by Google · Gemini 3.5 family · best for high-volume subagents and document parsing
Gemini 3.5 Flash-Lite is Google's newest Flash-Lite model and the successor to Gemini 3.1 Flash-Lite, released as a stable GA model on 2026-07-21 in the Gemini API, Google AI Studio and Gemini Enterprise Agent Platform. Google built it as the fast worker in a two-tier agent stack: a Flash model plans, Flash-Lite runs the high-volume subtasks, document parsing and extraction. On Google's own evaluations it is a large step up from 3.1 Flash-Lite in coding and agent work (Terminal-Bench 2.1 54.0% vs 31.0%, SWE-Bench Pro 54.2% vs 38.3%) and it keeps the 1M-token window and native audio, video and image input. For a buyer: it now handles agent work that needed a Flash model last generation, at $0.30 in / $2.50 out per 1M tokens, but it costs more than the model it replaces and nearly all of its capability numbers come from Google.
gemini-3.1-flash-lite, whose earliest shutdown date is 2027-05-07.| Benchmark | Score | Source |
|---|---|---|
| LMArena Elo | 1456 | lmarena.ai 2026-09-13T00:00:00.000Z |
| Terminal-Bench | 54% | deepmind.google 2026-07-21T00:00:00.000Z |
| MRCR Long Context | 72.2% | deepmind.google 2026-07-21T00:00:00.000Z |
| LMArena Coding Elo | 1504 | lmarena.ai 2026-09-13T00:00:00.000Z |
Six personas, six verdicts — the same panel that reviews every product on TopReviewed.
“A GA, named successor that matches GPT-5.4 mini on SWE-Bench Pro at 40% of the input price. An easy default for subagent traffic on Google.”
For a CTO, 3.5 Flash-Lite is the low-risk default for high-volume agent traffic on Google Cloud. It is GA, Google names it the replacement for 3.1 Flash-Lite, and Agent Platform lists its retirement as no earlier than July 21, 2027. On Google's card it posts SWE-Bench Pro 54.2% and OSWorld-Verified 74.0%, roughly GPT-5.4 mini territory at $0.30 / $2.50. Governance matches the rest of Gemini: CMEK, VPC Service Controls and US or EU residency. The risks are cadence and cost drift: Google shipped four Flash-tier models between July and September 2026, and this generation's output price is 67% above the model it replaces.
“Google has turned Flash-Lite from a cheap classifier into a real subagent tier, doing work that needed a full Flash model last generation.”
In market terms 3.5 Flash-Lite moves the Lite tier up a class. Google's pitch is a two-tier agent stack, with a Flash model orchestrating and Flash-Lite doing the cheap, fast work, and its published numbers support it: it beats Gemini 3 Flash on SWE-Bench Pro (54.2% vs 49.6%) and OSWorld-Verified (74.0% vs 65.1%). Its moat is breadth in one cheap SKU: audio and video input, Search and Maps grounding, computer use and a 1M window. The positioning is muddied by Google's own pricing, since 3.6 to 3.8 Flash sit at an introductory $0.75 / $3.75 until the end of 2026, and by cheaper text-first rivals such as GPT-5.6 Luna and Mistral Small 4.
“$0.30 in and $2.50 out with half-price Batch and Flex is still cheap, but output costs 67% more than 3.1 Flash-Lite. Rerun the unit math.”
At $0.30 input and $2.50 output, thinking tokens included, with cache reads at $0.03 and Batch or Flex at $0.15 / $1.25, 3.5 Flash-Lite is one of the cheapest ways to buy this level of agent capability. Audio now bills at the text rate, a 40% cut from 3.1 Flash-Lite's $0.50. The catch is output: $2.50 is 67% above the predecessor and roughly double GPT-5.6 Luna ($1.20) or DeepSeek V4.1-Flash at peak ($1.20), and until December 31, 2026 Gemini 3.8 Flash costs only 1.5 times as much per output token. The default minimal thinking level keeps output lean; the high level Google used for its benchmarks will raise spend.
“Computer use, code execution, file search and Search grounding in a Lite model is great. Set thinking to medium before trusting tool loops.”
For builders, 3.5 Flash-Lite exposes nearly the full Gemini tool surface: function calling, structured outputs, code execution, URL context, file search, Search and Maps grounding, and computer use in preview, plus Batch, Flex and Priority consumption. Google Cloud's docs say plainly that the default minimal thinking level can end multi-step tool runs early, so subagents should run at medium or high. Migration has sharp edges: temperature, top-K and top-P are ignored, frequency and presence penalties throw errors, and requests ending on a model turn fail. Once those are handled, it is a model-ID swap across the Gemini API, the Interactions API and Agent Platform.
“Quick, and clearly smarter than older Lite models. On LMArena it sits level with GPT-5.6 Luna, though it is still no deep-thinking model.”
Most people will meet 3.5 Flash-Lite indirectly: Google lists the Gemini app as a distribution channel and is rolling the model out in Google Search. Where users compare it head to head, it holds up: LMArena's text leaderboard rates it 1456 as of September 13, 2026, about 24 points above 3.1 Flash-Lite and level with GPT-5.6 Luna. Google quotes 350 output tokens per second from Artificial Analysis; treat that as a vendor-relayed figure until you time your own calls. The March 2026 cutoff is fresher than most peers, though Google warns some domains stop at January 2025. For hard reasoning, step up to a Flash model.
“Every agentic number is Google-run at high thinking while the default is minimal, and recall at 1M tokens is about 20%. Test before migrating.”
The direction of the upgrade is real: two independent boards, LMArena and Context Arena, both place it clearly above 3.1 Flash-Lite. But every coding and agentic score on the card is Google-computed, on Google's harnesses, at high thinking, while the model ships defaulting to minimal. Context Arena shows the cost of that gap: recall to 128K falls from 64.4% AUC at high thinking to 24.2% at minimal. At the full 1M window it recalls about 20%. The 350 tokens-per-second speed claim reaches buyers only through Google's blog, the knowledge cutoff is only partly March 2026, unjustified refusals regressed 5.32 points on Google's own eval, and the price went up.
The full research notes behind this review — verified against primary sources.
Google's model card says 3.5 Flash-Lite is based on Gemini 3.1 Flash-Lite, which in turn is based on Gemini 3 Pro, and Google describes Gemini 3 Pro as a sparse mixture-of-experts transformer with native multimodal input. That chain is the basis for architecture_type: moe. Parameter counts, expert counts, layers, attention type, tokenizer and training scale are undisclosed and left null. Disclosed facts: a 1,048,576-token input window, 65,536 output tokens, native input of text, images, audio, video and PDFs, and adjustable thinking levels. Google Cloud documents per-request limits of up to 3,000 images, 3,000 PDF pages per file, roughly 45 minutes of video with audio (about an hour without) and about 8.4 hours of audio.
The headline is agent work (cap_agentic 8.5): OSWorld-Verified 74.0% edges GPT-5.4 mini (72.1%) and far exceeds Claude Haiku 4.5 (50.7%) on Google's card, Terminal-Bench 2.1 54.0% trails GPT-5.4 mini (59.2%) but beats Haiku (44.2%), and GDPval-AA v2 1140 sits near GPT-5.4 mini (1171). Function calling and built-in tools (cap_function_calling 8.5) cover computer use, code execution, file search, URL context and structured outputs. Coding (cap_coding 8.0) rests on SWE-Bench Pro 54.2%, level with GPT-5.4 mini (54.4%) and well ahead of Haiku 4.5 (39.5%). Long context (cap_long_context 8.0) is strong to 128K (72.2% Google, 64.4% AUC on Context Arena) and weak at the full 1M (about 20%). Vision (cap_vision 8.5) is steady rather than transformed: CharXiv Reasoning 74.5% vs 73.2% for 3.1 Flash-Lite, below GPT-5.4 mini's 80.3%. Document understanding (cap_document_ocr 8.5) is a stated design target: Google's API page names document parsing, and Google Cloud describes the model as optimized for precise document understanding. Reasoning (7.5) and math (7.0) carry over from 3.1 Flash-Lite's level because Google published no GPQA, HLE or AIME figures for this model. Multilingual (8.0) reflects Google's translation positioning without a new public multilingual score. Instruction following (8.0) and creative writing (7.0) are supported by its LMArena text rating of 1456. Safety calibration (7.5) reflects fewer policy violations but more unjustified refusals on Google's automated evals. Real-time data (cap_realtime_data 9.0) comes from built-in Google Search and Maps grounding.
| Benchmark | Score | vs Predecessor | vs Top Competitor | Source |
|---|---|---|---|---|
| SWE-Bench Pro (Public) | 54.2% | +15.9 pts vs 3.1 Flash-Lite (38.3%) | GPT-5.4 mini 54.4%; Claude Haiku 4.5 39.5% | model card |
| Terminal-Bench 2.1 (Terminus-2) | 54.0% | +23.0 pts (31.0%) | GPT-5.4 mini 59.2%; Claude Haiku 4.5 44.2% | model card |
| MLE-Bench (Partial 30) | 39.2% | +17.2 pts (22.0%) | No competitor figure published | model card |
| GDPval-AA v2 (Elo) | 1140 | +498 (642) | GPT-5.4 mini 1171; Claude Haiku 4.5 907 | model card |
| OSWorld-Verified | 74.0% | +19.7 pts (54.3%) | Ahead of GPT-5.4 mini 72.1% and Claude Haiku 4.5 50.7% | model card |
| CharXiv Reasoning (no tools / with tools) | 74.5% / 76.5% | +1.3 / +0.9 pts | GPT-5.4 mini 80.3% leads; Claude Haiku 4.5 61.7% | model card |
| GDM-MRCR v2 8-needle (128K avg / 1M) | 72.2% / 21.3% | +12.1 / +9.0 pts | GPT-5.4 mini 42.7%; Claude Haiku 4.5 35.3% at 128K | model card |
| Context Arena MRCR v2 8-needle, high thinking (AUC to 128K / at 1M) | 64.4% / 19.8% | 3.1 Flash-Lite Preview (high): 48.0% / 12.4% | GPT-5.6 Luna (max) 74.8% to 128K | Context Arena |
| LMArena Text Elo | 1456 ±5 | +24 vs 3.1 Flash-Lite Preview (1432) | GPT-5.6 Luna (xhigh) 1452; Claude Haiku 4.5 1415 | LMArena |
All model-card scores are vendor-reported and Google-computed, run at high thinking with default sampling as of July 2026. Competitor figures in those rows are as reported on Google's card, and the GDPval-AA v2 figures are Artificial Analysis leaderboard numbers relayed by Google. The Context Arena and LMArena rows are independent (LMArena as of 2026-09-13, 26,165 votes). GPQA Diamond, Humanity's Last Exam, MMMU-Pro, AIME 2025 and SWE-bench Verified are not published for this model and are left null.
Google calls 3.5 Flash-Lite the fastest model in the 3.5 series and quotes 350 output tokens per second, a figure it attributes to Artificial Analysis. That number is vendor-relayed, so output_speed_tps and time-to-first-token stay null until an independent measurement is published. Speed depends heavily on the thinking level: the default minimal level is tuned for fast classification, routing and JSON extraction, while medium and high add reasoning tokens before the answer. For latency-critical paths Google sells a Priority tier ($0.54 / $4.50); for work that can wait, the Flex and Batch tiers halve the price.
| Surface | Cost | Notes |
|---|---|---|
| API input | $0.30 / 1M tok | Same rate for text, image, video and audio |
| API output | $2.50 / 1M tok | Thinking tokens billed as output |
| Cached input | $0.03 / 1M tok | 90% off input; plus $1.00 / 1M tokens per hour of storage |
| Batch (in/out) | $0.15 / $1.25 | Cached input $0.02 |
| Flex (in/out) | $0.15 / $1.25 | Cheaper, lower-priority interactive tier |
| Priority (in/out) | $0.54 / $4.50 | Cached input $0.05; for latency-sensitive traffic |
| Search and Maps grounding | 5,000 requests per month free each (shared across Gemini 3.x), then $14 / 1,000 | Paid tier only; free tier can test in AI Studio |
| Direct UI | No named plan | Google lists the Gemini app and Google Search as channels but does not name a plan that selects this model |
| Free tier | Input and output free of charge | No caching; free-tier content may be used to improve Google products |
| Rate limits | Per project, shown in AI Studio | Batch enqueued tokens: 10M (Tier 1), 500M (Tier 2), 1B (Tier 3) |
Proprietary and closed-weights. It is available through the Gemini API (Google AI Studio, including the Interactions and GenerateContent APIs), Android Studio, and Gemini Enterprise Agent Platform, Google Cloud's managed platform formerly branded Vertex AI. On Agent Platform it runs in the global endpoint and the US and EU multi-regions, with Standard, Flex and Priority pay-as-you-go, Provisioned Throughput and batch inference, and data residency, CMEK, VPC Service Controls and Access Transparency for online, batch and cached traffic. Google Cloud lists tuning as not supported. OpenRouter resells it as google/gemini-3.5-flash-lite. Migration from 3.1 Flash-Lite is a model-ID change plus the Gemini 3.5-generation API changes (no sampling or penalty parameters, no trailing model turn).
Google's model card reports that, against 3.1 Flash-Lite, automated text-to-text policy violations fell 8.14 points and multilingual violations 0.92 points, with image-to-text unchanged. Unjustified refusals rose 5.32 points, which Google attributes to more caution on sensitive topics. For Frontier Safety, Google relies on its Gemini 3.1 Pro assessment, which reached no Critical Capability Levels, and states that 3.5 Flash-Lite adds no meaningful new capability in those domains. Google's pricing page marks paid-tier Gemini API content as not used to improve its products, and free-tier content as used; Agent Platform traffic falls under the Google Cloud Platform terms. Google Cloud's platform certifications (SOC 2, ISO 27001, HIPAA, FedRAMP, GDPR, CCPA) apply at the platform level, so confirm this model's audit scope with Google before relying on a specific attestation. Built-in safety filters are configurable.
Official Google GenAI SDKs cover Python, JavaScript/TypeScript, Go and Java, and the model is also reachable through Google's OpenAI-compatible chat-completions surface on Agent Platform. Frameworks that target the Gemini API, including Google ADK, Genkit, LangChain, LlamaIndex and the Vercel AI SDK, use it by model ID. Google is rolling it out in Google Search and lists the Gemini app as a channel, and OpenRouter resells it. It had about 26,000 LMArena votes by mid-September 2026, and Google is building it into its own products.
Google repriced the tier to $0.30 / $2.50 from $0.25 / $1.50 alongside large agentic gains (Terminal-Bench 2.1 54.0% vs 31.0%). Output-heavy workloads feel the 67% output increase most; audio-heavy workloads get cheaper because audio input drops from $0.50 to $0.30. Model the change on your own input-to-output ratio.
Until December 31, 2026, 3.8 Flash costs $0.75 / $3.75, only 2.5 times Flash-Lite's input and 1.5 times its output. For hard, multi-step work that gap is small, so test 3.8 Flash first. Flash-Lite wins on high-volume, input-heavy and latency-sensitive calls, and 3.8 Flash rises to $1.50 / $7.50 on January 1, 2027.
Google Cloud documents minimal as the default and recommends it for classification, routing and JSON extraction. For subagents that write code, run terminal commands or call APIs, Google recommends medium or high, because minimal can end multi-step tool runs early. Google's published benchmarks were run at high.
Only partly. Recall is strong to 128K (72.2% on Google's MRCR v2 run, 64.4% AUC on Context Arena) but falls to about 20% at 1M on both. Chunk or retrieve for precise lookups in very long inputs.
Custom temperature, top-K and top-P values are ignored, frequency and presence penalties return an error, and requests whose last turn is a model turn fail. Beyond that, the switch is a model-ID change.
Not on the paid tier: Google's pricing page marks paid-tier Gemini API content as not used to improve its products. Free-tier content is used, so keep sensitive data off the free tier. Agent Platform usage is governed by the Google Cloud Platform terms.
Google lists May 7, 2027 as its earliest shutdown date and names 3.5 Flash-Lite as the replacement. 3.5 Flash-Lite has no announced shutdown date in the Gemini API, and Agent Platform lists its retirement as July 21, 2027 or later.
No. Google Cloud lists tuning as not supported for this model, and the weights are closed; access is through the Gemini API, Agent Platform or resellers such as OpenRouter.
Cheaper at $0.20 / $1.20 with a 1.05M window (922K max input) and 128K max output, but text and image input only, and prompts over 272K input tokens bill at 2x input and 1.5x output. The two are level on LMArena text (1452 vs 1456), and Luna's max-effort run recalls better to 128K on Context Arena (74.8% vs 64.4% AUC). Luna wins on price and output length; 3.5 Flash-Lite wins on audio and video input, Search and Maps grounding, and built-in computer use.
Costs over three times as much on input and twice as much on output ($1 / $5) with a 200K window. On Google's card Flash-Lite leads it on SWE-Bench Pro (54.2% vs 39.5%), Terminal-Bench 2.1 (54.0% vs 44.2%) and OSWorld-Verified (74.0% vs 50.7%), and it leads LMArena text by about 40 points (1456 vs 1415). Haiku remains the pick for teams standardised on Claude tooling.
Same $0.30 peak input but less than half the output price ($1.20 peak, $0.60 off-peak), 384K max output and MIT open weights you can self-host. It takes text and image input, with no audio, video or Google-style grounding. Pick DeepSeek for text-heavy volume where output tokens dominate the bill; pick 3.5 Flash-Lite for multimodal, grounded work under Google Cloud's US and EU residency controls.
Primary references used to verify this review.
Does not train on API inputs by default
Last verified 2026-09-21