Google's budget Flash-Lite tier in the 3.5 line, released 2026-07-21. Built for high-volume, latency-sensitive classification, extraction, and summarisation at a fraction of Flash pricing.
Gemini 3.5 Flash-Lite is Google's budget tier in the 3.5 line, released 21 July 2026 alongside Gemini 3.6 Flash. Google's own one-line framing is "our fastest, most cost-effective 3.5 model for high-throughput execution". It keeps the full 1,048,576-token input window and 65,536-token output limit of its bigger siblings, and supports flexible thinking levels so callers can trade latency against quality per request.
The value case is the price-to-capability jump over the previous budget tier. Against Gemini 3.1 Flash-Lite, Google's model card shows 54.2% vs 38.3% on SWE-Bench Pro and 74.0% vs 54.3% on OSWorld-Verified, for $0.30 per 1M input and $2.50 per 1M output — a fifth of Gemini 3.5 Flash's input rate. On TheRouter it is served over the OpenAI-compatible chat completions API at Google's published Standard-tier rate.
| Type | Rate |
|---|---|
| Input | $0.324 / 1M tokens |
| Output | $2.70 / 1M tokens |
| Release date | 2026-07-21deepmind.google ↗ | verified |
| Knowledge cutoff | March 2026deepmind.google ↗ | verified |
| Input token limit | 1,048,576ai.google.dev ↗ | verified |
| Output token limit | 65,536ai.google.dev ↗ | verified |
| Native modalities (upstream) | Input: text, image, video, audio, PDF. Output: text only. TheRouter exposes text, image, and PDF input with text output.ai.google.dev ↗ | verified |
| Supported capabilities | Thinking (flexible levels), function calling, structured outputs, code execution, context caching, Batch API, Google Search grounding, built-in computer-use tool.ai.google.dev ↗ | verified |
| Not supported upstream | Live APIai.google.dev ↗ | verified |
| Vendor list price (Standard tier) | $0.30 / 1M input (no audio premium), $2.50 / 1M output, $0.03 / 1M cached input, plus $1.00 / 1M tokens per hour cache storage charged by Googleai.google.dev ↗ | verified |
| Output throughput | 350 output tokens/second, per Google citing Artificial Analysisblog.google ↗ | to verify |
| Benchmark | Distribution | Score | Source |
|---|---|---|---|
SWE-Bench Pro Gemini 3.1 Flash-Lite 38.3%; GPT-5.4 mini 54.4%; Claude Haiku 4.5 39.5% | 54.2%% | deepmind.google ↗ | |
Terminal-Bench 2.1 Gemini 3.1 Flash-Lite 31.0%; GPT-5.4 mini 59.2%; Claude Haiku 4.5 44.2% | 54.0%% | deepmind.google ↗ | |
MLE-Bench Gemini 3.1 Flash-Lite 22.0% | 39.2%% | deepmind.google ↗ | |
OSWorld-Verified Gemini 3.1 Flash-Lite 54.3%; GPT-5.4 mini 72.1%; Claude Haiku 4.5 50.7% | 74.0%% | deepmind.google ↗ | |
CharXiv (no tools) Gemini 3.1 Flash-Lite 73.2%; GPT-5.4 mini 80.3%; Claude Haiku 4.5 61.7% | 74.5%% | deepmind.google ↗ | |
GDM-MRCR v2 (128k) Gemini 3.1 Flash-Lite 60.1%; GPT-5.4 mini 42.7%; Claude Haiku 4.5 35.3% | 72.2%% | deepmind.google ↗ | |
GDM-MRCR v2 (1M pointwise) Gemini 3.1 Flash-Lite 12.3% | 21.3%% | deepmind.google ↗ | |
GDPval-AA v2 Gemini 3.1 Flash-Lite 642; GPT-5.4 mini 1171; Claude Haiku 4.5 907 | 1140Elo | deepmind.google ↗ |
Use the global api.therouter.ai endpoint shown below for new integrations; the legacy China accelerated endpoint is retired.
curl https://api.therouter.ai/v1/chat/completions -H "Content-Type: application/json" -H "Authorization: Bearer $THE_ROUTER_API_KEY" -d '{
"model": "google/gemini-3.5-flash-lite",
"messages": [
{"role": "user", "content": "Summarize the key points from this input."}
]
}'Standard OpenAI-compatible chat completion. Thinking is supported with flexible levels; thinking tokens are billed as output.
curl https://api.therouter.ai/v1/chat/completions \
-H "Authorization: Bearer $THEROUTER_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"google/gemini-3.5-flash-lite","messages":[{"role":"user","content":"Classify this ticket: billing or technical?"}]}'Released the same day as Gemini 3.6 Flash, Flash-Lite takes over the budget slot from Gemini 3.1 Flash-Lite with a large capability jump — Google's model card puts it 16 points ahead on SWE-Bench Pro and 23 points ahead on Terminal-Bench 2.1 — at $0.30 / $2.50 per 1M tokens and a claimed 350 output tokens per second.
When volume and latency matter more than the last few points of quality. Flash-Lite is a fifth of the input price and a third of the output price, but Google's own benchmarks put 3.6 Flash 24 points ahead on Terminal-Bench 2.1 and 20 points ahead on GDM-MRCR v2 at 128K. Classification, extraction, and routing are good fits; long-horizon agent runs are not.
Yes — Google's model reference lists a 1,048,576-token input limit, the same as Gemini 3.6 Flash. Quality at extreme length is lower, though: Google reports 21.3% on GDM-MRCR v2 at 1M versus 54.0% for Gemini 3.6 Flash, so treat the full window as available capacity rather than guaranteed recall.
Not upstream: Google's pricing page lists a single $0.30 per 1M input rate covering text, image, video, and audio, unlike Gemini 3.1 Flash-Lite which charged audio separately. TheRouter currently exposes text, image, and PDF input for this model.
| source | URL | retrieved | |
|---|---|---|---|
| Release date | deepmind.google ↗ | 2026-07-28 | verified |
| Knowledge cutoff | deepmind.google ↗ | 2026-07-28 | verified |
| Input token limit | ai.google.dev ↗ | 2026-07-28 | verified |
| Output token limit | ai.google.dev ↗ | 2026-07-28 | verified |
| Native modalities (upstream) | ai.google.dev ↗ | 2026-07-28 | verified |
| Supported capabilities | ai.google.dev ↗ | 2026-07-28 | verified |
| Not supported upstream | ai.google.dev ↗ | 2026-07-28 | verified |
| Vendor list price (Standard tier) | ai.google.dev ↗ | 2026-07-28 | verified |
| Output throughput | blog.google ↗ | 2026-07-28 | to verify |
| SWE-Bench Pro | deepmind.google ↗ | 2026-07-28 | verified |
| Terminal-Bench 2.1 | deepmind.google ↗ | 2026-07-28 | verified |
| MLE-Bench | deepmind.google ↗ | 2026-07-28 | verified |
| OSWorld-Verified | deepmind.google ↗ | 2026-07-28 | verified |
| CharXiv (no tools) | deepmind.google ↗ | 2026-07-28 | verified |
| GDM-MRCR v2 (128k) | deepmind.google ↗ | 2026-07-28 | verified |
| GDM-MRCR v2 (1M pointwise) | deepmind.google ↗ | 2026-07-28 | verified |
| GDPval-AA v2 | deepmind.google ↗ | 2026-07-28 | verified |
| Gemini 3.5 Flash-Lite lands as the new budget tier | blog.google ↗ | 2026-07-28 | verified |
| When should I use Flash-Lite instead of Gemini 3.6 Flash? | deepmind.google ↗ | 2026-07-28 | to verify |
| Does it really have the full 1M-token context window? | deepmind.google ↗ | 2026-07-28 | to verify |
| Is audio input charged at a premium? | ai.google.dev ↗ | 2026-07-28 | to verify |