1. This is not “ask AI which stock to buy”; it is “turn the research lab into software”
The interesting part of autonomous investing is not asking a chatbot what to buy today.
The real shift is automating the research loop itself:
collect ideas, turn them into hypotheses, write code, test them on data, reject failures, compare survivors with existing strategies, add useful ones to a portfolio, then search again.
Traditional algorithmic trading often meant humans designed the strategy and machines executed it. The newer model is different: humans define research directions and evaluation rules while AI agents behave like a team of junior researchers.
This is not an oracle booth.
It is a 24-hour research laboratory.
The researchers do not sleep. They occasionally faint from rate limits.
2. J-Quants matters because it removes the boring distance between an idea and a test
In Japanese equity research, data plumbing has often been a major tax before the actual research begins.
Security masters, prices, financials, margin data, short selling, dividends, futures and options can turn a simple hypothesis into a scavenger hunt.
J-Quants reduces that friction by exposing standardized market data through APIs.[1]
As of October 1, 2026, Standard costs ¥3,300 per month and Premium ¥16,500. Premium provides the full available history and a broader set of data, including morning/afternoon session detail.
The important claim is not “¥16,500 buys alpha.”
It buys fewer archaeological expeditions for CSV files.
Once data access is programmable, an agent can go directly from hypothesis to retrieval, Python implementation, backtest and stored results.
That is what makes high-volume experimentation practical.
3. When implementation stops being scarce, ideas and evaluation become scarce
Previously, ten good ideas might produce only one or two implementations.
Coding, debugging, reading API docs, cleaning data and plotting results created a natural bottleneck.
Strong coding agents change the conversation:
“Try this.”
“Implemented.”
“Now vary the conditions.”
“Done.”
“Walk-forward?”
“Done.”
“Correlation?”
“Done.”
Humanity finally received the mythical creature that turns random thoughts into code.
Ten minutes later, a new problem appears.
You need more useful thoughts.
Then AI starts generating those too, and the bottleneck moves again: how do you select the good ones?
Implementation gives way to search-space design, falsification and evaluation.
4. The dangerous superpower is producing 1,000 bad strategies at light speed
Autonomous research accelerates overfitting as efficiently as it accelerates discovery.
Generate hundreds of slightly different strategies, keep only the lucky historical winners, and the backtest can look like a genius hedge fund.
Live trading may look more like a documentary about extinction.
So autonomous research needs rejection machinery before it needs more idea machinery.
Useful gates include out-of-sample tests, walk-forward validation, transaction costs, slippage, year-by-year stability, regime sensitivity, correlation with existing strategies, turnover, liquidity, capacity, and duplicate-idea detection.
If you hire a hundred AI researchers, enlarge the trash can before the HR department.
A lab without rejection criteria is not a lab.
It is a backtest zoo.
5. What does a serious setup cost in 2026?
Start with fixed costs.
J-Quants Standard is ¥3,300 per month and Premium ¥16,500.[1]
Claude Max costs $100 per month for Max 5x and $200 for Max 20x, and includes interactive Claude Code usage.[2][3]
But $200 is not a magic ticket for unlimited 24/7 Claude Code. Max has five-hour session limits and weekly limits.[2]
Since June 15, 2026, Claude Agent SDK and claude -p usage have been separated from normal plan limits for eligible users, with monthly credits: $100 for Max 5x and $200 for Max 20x.[4]
So interactive coding and unattended agent execution can share the Claude ecosystem while having different budget mechanics.
API pricing currently places Claude Sonnet 5.5 at $2 per million input tokens and $10 per million output tokens; Opus 5.5 is $4 and $20.[5]
OpenAI GPT-6.1 Sol is also $2/$10 under Standard processing, while GPT-6 Astra is $10/$50.[6]
“AI” is not one labor price.
The researcher and the research director can cost five times apart.
6. Is ¥900,000 per month ridiculous? No. It is surprisingly easy to manufacture
Using an uncached $2/$10 model:
50M input + 5M output costs about $150.
200M + 20M costs about $600.
500M + 50M costs about $1,500.
Painful, but still within individual-professional territory.
Now use a $10/$50 frontier model with the same behavior:
50M + 5M costs about $750.
200M + 20M costs about $3,000.
400M + 40M costs about $6,000.
At a deliberately rounded ¥150 per dollar, $6,000 is roughly ¥900,000.
The ¥900,000 monthly AI bill is not science fiction.
You can summon it by running expensive models in parallel for long enough.
The final boss of invoices needs only one command.
Real systems can cut this sharply through caching, Batch, cheaper routing and early stopping. They can also exceed it through long context, tool calls, retries and agents that continue failing unattended.
So “what does the API cost?” is the wrong planning question.
The useful question is: how many research loops are you willing to authorize per month?
7. Will unit prices fall? Probably. Total spending can still rise
This is the counterintuitive part.
Epoch AI reported in September 2026 that the cost of achieving a given level of AI performance had fallen about 47% per quarter since 2023, roughly a 13-fold annual decline.[7]
There is no guarantee that pace continues.
But the direction is clear: the same level of capability has been getting dramatically cheaper.
Does that mean AI budgets collapse?
Cost per task may collapse.
Total spending may not.
If a task falls from $1 to $0.10, users do not necessarily run ten tasks and stop. They may run a hundred.
Cheaper agents enable one agent with ten ideas to become one hundred agents with one hundred ideas each, plus critic agents, plus a meta-agent searching unexplored regions, plus another wave of researchers.
That resembles the Jevons paradox.
Better efficiency does not guarantee lower aggregate consumption.
AI gets cheaper.
We respond by using much more AI.
8. The winner may be the person who routes intelligence well, not the person who buys the smartest model
There is little reason to send every task to the most expensive model.
Doing that creates an invoice-generation system, not a research system.
A sensible stack is hierarchical.
Use cheap models for broad hypothesis generation.
Use Python, SQL and deterministic filters for anything that does not need language reasoning.
Use a strong mid-tier model to implement, debug and interpret.
Apply statistical gates again.
Send only the small surviving set to the most capable model for adversarial review.
Numerical loops belong on CPUs and batch compute. Let the LLM write the program; let the computer repeat it.
The rule is simple:
Do not hire the expensive model as a factory worker.
Hire it as the research director.
The core management skill becomes allocating the right intelligence price to each job.
Conclusion
Autonomous agents plus J-Quants materially lower the barrier to systematic Japanese equity research.
As data acquisition, implementation, validation and reporting become automated, the human role shifts from manual execution toward designing search spaces and deciding what deserves belief.
The next bottleneck is AI spending.
But the likely future is not simply “AI gets more expensive.”
Capability-adjusted prices fall.
The set of possible tasks expands.
Desired usage expands even faster.
Total spending can rise.
So the future question is not:
“How much does AI cost?”
It is:
“How much intelligence should this research lab buy each month before the marginal value stops paying?”
AI may remove the implementation wall.
On the other side of that wall is freedom.
Standing next to freedom is the usage statement.
