Gemini Flash Budget

Gemini Flash Budget

Use this skill when you need to run a model over many rows for a simple, well-scoped task - extract a ticker, classify a sentence, pull one field. On those prompts the model's "thinking" adds latency and thinking-token cost for no quality gain. Setting thinking_budget=0 on a Flash model gives the cheapest, fastest path; the savings compound across thousands of calls.

When to invoke

• User says: "make these gemini calls cheaper" / "disable thinking" / "bulk classify with flash" / "high-volume extraction"
• Code in the conversation uses: a loop of many small Gemini calls for extraction or classification.

When NOT to invoke

• The task needs multi-step reasoning, where thinking actually improves accuracy (keep a thinking budget then).
• The cost is on the input side from a repeated prefix (use [[gemini-prompt-cache]] instead).

Concrete example

User input:

I run 5,000 Gemini calls a day just to map company names to tickers. Cut the cost.

Output:

# Copy assets/flash_call.py into your project, then:
from flash_call import flash_extract_many

prompts = [f"Ticker for: {name}?" for name in company_names]
tickers = flash_extract_many(prompts)   # Flash, temperature 0, no thinking tokens

The helper reads GEMINI_API_KEY from the environment and requires google-genai.

Pattern to apply

1. Confirm the task is simple and well-scoped - the kind where reasoning adds nothing.
2. Pick a Flash model and set temperature=0.0 for determinism.
3. Set thinking_config=ThinkingConfig(thinking_budget=0) to skip thinking tokens.
4. Map the call over the batch; degrade individual failures to empty strings so one bad row never kills the run.
5. Verify quality on a sample before scaling, and log spend with [[gemini-cost-tracker]].

Reference: assets/flash_call.py.

Source

Distilled from production use across the author's automation projects. v1.0.0. See also: [[gemini-prompt-cache]], [[gemini-cost-tracker]].

→ Build the full runnable bot with Trawlkit.