Why did Gemini 4 Argon return empty answers?

Google's documentation says a non-streaming REST request returns up to 32,768 tokens, down from 65,536 on Gemini 3.x, and that cap includes reasoning tokens. On hard problems the model used up the whole budget thinking and produced no answer text.

How was the blank-answer problem fixed?

The author switched to streaming via streamGenerateContent, where the cap rises to 65,536, and after that 5 of 6 runs scored full marks with reasoning between 39,000 and 54,000 tokens.

Can the AI models actually be ranked from this test?

Not cleanly. The results were nearly tied, 10 of the 12 problems were solved perfectly by all four models, and GPT-6 Astra could only be run once per problem because of usage limits, so the author says the numbers cannot be compared as they stand.

Large Language ModelsAI Coding AssistantsZenn AI/ML1d agoOct 15, 2026, 10:00 JST

Gemini 4 Argon's 32,768-token cap trips up a three-AI test

Send on LINE

3 Key Points

What happened
On 12 programming problems run 120 times, Claude Opus 5.5 scored 36 of 36, Gemini 4 Argon 35 of 36, GPT-5.6 Sol 33 of 36 and GPT-6 Astra 11 of 12.
Why it matters
Gemini 4 Argon returned blank answers on hard problems because its non-streaming REST cap is 32,768 tokens, counting reasoning, down from 65,536 on Gemini 3.x.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, in your inbox every morning.

Get news like this every morning, free

FAQ

Why did Gemini 4 Argon return empty answers?: Google's documentation says a non-streaming REST request returns up to 32,768 tokens, down from 65,536 on Gemini 3.x, and that cap includes reasoning tokens. On hard problems the model used up the whole budget thinking and produced no answer text.
How was the blank-answer problem fixed?: The author switched to streaming via streamGenerateContent, where the cap rises to 65,536, and after that 5 of 6 runs scored full marks with reasoning between 39,000 and 54,000 tokens.
Can the AI models actually be ranked from this test?: Not cleanly. The results were nearly tied, 10 of the 12 problems were solved perfectly by all four models, and GPT-6 Astra could only be run once per problem because of usage limits, so the author says the numbers cannot be compared as they stand.

Zenn AI/MLRead Original Article

Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Get it free

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.