gemini-3.5-flash-lite
Meant to excel at
-
Google positions it as a low-latency, minimal-cost model optimized for coding, precise document reading and small agent pipelines.
Vendor's own site docs.cloud.google.com -
Cheapest current Gemini ($0.30/$2.50 per 1M) with very high throughput (about 490 tokens/s) and a 1M-token window.
Independent evaluator blog.buildfastwithai.com -
Google-reported SWE-Bench Pro 54.2% and OSWorld-Verified 74.0%, large jumps over 3.1 Flash-Lite.
Vendor claim, relayed blog.buildfastwithai.com -
Good fit for classification, extraction, routing, translation and agentic search in high-volume or batch pipelines.
Independent evaluator geotoolbox.ai -
Strong on OCR transcription (87.5% at low effort) and very fast per image, about 3.9s per sample.
Independent evaluator playground.roboflow.com
Watch out for
-
Slow first token (about 11.7s per Artificial Analysis), so interactive chat feels sluggish.
Independent evaluator blog.buildfastwithai.com -
Capability ceiling: intelligence index 36 versus 50 for 3.6 Flash; long-context MRCR 72.2% versus 91.8%.
Independent evaluator blog.buildfastwithai.com -
Weak at counting in images (52.7%, rank 44 of 63); reasoning needs high effort at 3 to 5x cost.
Independent evaluator playground.roboflow.com -
Poor fit for high-stakes legal, medical or financial calls, long-document reasoning and frontier coding; escalate quality-critical steps.
Independent evaluator blog.buildfastwithai.com
Benchmarks
Agentic coding percentile 0.06
| Benchmark | Score | Percentile | Source | Date |
|---|---|---|---|---|
| swe-bench pro | 54.2 | 0.13 | Vendor claim, relayed datacamp.com | 2026-07-21 |
| terminal-bench 2.1 | 54 | 0.00 | Vendor claim, relayed datacamp.com | 2026-07-21 |
Computer use percentile 0.50
| Benchmark | Score | Percentile | Source | Date |
|---|---|---|---|---|
| osworld-verified | 74 | 0.50 | Vendor claim, relayed datacamp.com | 2026-07-21 |
Overall percentile 0.00
| Benchmark | Score | Percentile | Source | Date |
|---|---|---|---|---|
| artificial analysis intelligence index | 22 | 0.00 | Independent evaluator artificialanalysis.ai | 2026-10-19 |
| gdpval-aa v2 | 1140 | 0.00 | Vendor claim, relayed datacamp.com | 2026-07-21 |
Reasoning percentile 0.00
| Benchmark | Score | Percentile | Source | Date |
|---|---|---|---|---|
| humanity's last exam | 18 | 0.00 | Vendor's own site ai.google.dev | 2026-10-13 |
Percentile is the share of tracked models this model beats on that benchmark, averaged across the group. Only benchmarks at least four tracked models share are scored.
Practitioner reports
- +/-
An analysis positions Flash-Lite as a subagent tier: at $0.30/$2.50 per 1M and about 350 tok/s it makes fan-out cheap, but failed or escalated calls erode savings and cost figures are the author's own arithmetic.
Task: agentic-long-running | Compared with gemini-3.6-flash | 2026-07-22 older than 90 days
Practitioner report digitalapplied.com - +
Author estimates a million small documents costs about $1,100 on Flash-Lite versus $4,500 on 3.6 Flash, and recommends it for first-pass filtering with escalation to a stronger model.
Compared with gemini-3.6-flash | 2026-07-22 older than 90 days
Practitioner report blog.buildfastwithai.com - +/-
Across five vision tasks it averaged 61.7% (rank 33 of 63) but was cheap and third fastest; OCR transcription was strong while counting was its weakest area.
undated
Practitioner report playground.roboflow.com - +/-
Reviewer warns that first-token latency over 11 seconds and weaker reasoning make it a bad choice for live chat or work needing cleanup.
Compared with gemini-3.6-flash undated
Practitioner report geotoolbox.ai