Gemini 3.5 Flash vs Gemini 3.1 Flash-Lite: Cost and API Migration
Keep Gemini 3.1 Flash-Lite for routine work that passes validation. An old 3.5 Flash API configuration now serves 3.6 Flash; compare current costs before switching.
On this page

For routine extraction, translation, classification, and short summaries, start with Gemini 3.1 Flash-Lite and keep it when its answers pass your checks. For coding agents and demanding tool workflows, evaluate the current Flash route. The important change is that Google's Developer API now sends requests for gemini-3.5-flash to gemini-3.6-flash. You are no longer choosing between two independently served versions of the original pair. Google's 3.5 Flash model page documents that routing change.
This comparison uses the Gemini Developer API documentation checked on October 13, 2026. It explains how to preserve a useful low-cost route, identify what an old Flash configuration actually requests, and decide whether paying more produces a better finished result. Recommendations below follow the documented capabilities and pricing; they are not results from our own model or billing tests.
Which route should you choose?
| Your workload | Start with | What would justify changing it? |
|---|---|---|
| Repeated translation, ticket classification, simple extraction | gemini-3.1-flash-lite | Errors require enough retries or human correction to erase its cost advantage. |
| Code changes and agents that must coordinate several tools | gemini-3.6-flash | Keep it if the completed work meets your requirements at an acceptable total cost. |
| Long-running software engineering or autonomous enterprise workflows | Evaluate gemini-3.8-flash as another Flash candidate | Its results on your workload justify a migration from your existing route. |
An application configured with gemini-3.5-flash | Inspect the returned version, then make the replacement explicit | Your configuration and logs should describe the current route instead of implying the old model still runs independently. |
Google positions 3.1 Flash-Lite for lightweight, high-frequency work, 3.6 Flash for coding and rapid agentic loops, and 3.8 Flash for longer software engineering and autonomous workflows. Those descriptions identify sensible candidates. They do not establish a universal quality or speed winner.
A useful first split is whether you can verify the output cheaply. A ticket label can be checked against a short set of categories. A code change may require tests, review, and another debugging cycle. The latter can justify a more expensive call, but only when the resulting work improves. A larger model that makes the same mistake simply costs more.
Does Gemini 3.5 Flash still work in the API?
Its Developer API endpoint is not shutting down under the current published schedule. Google lists gemini-3.5-flash as deprecated, removes it from AI Studio's model selector, and automatically routes API requests to gemini-3.6-flash. The deprecations table explicitly marks its shutdown date as N/A and explains the alias behavior. Deprecated does not mean this particular endpoint now returns a shutdown error.
For new configurations, name gemini-3.6-flash directly, or evaluate gemini-3.8-flash if its intended workload fits. Do not benchmark gemini-3.5-flash against gemini-3.6-flash as though those request strings guarantee two different models.
The Lite side has a different lifecycle:
| Identifier | Current documented status |
|---|---|
gemini-3.1-flash-lite | Stable; released May 7, 2026. Earliest possible shutdown is May 7, 2027. |
gemini-3.1-flash-lite-preview | The separate March preview was shut down on May 25, 2026. |
gemini-3.5-flash-lite | A separate stable Lite model and Google's recommended replacement for 3.1 Flash-Lite. It is not the old 3.5 Flash alias. |
These dates and replacements come from Google's lifecycle schedule. An earliest shutdown date is a planning boundary, not an availability guarantee or service-level agreement. If 3.1 Flash-Lite still meets your needs, there is time to evaluate its replacement deliberately rather than treating the closed preview as evidence that stable Lite is unavailable.
This routing statement applies to Google's Developer API. A third-party gateway or a Gemini app label may have its own routing and prices. Google's alias documentation alone cannot identify that provider's backend or your invoice.
What capabilities actually differ?
The current 3.1 Flash-Lite, 3.6 Flash, and 3.8 Flash pages list many of the same features. Their limits and output types are worth separating from their workload positioning:
| Capability | 3.1 Flash-Lite | Current Flash candidates: 3.6 / 3.8 |
|---|---|---|
| Input | Text, images, video, audio, PDF | Text, images, video, audio, PDF |
| Output | Text | Text |
| Token limits | 1,048,576 input; 65,536 output | 1,048,576 input; 65,536 output |
| Structured output, function calling, code execution | Supported | Supported |
| Caching, file search, URL context, Search and Maps grounding | Supported | Supported |
| Batch, Flex, Priority | Supported | Supported |
| Computer Use | Not supported | Supported in Preview |
| Image generation, audio generation, Live API | Not supported | Not supported |
A million-token input window does not mean a million-token response. For a long translation or report, divide the output into manageable sections and check completeness. Likewise, accepting an image or audio recording does not turn a text-output model into an image generator or real-time voice model.
Thinking configuration also matters. Google's thinking guide gives 3.6 and 3.8 Flash a medium default. The 3.6 model supports minimal, low, medium, and high; 3.8 supports only low, medium, and high. Passing minimal to 3.8 returns an error. The 3.1 Flash-Lite model guide demonstrates high, but the current general thinking table does not specify its default. Avoid assuming Lite has thinking disabled when estimating latency or cost.
Current API prices: model, mode, and date
Paid prices below are USD per 1 million tokens, from Google's pricing page, checked October 13, 2026. Output includes thinking. The Flash columns quote the direct 3.6 and 3.8 models, not an observed bill for the legacy 3.5 alias.
| Mode | 3.1 Lite input: text / image / video | 3.1 Lite audio input | 3.1 Lite output | Direct 3.6 / 3.8 input through Dec. 31, 2026 | Direct 3.6 / 3.8 output through Dec. 31, 2026 |
|---|---|---|---|---|---|
| Standard | $0.25 | $0.50 | $1.50 | $0.75 | $3.75 |
| Batch | $0.125 | $0.25 | $0.75 | $0.375 | $1.875 |
| Flex | $0.125 | $0.25 | $0.75 | $0.375 | $1.875 |
| Priority | $0.45 | $0.90 | $2.70 | $1.35 | $6.75 |
From January 1, 2027, the published direct 3.6/3.8 rates double: Standard becomes $1.50/$7.50, Batch and Flex $0.75/$3.75, and Priority $2.70/$13.50. The old blanket “six times cheaper” comparison no longer describes these current rates.
Free access also varies by model and mode. The current table marks 3.1 Lite input/output free of charge across these modes, while 3.6/3.8 Batch and Flex are unavailable on the Free tier. A free price column does not promise that your project's quota or access will support the workload. Check the exact model and mode on the pricing page and your active limits in AI Studio.
Budget the finished task, not just the visible answer

Consider an illustrative text workload of 20,000 tasks, each using 1,800 input tokens and 500 generated tokens, including any thinking. With uncached Standard calls at the current paid rates:
Input: 20,000 × 1,800 = 36 million tokens
Generated: 20,000 × 500 = 10 million tokens
3.1 Flash-Lite: 36 × $0.25 + 10 × $1.50 = $24.00
Direct Flash: 36 × $0.75 + 10 × $3.75 = $64.50This is arithmetic under stated assumptions, not a measured workload, bill, or guaranteed task allowance. It excludes caching, tools, embeddings, tax, retries, fallback, and human review. Using the published January Standard rates, the same assumed direct Flash token volume would cost $129.
The current $40.50 premium is the amount the more expensive route must justify. At an assumed reviewer cost of $30 per hour, it equals 81 minutes of review time across the entire workload. If Flash saves that much work without introducing new failures, the premium can pay for itself. If both routes produce equally usable results, Lite preserves the lower model cost.
For your actual decision, use:
Cost per accepted task =
(model charges + billed retries/fallback + tools/cache + human review cost)
÷ tasks that meet your acceptance criteriaInclude failed attempts according to your route's actual billing records; an HTTP error alone does not establish whether a charge occurred. Define acceptance first: valid JSON alone may be enough for one classifier, while extraction may also require every field to match its source.
Batch and Flex change how the work runs
Use Batch for independent, non-urgent work that can complete asynchronously. It supports generateContent, targets a 24-hour turnaround, and charges half the equivalent Standard token price. That target is not a completion guarantee. Submit a job once: creation is not idempotent, so submitting it again creates another job. Inspect each result for either a response or a row-level error.
Flex is a Preview synchronous option for latency-tolerant work, including sequential jobs that need the previous answer. It trades its discount for variable latency and best-effort capacity; its 1–15 minute target is not a promise. Capacity can be shed, and requests may return 503 or 429. Google does not automatically upgrade Flex to Standard. If your client retries or falls back, make that decision explicit and include the resulting cost.
Caching needs its own estimate. Cache reads and storage are different units, and storage accumulates over token-hours. There is also an unresolved documentation detail: the Batch guide says cache hits use Standard caching rates, while the model-specific pricing columns list different Batch cache rates. Confirm the applicable treatment before budgeting Batch caching; do not automatically halve every cache-related charge. Batch caching guidance, model-specific prices.
Make the API migration observable
Preserve one low-cost route and one Flash candidate in configuration, then evaluate them on the same useful tasks. Start by making the requested ID explicit; record the returned version separately.

The following is a documentation-based example for the native Developer API, not a call we executed. It requires an existing authorized Gemini API key in GEMINI_API_KEY, curl, and jq. Your account must meet Google's region and age requirements, and the project must have access and quota for the selected model. New AI Studio keys have been auth keys by default since May 28, 2026; unrestricted standard keys are rejected. See Google's key guidance.
Run one chosen model at a time. Change MODEL to gemini-3.6-flash to compare the same input. high is supported by both models here and makes the thinking choice explicit; it is not a recommendation to use high thinking for every ticket.
MODEL='gemini-3.1-flash-lite'
: "${GEMINI_API_KEY:?Set your authorized Gemini API key first}"
cat > request.json <<'JSON'
{
"contents": [{
"role": "user",
"parts": [{"text": "Classify this support ticket as billing, delivery, or product. Return JSON with category and a short reason. Ticket: I paid for express shipping, but the parcel arrived five days late."}]
}],
"generationConfig": {
"responseMimeType": "application/json",
"thinkingConfig": {"thinkingLevel": "high"}
}
}
JSON
curl --fail-with-body --show-error \
"https://generativelanguage.googleapis.com/v1beta/models/${MODEL}:generateContent" \
-H "x-goog-api-key: ${GEMINI_API_KEY}" \
-H 'Content-Type: application/json' \
--data-binary @request.json \
--output "${MODEL}.json"
jq '{modelVersion, responseId, usageMetadata, promptFeedback,
candidates: [.candidates[]? | {finishReason, content}]}' "${MODEL}.json"The generateContent reference documents this endpoint and response fields. Inspect the HTTP result and saved JSON before counting the task as successful. An error, blocked prompt, empty candidate, truncated answer, or incorrect classification is not an accepted result. This ambiguous ticket is useful because an answer can be well-formed yet fail your team's category policy.
Keep the request ID, requested model, returned modelVersion, endpoint, mode, date, configuration, latency, and acceptance outcome. In usageMetadata, promptTokenCount includes cached content; do not add cachedContentTokenCount again. candidatesTokenCount and thoughtsTokenCount describe generated usage, while totalTokenCount also includes the prompt. Do not price that total as if it were all output. UsageMetadata reference.
If your integration uses the newer Interactions API, preserve that API's contract: /v1beta/interactions uses model, input, and generation_config, with its own output and usage fields. The payload and response parser above are for generateContent. The current getting-started guide uses Interactions. Changing a model does not require silently changing the API or moving to Vertex AI. For a separate governance decision, see Gemini Developer API versus Vertex AI.
Before promoting a route, keep the same inputs, retrieval material, tools, supported settings, and acceptance criteria. Include the failure cases that matter to your product. Check that gains survive the full workflow, and keep a rollback configuration. There is no fixed number of evaluation rounds that makes a migration safe.
FAQ
Is Gemini Flash or Flash-Lite faster?
3.1 Flash-Lite is designed for low-latency, lightweight work, but that does not establish a speed ratio for your requests. Thinking level, input size, service mode, and tool calls affect the result. Measure time to the first usable response and time to an accepted result under comparable settings. An old 3.5 Flash request now routes to 3.6 in Google's Developer API, so historical 3.5 speed charts describe a different comparison. Lite model guide, 3.5 alias notice.
Is Gemini 3.5 Flash-Lite the same as Gemini 3.5 Flash?
No. gemini-3.5-flash-lite is a separate stable model for low-cost, high-throughput work and a recommended migration option for 3.1 Lite. gemini-3.5-flash is the deprecated Flash identifier that automatically routes to 3.6 Flash. The word “Lite” changes which model you request. 3.5 Flash-Lite model page, deprecation schedule.
Can either route generate images or run a live voice conversation?
No. The text models compared here can understand multimodal input and return text, but their current cards do not support image generation, audio generation, or the Live API. Choose a model that explicitly supports the output or runtime you need. 3.1 Lite capabilities, 3.6 Flash capabilities.
Will more API keys let Flash-Lite process more requests?
Not within the same project. Gemini limits apply per project, with RPM, input TPM, and RPD all enforced; daily requests reset at midnight Pacific time. Active numbers are shown in AI Studio and capacity is not guaranteed. A 429 can reflect quota or spend limits, so inspect those before interpreting it as a model failure. Google's rate-limit guide.
Should I replace Flash-Lite with Flash everywhere?
Keep Lite where results are already correct and cheap to verify. Use explicit current Flash IDs for the workloads where better completed results justify their extra cost. Route by observed failures and accepted-task cost, and preserve a fallback that you have evaluated for that task. A release name alone is not a reason to replace a working pipeline.
Sources15
External pages this guide links to, in the order they appear. Last updated Oct 13, 2026.
Sources15
External pages this guide links to, in the order they appear. Last updated Oct 13, 2026.
- 1.Google's 3.5 Flash model pageai.google.dev/gemini-api/docs/models/gemini-3.5-flash
- 2.3.1 Flash-Liteai.google.dev/gemini-api/docs/models/gemini-3.1-flash-lite
- 3.3.6 Flashai.google.dev/gemini-api/docs/models/gemini-3.6-flash
- 4.3.8 Flashai.google.dev/gemini-api/docs/models/gemini-3.8-flash
- 5.deprecations tableai.google.dev/gemini-api/docs/deprecations
- 6.Google's thinking guideai.google.dev/gemini-api/docs/thinking
- 7.Google's pricing pageai.google.dev/gemini-api/docs/pricing
- 8.Batchai.google.dev/gemini-api/docs/batch-api
- 9.Flexai.google.dev/gemini-api/docs/flex-inference
- 10.region and age requirementsai.google.dev/gemini-api/docs/available-regions
- 11.Google's key guidanceai.google.dev/gemini-api/docs/api-key
- 12.generateContent referenceai.google.dev/api/generate-content
- 13.current getting-started guideai.google.dev/gemini-api/docs/get-started
- 14.3.5 Flash-Lite model pageai.google.dev/gemini-api/docs/models/gemini-3.5-flash-lite
- 15.Google's rate-limit guideai.google.dev/gemini-api/docs/rate-limits





