Gemini 3.5 Flash-Lite: A Subagent Model Worth Its Price Rise

The cheapest tier of a model family is usually judged on one question: is it good enough to stop paying for the next tier up? For most of the Lite line's history the answer was "for classification, yes; for anything agentic, no". Gemini 3.5 Flash-Lite is the first release where that second answer needs re-checking.

Google released it on 21 July 2026, the same day as Gemini 3.6 Flash. It is a stable model, not a preview, and Google positions it for high-throughput subagent work, document parsing, translation and classification.

Cheap, but not as cheap as the last one

Google lists 3.5 Flash-Lite at $0.30 per million input tokens and $2.50 per million output, with cached input at $0.03. Its predecessor, Gemini 3.1 Flash-Lite, is $0.25 and $1.50. Input went up by a fifth; output went up by two thirds.

That matters more than it looks because Lite workloads tend to be output-light and volume-heavy. A classifier that returns a label pays almost entirely for input, so the increase is modest. An extraction job that writes a JSON record per document pays for output, and that bill rises noticeably.

The :batch variant listed on OpenRouter is the same model through Google's asynchronous Batch API, at half price: $0.15 input and $1.25 output. For backfills, nightly tagging runs and anything else nobody is waiting on, that discount is the first lever to pull. Batch API savings math covers when the latency trade is worth it.

The jump over 3.1 Flash-Lite is large

The model card says 3.5 Flash-Lite is based on 3.1 Flash-Lite, and its published comparison is unusually lopsided for a point release:

• Terminal-bench 2.1: 54.0%, up from 31.0%
• SWE-Bench Pro (public): 54.2%, up from 38.3%
• OSWorld-Verified: 74.0%, up from 54.3%
• GDPVal-AA v2: 1140 Elo, up from 642
• MRCR v2 at 128K: 72.2%, up from 60.1%

Chart reasoning barely moved: CharXiv without tools went from 73.2% to 74.5%. The gains are concentrated in tool use, terminal work and multi-step tasks, which is exactly where earlier Lite models were not trustworthy.

Google also reports that it beats the older and larger Gemini 3 Flash on SWE-Bench Pro (54.2% against 49.6%) and OSWorld-Verified (74.0% against 65.1%). Artificial Analysis measured it at around 350 output tokens per second, according to DataCamp's launch coverage, which makes it the fastest model in the 3.5 generation.

Against other small models

The model card compares it with OpenAI's GPT-5.4 mini and Anthropic's Claude Haiku 4.5, and lists their prices as $0.75/$4.50 and $1.00/$5.00 respectively. On Google's numbers, 3.5 Flash-Lite roughly ties GPT-5.4 mini on SWE-Bench Pro (54.2% against 54.4%), trails it on Terminal-bench 2.1 (54.0% against 59.2%), chart reasoning and GDPVal-AA, and leads it on OSWorld-Verified and long-context retrieval. Against Haiku 4.5 it leads on every row Google published.

These are vendor tables, so expect the margins to shrink on an independent harness. But a model that is competitive with GPT-5.4 mini at well under half the output price is worth a slot in any evaluation.

Flash-Lite or 3.6 Flash: a routing decision

The sibling to compare it with is Gemini 3.6 Flash, released the same day. Google currently lists 3.6 Flash at an introductory $0.75 input and $3.75 output, so Flash-Lite is two and a half times cheaper on input and one and a half times cheaper on output. From 1 January 2027 Google's pricing page says 3.6 Flash doubles to $1.50 and $7.50, which widens the gap considerably.

What you give up is visible in the model cards. On Terminal-bench 2.1, 3.6 Flash scores 78.0% to Flash-Lite's 54.0%. On SWE-Bench Pro it is 58.7% to 54.2%, a much narrower gap. On MRCR at 128K it is 91.8% to 72.2%.

A workable rule:

• Pick Flash-Lite for bounded tasks with a checkable result: classification, extraction into a schema, translation, routing decisions, summarising a single document, and subagents that do one focused job and return.
• Pick 3.6 Flash when the task is open-ended, involves long terminal sessions, or depends on finding details deep inside a large context.
• Use both where volume is high: Flash-Lite first, escalate failures. Model routing and fallbacks has the pattern, and the Gemini 3.6 Flash guide covers the upper tier.

Building a subagent tier around it

Google describes 3.5 Flash-Lite as suited to subagents that execute focused tasks inside larger multi-agent workflows. That is a sensible framing. An orchestrator running on a stronger model hands out narrowly scoped jobs, and a cheap, fast model does them in parallel. Subagents and delegation covers when that split actually helps.

A few specifics worth knowing before you build on it:

• It supports function calling, structured outputs, code execution, search grounding and URL context, plus computer use in preview.
• Thinking is adjustable down to a minimal setting, according to DataCamp's launch coverage. For classification, start at the lowest level and only raise it if accuracy measurably improves.
• It is one of three models Google enabled for agentic video understanding in September, where the model chooses which parts of a video to inspect rather than sampling every second.
• The model card reports a 5.3-point rise in unjustified refusals against 3.1 Flash-Lite. If you run moderation or content-classification pipelines, measure the false-refusal rate on your own data before switching.

The takeaway: 3.5 Flash-Lite is no longer only a classifier. Give it the bounded agentic jobs you were sending to a full Flash model, measure the failure rate, and keep the escalation path for the ones it misses.