benchr
Flash-Lite earns its page because cheap tokens change system design. At $0.30 per million input tokens and $2.50 per million output tokens, you can afford to put more bounded tasks behind a model. You still cannot afford silent extraction errors, malformed tool arguments, or an autonomous loop that keeps calling tools.
Route by failure shape, not by prompt length
A short request can be a dangerous decision, and a long document can be a routine extraction. The useful dividing line is whether failure is easy to detect. Schema validation can catch a missing invoice field. A human can approve a drafted tag before it reaches a customer. An open-ended compliance judgment has no equally cheap guardrail.
| Workload | Starting choice | Required guardrail |
|---|---|---|
| Document fields to strict JSON | Good first candidate | Schema validation plus a labeled holdout set |
| Subagent with one or two tools | Test at minimal thinking | Tool argument checks, turn cap, and stop condition |
| High-volume classification | Good candidate | Per-class error thresholds and fallback routing |
| Long autonomous investigation | Use a stronger baseline too | Full task-success scoring and cost ceiling |
| Legal, medical, or account action | Do not automate unattended | Qualified human approval and source checks |
Minimal thinking is a routing control
Google sets the default thinking level to minimal. That is the right starting point for parsing, routing, and classification because it protects throughput. For subagents that plan several calls, Google recommends trying medium or high so the model does not stop too early.
Do not raise thinking globally. Split your evaluation by task class. If medium thinking improves the hard subagent cases but adds cost to deterministic extraction, keep two routes. One model ID can still serve different policies.
Those two results support Google's claim that this is not merely a renamed cheap tier, but they do not certify your extraction schema or your tool stack. Build the acceptance test around the exact failure you pay for.
Cost the completed job
The model accepts text, images, video, audio, and PDF, and returns text. Its input limit is 1,048,576 tokens and its output limit is 65,536. Large capacity is useful for document batches, but resending a giant conversation history can erase much of the price advantage.
Track four values per completed job: total input, total output including thinking, tool-call count, and retries. Then add review time and fallback-model cost. The calculator handles token scenarios, but your production logs must supply the real request shape.
The migration contract can fail before inference
Google documents two changes that deserve a request audit. Sampling controls such as temperature, top_p, and top_k are deprecated for this generation, and a conversation whose last non-empty turn is a prefilled model turn returns a 400 error. Remove unsupported assumptions before comparing output.
Find sampling controls, prefills, tool schemas, and thinking settings.
Reject malformed JSON and invalid tool arguments automatically.
Send low-confidence or high-risk items to a stronger model or a person.
Multimodal input still needs a provenance contract
Flash-Lite can accept PDFs, images, audio, and video, but its response is text. Decide what that text must point back to before you scale the queue. A document extraction should retain the file and page for every critical field. An audio summary should retain the source clip or timestamp. A video classification should keep the frame range that justified the label. Without those links, a reviewer cannot distinguish a model error from a bad file, OCR problem, or missing segment.
Use three output states instead of forcing every request into success: valid, needs review, and failed validation. Route malformed JSON and missing required fields to the last state automatically. Send ambiguous but well-formed cases to review. Only the first state should continue to a downstream system without a person. That routing discipline is what turns low token price into safe volume.
Who should skip it
Skip Flash-Lite as the only model for a new autonomous agent whose failure modes are still unknown. Skip it for tasks where a persuasive wrong answer is expensive and no reviewer checks the result. Also skip a migration driven only by input price; a stronger model may finish in fewer turns and cost less per successful job.