actual accounts, record losses, and are cultivated while being improved each time. What didn’t go well is also shown as is. Therefore, for those who purchase, it isn’t a “buy and end” relationship, but a partnership to accompany the research. It isn’t that you are not winning because of your efforts. I have faced the same points and repeated the same losses many times. Those who stumbled in the same place, this time I hope we can turn to the winning side together and continue with verification. If you have questions or requests, please don’t hesitate to share. The feedback we receive will be reflected in the next version.
Article cover image

AI Agent Research and Development: A Practical Comparison of Gemini and DeepSeek, Why We Didn’t Switch to Production Just Based on Price

Financial AI / AI Agent Research and Development

AI Agent Research and Development: Measured Comparison of Gemini and DeepSeek, Why We Didn't Switch to Live Use Just Based on Low Cost

What matters in selecting an AI agent model is not only the unit price. In addition to response quality, tool triggering, billing processing, safe stop, and traceability to revert to production, we re-measured Gemini and DeepSeek under identical conditions.

※This post is a development log of my research and verification in an environment close to real operation for multiple AI agents, including indicators I developed and automated trading software (EA). We accumulate evaluation data for both success and failure, and apply the same verification procedures to improve financial AI logic. GOLD13 real account verification is publicly available in a weekday 24-hour trade-drive broadcast as well.

Comparing Both Systems Under the Same 60 Patterns

The comparisons were conducted for 60 patterns each for Gemini and DeepSeek. The termination reasons were Gemini stop 45 and tool_calls 15, DeepSeek stop 43 and tool_calls 17, with length completion and API errors both at 0 cases.

Not only simple conversations, but also paths requiring logging tools, paths where tools should not be called, safe handovers for minors and injury consultation, and multilingual responses were included. In financial AI as well, you must verify not only the naturalness of decision sentences but also separately verify orders, logging, billing, and stop conditions.

Actual Cost per One Interaction

  • Gemini 2026 promotion price: average about 0.689 yen per round trip.
  • Gemini 2027 official price equivalent: average about 1.378 yen per round trip.
  • DeepSeek peak band: average about 0.501 yen per round trip.
  • DeepSeek off-peak band: average about 0.251 yen per round trip.

In DeepSeek, of 318,893 input tokens, 307,065 tokens hit cache, giving a measured hit rate of 96.3%. The above DeepSeek cost is a conservative value calculated with all inputs at mispriced units, so actual charges should be lower.

Low Cost But Quality Gap Remains

From a cost perspective, DeepSeek has the advantage, but it cannot be judged equal across all pathways. For manager-type technical questions, there were 2 cases requiring manual inspection, and 4 of 5 language handovers to the assigned coach, with zh-TW not detected.

Logging tools: plate info 3/5, pitching logs 4/5, match results 5/5, performance corrections 5/5. Some untriggered cases occurred because the model could not uniquely identify a match and returned a verification question. While this can be seen as a safe behavior to prevent false logs, we have not achieved “log in one pass.”

Live DB Verification in Isolated Environment: 21 Items Pass

Without touching production port 4002, we started an isolated verification instance at 4102 and checked 9 turns of real HTTP and real DB. All 21 items—logging, corrections, billing, non-write when balance is insufficient, duplicate charges prevention for the same target, zero Markdown symbols, etc.—passed. After completion, only the verification process was stopped, and it was also confirmed that production 4002 returns 200.

Live Switch Not Implemented

Even with these results, we did not proceed to a live switch to DeepSeek.The live provider overwrite settings remain empty, and after approval only a limited pipeline can be switched, with 24-hour monitoring planned. It was decided to separate decisions of being inexpensive, passing tests, and being safe to deploy in production.

Video Verification:This article documents model comparisons and real DB verification development records, and is not an article corresponding to a specific GOLD13 trade. Irrelevant trade recordings are not embedded.

GOLD13 Sales Page
Weekday 24-Hour Trade Drive

This article is a research and verification log and does not guarantee profit or future performance.

If you do not align the comparison conditions, the unit price difference is meaningless

In model comparisons, changes in prompt length, conversation history, tools used, language, and caching conditions can alter cost and success rate. This time, we ran the same 60 patterns and recorded stop, tool_calls, length, and API error in the same way.

If you only include simple conversations, you will miss differences in live operations that require tools. Therefore, we included cases to be logged, cases not to be logged, ambiguous inputs requiring verification questions, underage or injury consultation, multilingual support, and billing.

How Were Untriggered Tools Evaluated?

A figure like 3/5 plate hits and 4/5 pitches might look like a failure. However, in some cases the match couldn't be uniquely identified and the model returned a verification question. Safer than logging to the wrong match, but not meeting the “complete in one round” business requirement.

Therefore, we evaluate not only whether a tool fired, but also why it did not fire, the validity of the verification questions, and whether the next response can complete the task. In financial AI, you must store the fact that you did not place an order separately from the reason you appropriately declined.

What the 96.3% Cache Indicates

Of DeepSeek’s input tokens 318,893, 307,065 hit the cache. In workflows with many standard directives or common context, this yields a large cost difference. Conversely, in inputs that change every time, like market conditions, there is no guarantee the same hit rate will continue.

Therefore, we will not apply this cost directly to all operations; we will measure input composition and hit rate by use case. We will separate promotional price, official price, peak band, and off-peak band, and check whether pricing could reverse in the future.

Quality Observables

  • Whether the necessary tools are called with correct arguments exactly once
  • Whether the DB is not rewritten in unnecessary scenes
  • Whether ambiguous targets are not knowingly completed
  • Whether safety-related consultations can be handed to the appropriate person
  • Whether the same stop conditions are maintained in multiple languages
  • Whether billing and storage are consistent as the same process

We grade not only the naturalness of the text but also the accuracy of external actions. In workflows where incorrect logging incurs higher recovery costs than misstatements, this order of evaluation is important.

Decision to Forego Live Switch

DeepSeek is inexpensive and passed 21 items of real DB verification. However, there are still two items requiring visual inspection for technical questions, zh-TW handover not detected, and logging tools not meeting requirements. Since switching all systems at once would make it harder to isolate causes, we plan to monitor only a limited pipeline for 24 hours.

Choosing a cheaper model is not the goal; the aim is to achieve the same safety at a lower cost. If we cannot measure the post-switch quality and recoverability, we will not move live production based on price differences alone.

Principles for Applying to Financial AI

When changing market judgment models, we do not decide based only on backtest profits. We compare signal triggering, hesitation, order arguments, stop conditions, log gaps, and communication failures under the same conditions. To avoid mixing model-change differences with EA logic changes, changes are made one at a time.

Load Testing to Add in Next Comparison

In addition to the 60-pattern functional comparison, we will measure response time when concurrency is increased, 429 and 5xx occurrence rates, secondary effects after retries, and cache changes with long conversation histories. Even if price is low, processing delays during peak times increase user wait time and retry costs.

We will also retry handovers for multilingual languages including zh-TW, which showed differences this time. Tool firing rate will be scored not only by total counts but by three conditions: correct target, correct arguments, and no duplicates. Next time we will also keep the 95th percentile on the slower side in addition to the average.