Gemini 3.8 Flash: pricing, API limits, benchmarks and 3.7 migration

Platform context: For developer deployments evaluating Google’s next iteration, refer to the Gemini 4 Argon guide.

Checked October 9, 2026. Gemini 3.8 Flash costs $0.75 per 1 million input tokens and $3.75 per 1 million output tokens in the paid standard Gemini Developer API through December 31, 2026. On January 1, 2027, Google says those rates will double to $1.50 and $7.50. Batch and Flex cost $0.375/$1.875 and Priority $1.35/$6.75 through December 31; output billing includes thinking tokens. ¹²

Gemini 3.8 Flash is an official Google release, not a leak or an internal-only test. The stable model ID is gemini-3.8-flash. Google documents a 1,048,576-token input limit, a 65,536-token maximum output, multimodal input across text, image, video, audio and PDF, and text output. ¹³

That makes 3.8 Flash unusually easy to compare with 3.7 Flash on list price, but not necessarily on cost per completed task. Google explicitly says the new model may work harder, take extra reasoning steps and use more tokens at higher effort levels. ¹⁴

Gemini 3.8 Flash: verified specs

Item	Gemini 3.8 Flash
Release date	September 2, 2026
Model ID	gemini-3.8-flash
Status	GA / production-ready
Input context	1,048,576 tokens
Maximum output	65,536 tokens
Thinking levels	low, medium, high
minimal thinking	unsupported; returns an error
Standard price through Dec. 31	$0.75 input / $3.75 output per 1M tokens
Standard price from Jan. 1, 2027	$1.50 input / $7.50 output
Batch/Flex through Dec. 31	$0.375 input / $1.875 output
Priority through Dec. 31	$1.35 input / $6.75 output
Inputs	text, image, video, audio, PDF
Output	text
Live API	not supported
Image generation	not supported
Computer use	Preview

What changed from Gemini 3.7 Flash?

Google describes 3.8 Flash as its most intelligent Flash model, targeting long-horizon software engineering, autonomous agents and complex enterprise workflows. The model card says it is based on Gemini 3.7 Flash rather than a completely separate lineage. ³⁴

The design trade-off matters more than the branding. Google says 3.8 can execute additional reasoning steps and iterative tool calls on difficult prompts. That can improve completion rates on agentic work, but it also means a fixed per-token price does not guarantee a fixed cost per task. ¹

For production teams, the right question is therefore not simply “Is 3.8 smarter?” It is: Does 3.8 complete more of my real tasks successfully for an acceptable total token, latency and tool budget?

Thinking levels and migration traps

The supported thinking levels are LOW, MEDIUM and HIGH. Google documents MEDIUM as the default on its Agent Platform guide. MINIMAL is explicitly unavailable and returns a validation error. ⁵

Google also documents stricter API behavior for the current 3.8 path on Agent Platform. Deprecated sampling controls such as temperature, top_k and top_p are described as ignored by the backend, while frequency_penalty, presence_penalty and candidate_count can produce errors. Function-call histories are also validated more strictly. ⁵

Those rules are important migration signals, but they should be applied to the exact Google API surface you are using rather than assumed to describe every third-party provider wrapper.

The example follows Google’s current Gemini Developer API documentation for the Interactions API:

from google import genai

client = genai.Client()

interaction = client.interactions.create(
    model="gemini-3.8-flash",
    input="Review this service for concurrency bugs and propose a safe rewrite.",
    generation_config={
        "thinking_level": "medium"
    },
)

print(interaction.output_text)

Source: Google AI for Developers. ⁶

Google’s 3.8 vs 3.7 benchmark table

Google publishes a direct developer-evaluation comparison between the two Flash releases. These are vendor results, but the matched table is still useful because the values were reported side by side. ⁵

Benchmark	3.8 Flash	3.7 Flash	Change
Terminal-bench 2.1	90.8%	81.6%	+9.2 pp
SWE-Bench Pro	61.6%	60.4%	+1.2 pp
SWE-Atlas	51.9%	48.0%	+3.9 pp
τ³-bench Banking	38.1%	30.9%	+7.2 pp
CharXiv	86.2%	84.5%	+1.7 pp
GDP.pdf	35.0%	34.0%	+1.0 pp
Humanity’s Last Exam	45.4%	45.7%	−0.3 pp

This is a useful correction to simplistic “3.8 wins everything” coverage. The gains are large on some agentic and terminal-oriented tests, modest on others, and slightly negative on the HLE row in this table.

Google separately advertises 54.9% on HLE-Verified in the launch post. That figure should not be merged with the 45.4% HLE number above: the benchmark labels and evaluation variants are different. ¹⁵

Independent check: Artificial Analysis

Artificial Analysis gives Gemini 3.8 Flash at high reasoning 59 on its Intelligence Index, versus 56 for Gemini 3.7 Flash at high reasoning in the direct comparison. It measures roughly 304.6 output tokens per second for 3.8 High. ⁷⁸

The same dataset also shows why “same token price” is an incomplete cost story. Artificial Analysis calls the model verbose relative to its comparison set and reports $0.58 per Intelligence Index task for the High configuration. That is a benchmark-specific task cost, not a general API price per request. ⁷

Its direct comparison also separates generation speed from latency. 3.8 High generates output faster after streaming begins, while time to first token is higher than 3.7 High in that test. Teams building interactive tools should measure both.

Pricing through 2026 and after the scheduled increase

Google’s current Developer API page lists four consumption modes. ²

Mode	Input through Dec. 31, 2026	Output through Dec. 31, 2026	Input from Jan. 1, 2027	Output from Jan. 1, 2027
Standard	$0.75	$3.75	$1.50	$7.50
Batch	$0.375	$1.875	$0.75	$3.75
Flex	$0.375	$1.875	$0.75	$3.75
Priority	$1.35	$6.75	$2.70	$13.50

All rates are USD per million tokens. Standard context-cached input is $0.075 per 1M tokens through December 31, 2026 and $0.15 from January 2027. Keeping cached material also incurs $0.50 per 1M tokens per hour through December 2026, rising to $1.00 per 1M tokens per hour in 2027. Caching therefore does not guarantee lower overall cost if the prefix is rarely reused or retained too long. Google Search grounding in the paid tier includes 5,000 free search requests per month shared across Gemini 3.x, then $14 per 1,000 searches. One model response may use multiple search queries; do not equate model calls with billable search requests. Maps grounding has a separate allowance and may also cost extra. Google Gemini API rate card

Data use: Google’s rate card labels Free Tier requests as potentially used to improve its products, while Paid Tier content is not used for product improvement under the published table. This does not establish a blanket guarantee about regional processing or auxiliary tools. These are Google Developer API rates in USD (pre-tax); state and local sales tax is potentially additional depending on the billing jurisdiction. Google Developer API plans ²

Google Cloud’s Agent Platform page also distinguishes global and non-global inference. For standard 3.8 Flash through December 31, it lists $0.75/$3.75 globally versus $0.825/$4.125 non-globally. ⁹

What those rates mean in real workloads

These are our calculations, using the global standard token rates only. They exclude taxes, tool calls, cache storage and discounts.

Monthly workload	Through Dec. 31, 2026	From Jan. 1, 2027
5M input + 1M output	$7.50	$15.00
50M input + 10M output	$75.00	$150.00
500M input + 100M output	$750.00	$1,500.00

The formula for the middle case is:

50 × input price + 10 × output price

At introductory rates: 50 × 0.75 + 10 × 3.75 = $75.

The calculation does not capture changes in token usage. If 3.8 uses more reasoning output to solve the same task, the model can cost more per successful result even while the nominal token rate matches 3.7. ¹⁴

Should an existing 3.7 deployment migrate immediately?

For a new coding or agent workload, 3.8 is the stronger default candidate: it is GA, uses a stable model ID, costs the same as 3.7 during the introductory period and improves several directly relevant benchmarks. ³⁵²

For a mature high-volume deployment, a blind model-ID swap is less defensible. A useful A/B test should capture:

1. task success rate,
2. input, reasoning and answer tokens per successful task,
3. number of tool calls and retries,
4. time to first token,
5. end-to-end completion time,
6. total API and tool cost per successful task.

That evaluation can reveal whether the extra reasoning translates into fewer retries and higher completion rates, or mostly into higher output cost.

Hands-on migration from a Mac: Our Gemini 3.7-to-3.8 migration guide includes copyable Python calls to both cloud endpoints, actual input/answer/thinking-token metadata, and Google’s current guidance on thinking_level, deprecated sampling fields and multi-step tool interactions. These are instructions, not benchmark results. Validate your own appropriately scoped workload before changing production routing.

Known limitations and safety signals

Google’s model card still lists hallucinations, occasional slowness or timeouts, and higher token use at higher effort levels among the limitations. ⁴

The knowledge cutoff is also not one clean date. Google says users can expect March 2026 knowledge in some domains, while other domains can remain limited to January 2025. Time-sensitive workflows should therefore use grounding or external retrieval rather than assume the model itself knows events through launch day. ⁴

The safety table is mixed. In Google’s automated internal comparison against 3.7 Flash, text-to-text safety moves by −0.4 percentage points where lower is better, while multilingual safety moves +5.4 points, which Google describes as a regression. These are automated development evaluations, not a full human red-team result. ⁴

Where Gemini 3.8 Flash is available

Google lists the Gemini API, Google AI Studio, Gemini Enterprise Agent Platform, Gemini app, AI Mode and Antigravity among distribution channels. The launch post also names Android Studio and Stitch for developers, and Sheets for eligible consumer plans. Surface-level rollout can still vary by account or region. ¹⁴

A Business Insider report from August 27 described employee testing of the next Flash model. That report is now historical context: the model was formally released on September 2. ¹¹⁰

What matters for production

Gemini 3.8 Flash looks like a meaningful 3.7 upgrade for coding and agentic work, not a universal benchmark sweep. Google’s own table contains both sizable gains and one small HLE decline; Artificial Analysis independently reports a higher intelligence score and faster output generation, while also showing a more verbose model and higher first-token latency in the direct comparison. ⁵⁷⁸

The biggest deployment question is cost per completed task. Through December 31, the standard rate remains $0.75/$3.75 per million input/output tokens, but 3.8 may spend more tokens to solve difficult work. On January 1, 2027, the published token rates are scheduled to double. ²

For production systems, test 3.8 and 3.7 on the same real tasks before moving all traffic.

Source notes

Footnotes

1. Google: launch ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸

2. Google AI: API pricing ↩ ↩² ↩³ ↩⁴ ↩⁵

3. Google: model specs ↩ ↩² ↩³

4. DeepMind: model card ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷

5. Google Cloud: 3.8 guide ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶

6. Google: latest models ↩

7. Artificial Analysis ↩ ↩² ↩³

8. Artificial Analysis: comparison ↩ ↩²

9. Google Cloud: pricing ↩

10. Business Insider ↩