Short answer: move a repeatable API workflow to Gemini 3.8 Flash only when the workflow needs its newer reasoning and tool behavior, and make the decision with a same-task pilot. The stable model ID is gemini-3.8-flash, the model supports low, medium and high thinking levels, and Google says it can consume more tokens on complex work. A lower token rate or a faster first response does not by itself prove a lower cost per accepted task.
Google made Gemini 3.8 Flash generally available on September 2, 2026. The launch announcement positions it for long-running coding and agent workflows, while the current migration guide gives the implementation changes. This guide turns those changes into a migration and buying decision for teams that run repeatable API work. It does not report an executed API test or an account-level performance result.
What Gemini 3.8 Flash changes
The model page lists gemini-3.8-flash as a stable model with a 1,048,576-token input limit and a 65,536-token output limit. It accepts text, image, video, audio and PDF inputs and returns text. The documented capability set includes function calling, structured outputs, code execution, file search, Search and Maps grounding and caching. Computer use is marked Preview; audio generation, image generation and Live API are not listed as supported capabilities for this model.
| Decision area | Current 3.8 fact | What to verify in your workflow |
|---|---|---|
| Model and context | Stable gemini-3.8-flash; 1,048,576 input and 65,536 output tokens. | Whether your prompts, files and tool results fit the context and remain reviewable. |
| Thinking | low, medium and high; medium is the default. minimal returns an error. | Which level meets the job’s quality and latency requirement without unnecessary token use. |
| Tools | Function calling, code execution, file search, structured outputs and Search/Maps grounding are listed; computer use is Preview. | Account entitlement, tool fees, retries, permissions and the review boundary around each tool. |
| Workflow behavior | Google says complex tasks may use more tokens for iterative reasoning, tool calls and verification. | Tokens, tool calls, loops, retries, reviewer time and accepted outputs per task. |
Keep the API surface clear
The model ID identifies the Gemini API model. It does not make consumer-app access, Google AI Studio behavior, Managed Agents, or Cloud Agent Platform billing interchangeable. Check the exact product, account, region and terms before treating a successful experiment in one surface as evidence for another.
| Surface | Use this guide for | Boundary |
|---|---|---|
| Gemini API | Model ID, Interactions or generateContent integration, token rates, cache and grounding accounting. | Use the API pricing table and model documentation for the account and endpoint you will operate. |
| Google AI Studio or consumer products | Prompt exploration or a product-specific experience. | An app plan, interface or rollout does not prove API quota, tool access or API billing. |
| Cloud Agent Platform | Enterprise procurement on that platform. | The Cloud pricing table separates Global and Non-global rates. Do not generalize it to every Gemini API endpoint or infer a data-residency promise. |
Gemini 3.8 Flash Cyber is a separate restricted offering through Fairwind. It is not a general security or compliance tier for a marketing API workflow.
Migration checklist for an existing 3.7 API workflow
Start by freezing the current model ID, prompt version, tool list, output schema, token usage and acceptance rule. Then apply the model-specific checklist in Google’s latest-model guide:
- Change the model ID. Set the target to
gemini-3.8-flashand record the date and API surface. - Remove deprecated sampling fields. Google’s 3.8 checklist says to strip
temperature,top_pandtop_k, replacethinking_budgetwith the string settingthinking_level, and removecandidate_count. The generic Interactions API reference exposes broader configuration fields, so use the model-specific migration guidance for this target and revalidate the SDK version you deploy. - Choose a supported thinking level. Use
low,mediumorhigh; do not sendminimalto 3.8. - Standardize conversation state. For stateful Interactions conversations, chain turns with
previous_interaction_idinstead of prefilled model turns or an untracked client-side history. - Audit tool payloads. Follow the current function-calling rules for multimodal assets and inline instructions. If the integration uses
generateContentfunction responses, include bothcall_idandnameas required by the migration guide. - Run the same acceptance check. Compare the old and new model on the same inputs, tools, review rule and reporting window before changing the production route.
A small, current Python example
This example uses the official Interactions Python pattern for a single API turn. Google’s current setup pattern is pip install -U google-genai followed by a locally configured GEMINI_API_KEY environment variable; complete that setup before running it. The short brief is illustrative input, so the model has material to inspect and is told not to invent source links. The example sets the model and thinking level explicitly, then reads the documented output_text property. It is a migration reference and was not executed against a live account here.
from google import genai
client = genai.Client()
brief = """Illustrative campaign brief:
Goal: launch a spring search campaign for a new analytics course.
Constraint: use the approved landing page and report any missing facts.
"""
interaction = client.interactions.create(
model="gemini-3.8-flash",
input=(
"Extract decisions and missing facts from this brief. "
"Do not invent source links or unsupported claims.\n\n" + brief
),
generation_config={
"thinking_level": "low"
}
)
print(interaction.output_text)
For a multi-turn Interactions workflow, pass the prior interaction’s id as previous_interaction_id. Do not combine this snippet with a generateContent tool-result implementation without reworking the payload shape; the two APIs have different turn and function-response rules. Keep writes to a CMS, ad account, CRM or analytics property behind the same approval boundary used for the current workflow.
Choose the price mode before you compare bills
Google’s current pricing table lists dated token rates. The introductory window ends December 31, 2026; the standard rates below start January 1, 2027. Output pricing includes thinking tokens. These are token rates, not a forecast of the cost of a completed task.
| Mode | Input through Dec 31, 2026 | Output through Dec 31, 2026 | Input from Jan 1, 2027 | Output from Jan 1, 2027 | Intro cache read |
|---|---|---|---|---|---|
| Standard | $0.75 / 1M | $3.75 / 1M | $1.50 / 1M | $7.50 / 1M | $0.075 / 1M |
| Batch | $0.375 / 1M | $1.875 / 1M | $0.75 / 1M | $3.75 / 1M | $0.0375 / 1M |
| Flex | $0.375 / 1M | $1.875 / 1M | $0.75 / 1M | $3.75 / 1M | $0.0375 / 1M |
| Priority | $1.35 / 1M | $6.75 / 1M | $2.70 / 1M | $13.50 / 1M | $0.135 / 1M |
Cache storage is listed separately at $0.50 per million tokens per hour through December 31 and $1.00 from January 1 for these modes. The workflow contracts differ even when Batch and Flex share the same token rates: Batch is asynchronous through generateContent with a target turnaround of up to 24 hours; Flex is synchronous, Preview and best-effort with variable latency and possible 429 or 503 responses; Priority is synchronous premium processing for latency-sensitive work. Flex has no automatic fallback to Standard, so use it only where the workload can tolerate that behavior.
For a separate provider-specific price reference, DMT’s GPT-6 Astra API rates guide covers another model surface. It should not be used as a substitute for Gemini’s live table.
Worked token-only budget
Use a clean unit example before adding tools. Assume 10 million input tokens and 1 million output tokens, with output including thinking. Exclude cache storage, cache reads, Search grounding, other tool calls, retries and reviewer time. The arithmetic is:
| Mode or date | Arithmetic | Illustrative token total |
|---|---|---|
| Standard through Dec 31, 2026 | (10 × $0.75) + (1 × $3.75) | $11.25 |
| Batch or Flex through Dec 31, 2026 | (10 × $0.375) + (1 × $1.875) | $5.625 |
| Priority through Dec 31, 2026 | (10 × $1.35) + (1 × $6.75) | $20.25 |
| Standard from Jan 1, 2027 | (10 × $1.50) + (1 × $7.50) | $22.50 |
These totals are a baseline for eligible input and output usage at the listed rates. They are not an invoice forecast. Gemini 3.8 may use more output tokens on a complex workflow, and iterative tools, retry-associated usage, cache storage and human review sit outside this calculation. The same unit rates as Gemini 3.7 therefore do not guarantee the same cost per completed task.
Account for caching and grounding separately
The caching guide separates two patterns. Interactions supports implicit caching only. Explicit cache objects require the generateContent API. For Gemini 3.8, the documented minimum input size for implicit caching is 4,096 tokens, but a cache hit is not guaranteed. Inspect usage.total_cached_tokens in your own telemetry before crediting a cost reduction.
For a mode-aware estimate, use:
Express every token quantity in millions (or divide raw tokens by 1,000,000): uncached_input_M × mode_input_rate + cached_input_M × mode_cache-read_rate + output_M × mode_output_rate + explicit_storage_M_token_hours × storage_rate + measured grounding and tool charges. For an Interactions workflow using implicit caching, use observed cached-token usage and do not add a managed explicit-storage term. Record retry-associated tokens, tool calls and any provider charge as measured usage; do not invent a generic retry fee or count usage that is already included in the token totals twice.
Do not use the Standard cache-read rate for Batch, Flex or Priority. The pricing table lists different cache-read rates for each mode. Also separate API prompts from grounding queries: Google’s Google Search grounding billing guide says a Gemini 3 request can generate multiple Search queries and charges each individual query after the shared allowance. If one request generates three billable queries and the applicable rate is $14 per 1,000 queries, that adds $0.042 at the listed rate. The example shows the unit conversion; it is not an account invoice or a claim about how many queries your prompt will generate.
Decide between Gemini 3.7 and 3.8 by task
Google says Gemini 3.7 Flash remains fully supported. The current standard pricing table gives the two models the same dated standard input and output rates, but the model-specific behavior and token use can still differ. Teams already using Gemini 3.7 Flash for marketer workflows have a useful baseline to preserve while they test 3.8. Do not turn the launch into a universal quality ranking.
| Task pattern | Starting route | Acceptance evidence |
|---|---|---|
| Stable, short, high-volume transformation | Keep 3.7 or test 3.8 at low thinking. | Accepted-output rate, review minutes and token cost on the same inputs. |
| Long context, multi-step analysis or tool orchestration | Test 3.8 at medium, then high only when the job needs it. | Correctness, tool-call completion, loop/retry rate and cost per accepted task. |
| Latency-sensitive response | Test low thinking and a fallback plan that your own system controls. | Time to usable result, timeout rate and reviewer effort. |
| Grounded workflow | Count cached tokens and generated search queries separately. | Source coverage, query count, tool charges and accepted result quality. |
Use cost per accepted task as the migration decision
A model is useful when it produces an accepted result under the team’s quality, time and budget rules. Define “accepted” before the pilot. For a marketing operations job, that might mean required fields are present, each material claim has an approved source, links work, the output passes the review checklist and no substantive rewrite is required.
Report two views when useful:
| Measure | Record | Why it changes the decision |
|---|---|---|
| API-only cost | Input and output including thinking, observed cached tokens, explicit cache-storage hours, generated grounding queries, tool usage and retry-associated usage. Do not add a generic retry fee or double count usage already included in totals. | Shows the billable model and mode cost without hiding usage behind a flat prompt count. |
| Fully loaded cost | API-only cost plus reviewer minutes at the team’s chosen labor rate, if labor is included in planning. | Captures the correction burden that a cheaper token rate can hide. |
| Acceptance | Accepted tasks divided by all attempted tasks, with failure reasons. | Separates a usable result from a response that merely completed. |
| Operational behavior | Latency, 429/503 responses, retries, loops, tool calls and manual interventions. | Shows whether the route is reliable enough for the intended schedule. |
If at least one task is accepted, use cost per accepted task = total measured cost ÷ accepted tasks. If zero tasks are accepted, report the total measured cost and “no accepted result” instead of dividing by zero. Run the same representative jobs through the current 3.7 route and the proposed 3.8 route, hold the source material and acceptance rule constant, and keep the model ID and thinking level in the record. A sensible outcome may be a split route: 3.8 for complex work that clears the acceptance bar and 3.7 for simpler jobs where its efficiency is sufficient.
Questions to settle before production
Does the 3.8 launch require every team to switch?
No. Google says 3.7 remains supported. Use a controlled comparison when the workflow’s quality, latency, tool behavior or cost per accepted task could improve. Keep the current route when the migration work is larger than the measured benefit.
Can I set minimal thinking to reduce cost?
No. The current 3.8 model guide says minimal is unsupported and returns an error. Use low when the job can accept less reasoning effort, and measure the accepted result rather than assuming that a lower setting is automatically cheaper overall.
Do the introductory rates cover the whole workflow?
No. The token table covers eligible input and output usage. Cache storage, cache reads, grounding queries, tool calls, retries, platform-specific fees and review time need separate accounting. Recheck the December 31, 2026 expiry before a workflow runs into 2027.
Sources and scope
Primary sources for this guide are Google’s Gemini 3.8 Flash announcement, model reference, migration guide, Interactions text-generation guide, pricing table, caching guide, Google Search grounding billing guide, Batch documentation, Flex documentation, Priority documentation and Cloud Agent Platform pricing. The code example is source-checked but untested live. No account quota, latency, benchmark, cache-hit or accepted-task result is claimed here.