Gemini 3.8 Flash: Specs, Pricing, What Changed (2026)

Google shipped Gemini 3.8 Flash on 2 September 2026, three weeks after Gemini 3.7 Flash. The unusual part is the price tag: there isn't a new one. Every tier, every modality, the same context window, the same output ceiling, the same parameter surface. On paper, 3.8 Flash is a free upgrade.

That makes the question narrow and specific: if nothing commercial moved, what actually got better — and did anything get worse? This guide answers that from Google's documentation, its published benchmark slides, and three independent evaluators, keeping vendor claims and third-party measurements clearly separated.

What is Gemini 3.8 Flash?

Gemini 3.8 Flash is the current stable model in Google's Flash line — the tier positioned between the cheap Flash-Lite models and the frontier Pro/Ultra models. Google's Gemini API release notes describe the 2 September GA release as "our most intelligent Flash model, engineered for long-horizon software engineering, autonomous agents, and complex enterprise workflows."

It is a reasoning model by default. Thinking cannot be switched off entirely; it can only be dialled between three levels.

Two absences are worth flagging before you plan around this model. There is no Live API support, so 3.8 Flash cannot drive a bidirectional realtime voice session — it ingests audio as a file, it does not converse over a stream. And it produces text only: no image generation, no speech output. Those live on Google's separate Omni and image models.

Google also does not publish a knowledge cutoff date. The official model page lists every token limit and capability flag but omits the cutoff, as it does for 3.7 Flash. If your workload depends on recent world knowledge, use search grounding or URL context rather than assuming a cutoff you can't verify.

What changed from Gemini 3.7 Flash?

Start with what didn't change, because it's almost everything. Gemini 3.7 Flash, released 13 August 2026, has the identical model ID pattern, the identical 1,048,576-token context, the identical 65,536-token output ceiling, the identical five input modalities, the identical capability matrix, and — verified against Google's pricing page — the identical price on every single tier.

So the delta is entirely in model quality, and it is real but uneven. Here is the honest picture:

The pattern is clear once you line it up. Agentic and multi-step reasoning moved meaningfully — the Artificial Analysis agentic index is up nearly five points, the largest single movement anywhere in the comparison. Raw coding is flat: +0.2 on the coding index is noise. Human preference is flat too — a three-point LMArena gap with ±8 and ±9 error bars is statistically indistinguishable, and both scores are still marked preliminary.

There is also a genuine regression, and Google doesn't mention it. Time to first token got worse, from 12.01 to 13.21 seconds — about 10% slower to start responding, even though it streams faster once it starts. On the Design Arena website category, OpenRouter's mirrored scores put 3.8 Flash at 1311 Elo and a 56.3% win rate against 3.7 Flash's 1318 Elo and 57.1%. Slightly behind its predecessor on front-end design taste.

Google's most eye-catching claim is a customer quote, not a benchmark. On the DeepMind Flash page, Glean's Thai Tran says 3.8 Flash completed "more than three times as many tasks as Gemini 3.7 Flash in our evaluations." That is a document-heavy workload at one company on an undisclosed harness — directionally consistent with the agentic index jump, but not a number you should expect to reproduce.

We cover the head-to-head in depth in Gemini 3.8 Flash vs 3.7 Flash. For the previous generation's launch context, see our Gemini 3.7 Flash launch guide.

What does Gemini 3.8 Flash cost, really?

The list prices are simple. What people get wrong is the effective price, because of how thinking tokens are billed.

Two things in that table are easy to miss.

First, Google's pricing page states the output price "includes thinking tokens." Gemini 3.8 Flash reasons by default at medium, and every one of those internal tokens is billed at $3.75 per million — the output rate, not the input rate. You never see them in the response body, but you pay for them. The thinking documentation is explicit: "response pricing is the sum of output tokens and thinking tokens," and exposes the count as total_thought_tokens.

Second, the current prices are promotional through 31 December 2026. Google's pricing page shows standard rates doubling on 1 January 2027 — to $1.50 input and $7.50 output. If you are modelling annual spend, model two halves of the year.

A worked example that includes the reasoning tokens

Take a realistic agentic call: 30,000 input tokens of context, 800 tokens of visible answer, and 6,000 thinking tokens at the default medium level.

• Input: 30,000 × $0.75 / 1M = $0.0225
• Output + thinking: (800 + 6,000) × $3.75 / 1M = $0.0255
• Total: $0.0480 per call

The naive estimate — the one almost every cost calculator produces — counts only the 800 visible output tokens and lands on $0.0255. The real bill is 1.9× that. At 1,000 calls a day, the difference is roughly $22 a day, or about $660 a month, on a workload you had budgeted at half the price.

Three levers actually move this number:

• Drop thinking_level to low on tasks that don't need deliberation. This directly cuts the most expensive token class. It is the single highest-leverage cost control on this model.
• Use the Batch API for anything not user-facing. Half price on both input and output takes the example call from $0.0480 to $0.0240.
• Cache the stable prefix. Cached input reads cost $0.075 per 1M — a 10× discount. If 25,000 of the 30,000 input tokens are a fixed system prompt and document set, input drops from $0.0225 to $0.0056, and the call lands near $0.031. Watch the $0.50 per 1M-tokens-per-hour storage fee: caching only pays off if you re-read the cache often enough to beat it.

How good is Gemini 3.8 Flash?

Google's own numbers, from the DeepMind Flash page, position 3.8 Flash at the top of a field that includes frontier models several times its price:

• HLE-Verified: 54.9% — ahead of GPT-5.6 Sol (54.5%), Claude Opus 5 (54.4%), Gemini 3.7 Flash (53.6%), GPT-5.6 Terra (51.1%) and Claude Sonnet 5 (31.0%).
• Vals Finance Agent v2: 61.4% — ahead of 3.7 Flash (59.0%), Claude Opus 5 (58.6%), GPT-5.6 Terra (54.4%), Claude Sonnet 5 (53.9%) and GPT-5.6 Sol (53.8%).
• Harvey Legal Agent: 10.0% — ahead of 3.7 Flash (8.8%), Claude Opus 5 (6.7%), Claude Sonnet 5 (5.0%), GPT-5.6 Sol (2.5%) and GPT-5.6 Terra (0.8%). Note the absolute scores: this benchmark is brutally hard and a 10% pass rate is the current state of the art, not a good result in plain English.
• DeepSWE v1.1 — Google publishes a cost-versus-score chart only, with no numeric values. We are not going to invent them.

Independent measurements tell a more measured story. Artificial Analysis puts Gemini 3.8 Flash (high) at an Intelligence Index of 59, ranked #17 of 196 models, against a median of 36 — comfortably above average but not at the frontier. Its output speed of 298.6 tokens/second ranks #3 of 196, which is the more remarkable result and the reason to care about this tier at all. Time to first token is 13.21 seconds, a consequence of mandatory reasoning. Blended cost to run their evaluation suite: $0.58.

On LMArena's text leaderboard, Gemini 3.8 Flash High sits at #8 with 1494 ±9, still marked preliminary — behind Claude Fable 5 (1507), Claude Opus 4.6 High (1505) and Claude Opus 4.7 High (1502), and level with Claude Opus 4.7 (1494).

The fair summary: on blind human preference, a Flash-tier model now trades punches with last generation's flagships. On hard reasoning benchmarks it is strong for the price but not the best available. And its real advantage is that it delivers those scores at roughly 300 tokens per second for $0.75 in.

What can Gemini 3.8 Flash actually take as input?

This is where the Flash tier genuinely separates from its price-matched rivals. Most models at $0.75 per million input tokens take text and images. Gemini 3.8 Flash takes video and audio too, at meaningful lengths.

Run the arithmetic and something clicks: 3 hours of low-res video at 100 tokens/second is 1,080,000 tokens. One hour of high-res video at 300 tokens/second is also 1,080,000. And 9.5 hours of audio at 32 tokens/second is 1,094,400. All three caps are the 1M context window expressed in different units. They are not arbitrary product limits — they are the context ceiling, and they move together with it.

What that costs in practice, at the standard input rate:

• One hour of audio — 115,200 tokens, about $0.086 of input. Transcribing, summarising or extracting action items from an hour-long recording costs under nine cents before output.
• One hour of low-resolution video — 360,000 tokens, about $0.27 of input.

Per Google's audio documentation, incoming audio is downsampled to 16 Kbps and multi-channel audio is combined to a single channel, so don't expect speaker separation from the raw signal. Thirteen MIME types are accepted, including WAV, MP3, FLAC, AAC, Opus and M4A, and the model understands non-speech sound — birdsong, sirens — not just transcription.

On the video side, nine container formats are supported and public YouTube URLs can be passed directly. The free tier caps YouTube ingestion at 8 hours per day; the paid tier has no length-based limit. Gemini 3.8 Flash also supports agentic video understanding, added to the Flash line on 1 September 2026 — instead of statically tokenising every frame, the model navigates the timeline and requests transcripts, frames or audio tracks on demand, which Google says uses up to 88% fewer tokens on long-form content. On a three-hour video that is the difference between a $0.81 input bill and roughly a tenth of it.

How do you use Gemini 3.8 Flash?

The model string is gemini-3.8-flash. There is one stable version and no preview or experimental alias.

A minimal call against the Gemini API:

curl -s "https://generativelanguage.googleapis.com/v1beta/models/gemini-3.8-flash:generateContent" \
  -H "x-goog-api-key: $GEMINI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "contents": [
      { "parts": [ { "text": "Summarise this changelog into five bullets." } ] }
    ]
  }'

In Python, with the google-genai SDK, dialling thinking down for a cheap extraction task:

from google import genai
from google.genai import types

client = genai.Client()  # reads GEMINI_API_KEY

resp = client.models.generate_content(
    model="gemini-3.8-flash",
    contents="Extract every version number from this changelog as JSON.",
    config=types.GenerateContentConfig(
        thinking_config=types.ThinkingConfig(thinking_level="low"),
        max_output_tokens=1024,
    ),
)

print(resp.text)
print(resp.usage_metadata)  # inspect thought tokens before you scale up

Check usage_metadata on a representative sample of real traffic before committing to a budget. The thinking-token count decides your bill, and it varies far more by task than input length does.

Where you can run it:

• Google AI Studio — free tier available; the rate-limit documentation confirms 3.8 Flash is live for free-tier projects. Per-tier RPM/TPM figures are shown inside AI Studio rather than in the public docs.
• Vertex AI / Gemini Enterprise Agent Platform — for production workloads that need enterprise controls and regional deployment.
• OpenRouter — as google/gemini-3.8-flash, with google/gemini-3.8-flash:batch at exactly half price. OpenRouter exposes six endpoints (AI Studio and Vertex global, each in standard, flex and priority variants), and maps Google's thinking_level onto the OpenAI-style reasoning_effort parameter. Supported parameters there include reasoning_effort, structured_outputs, seed, tools, tool_choice, response_format, temperature and top_p.

Should you upgrade from Gemini 3.7 Flash?

For most workloads, yes — and mainly because there is no downside. Same price, same context, same modalities, same parameters. A model string change is the whole migration.

The decision rule, honestly stated:

• Upgrade now if your workload is agentic, multi-step, or long-horizon. That is where the entire measured gain sits: +4.9 on the independent agentic index and every one of Google's published wins.
• Upgrade, but don't expect a difference if you are doing straight code generation. The independent coding index moved 0.2 points. You will not notice.
• Test before you switch if you are latency-sensitive on first-token time. 3.8 Flash is 1.2 seconds slower to start. For a streaming chat UI, that is a perceptible regression the throughput gain does not fully offset.
• Stay on 3.7 Flash if you have front-end design output you have already tuned and validated. It scores marginally higher on Design Arena's website category, and there is no cost saving to be had from moving.

Gemini 3.7 Flash is not deprecated. Google's model list still shows it as stable, labelled the "previous-generation Flash model," with no shutdown date announced. There is no forced-migration clock, which is exactly why you can spend a week benchmarking on your own traffic rather than trusting anyone's index.

The most important number in this launch is not a benchmark at all. It is 1 January 2027, when promotional pricing ends and both models double to $1.50 and $7.50. Whatever you build on this tier, build the batch path and the caching path in now — those are the two levers that still work after the price change.

For where the Flash tier sits in Google's wider lineup, see our Gemini 3.5 complete guide. If you are cost-shopping this bracket, the closest open-weight rival is covered in our GLM-5.3-Flash guide.

FAQ

What is Gemini 3.8 Flash?

Gemini 3.8 Flash is Google's newest Flash-tier AI model, generally available on 2 September 2026 under the model ID gemini-3.8-flash. It accepts text, image, video, audio and PDF input, returns text only, has a 1,048,576-token context window with a 65,536-token output limit, and reasons by default at a configurable low/medium/high thinking level.

How much does Gemini 3.8 Flash cost?

$0.75 per 1M input tokens and $3.75 per 1M output tokens on the standard tier. The Batch API and Flex inference halve that to $0.375 and $1.875; Priority inference raises it to $1.35 and $6.75. Cached input reads cost $0.075 per 1M. These are promotional rates that double on 1 January 2027.

Is Gemini 3.8 Flash better than 3.7 Flash?

Yes, but unevenly. Artificial Analysis scores it 59 versus 56 on intelligence and 50.0 versus 45.1 on agentic tasks — a real gain. Coding is flat at 76.3 versus 76.1, and LMArena's 1494 versus 1491 is within the error bars. Time to first token is actually 1.2 seconds worse.

What is Gemini 3.8 Flash's context window?

1,048,576 input tokens — one full megatoken — with a separate 65,536-token maximum output. This is identical to Gemini 3.7 Flash. The context ceiling is also what caps every other modality: 3 hours of low-resolution video, 1 hour of high-resolution video and 9.5 hours of audio all work out to roughly 1M tokens.

Does Gemini 3.8 Flash support audio and video?

Yes, as input. It accepts up to 9.5 hours of audio per prompt at 32 tokens per second, and video up to 3 hours at low media resolution or 1 hour at high, plus public YouTube URLs. It also supports agentic video understanding, which Google says cuts token use by up to 88% on long-form content. It cannot generate audio or video.

Is Gemini 3.8 Flash free?

There is a free tier in Google AI Studio, and Gemini 3.8 Flash is available on it — with rate limits, and with an 8-hour-per-day cap on YouTube video ingestion. Exact free-tier RPM and TPM figures are shown inside AI Studio rather than published in the docs. Production use runs on the paid tiers above.

Is Gemini 3.7 Flash deprecated?

No. Google's model list still shows gemini-3.7-flash as a stable model, now labelled the previous-generation Flash model, with no announced shutdown date. The only near-term Gemini deprecation on the changelog is gemini-omni-flash-preview, retiring 30 September 2026. You can migrate to 3.8 Flash on your own schedule.

Do thinking tokens cost extra on Gemini 3.8 Flash?

They are billed at the output rate of $3.75 per 1M, and Google's docs state response pricing is the sum of output and thinking tokens. Because thinking is on by default at medium, real costs routinely run around twice a naive estimate based on visible output alone. Check total_thought_tokens in usage metadata, and drop thinking_level to low where deliberation isn't needed.