Gemini 3.8 Flash
Gemini 3.8 Flash is Google's latest Flash model, released on Sep 2, 2026 and described by the company as its most intelligent workhorse model. It costs US$ 0.75 per million input tokens and US$ 3.75 per million output until Dec 31, 2026, the same as 3.7 Flash, and twice that from Jan 1, 2027, with a 1,048,576 token context. In Google's table it beats 3.7 Flash on all 14 tests and leads 8 of them against much pricier models, but falls far behind on long terminal work and computer use.
Gemini 3.8 Flash in official numbers
Data published by the maker in its documentation and announcement. The source of each item is at the end of the page.
- Maker
- Google DeepMind
- API identifier
gemini-3.8-flash(stable, GA)- Release
- Sep 2, 2026
- Input
- US$ 0.75 per million until Dec 31, 2026; US$ 1.50 from Jan 1, 2027
- Output (includes reasoning)
- US$ 3.75 per million until Dec 31, 2026; US$ 7.50 from Jan 1, 2027
- Cache read
- US$ 0.075 per million (US$ 0.15 in 2027), plus hourly storage
- Long context
- single price, no surcharge for large prompts
- Context window
- 1,048,576 tokens
- Max output
- 65,536 tokens
- Reliable knowledge
- through March 2026; in some domains, through January 2025
- Reasoning
- low, medium (default) and high; minimal is not supported
- Input and output
- text, image, video, audio and PDF in; text only out
- Free tier
- yes, on the Gemini API and Google AI Studio
- Where it runs
- Gemini API, Google AI Studio, Google Cloud, Antigravity and, for AI Pro and Ultra subscribers, the Gemini app and AI Mode in Search, according to the announcement and model card
Where Gemini 3.8 Flash does well, where it falls behind and who it is for
Where it does well
- Long-horizon coding for the price: 73.7% on DeepSWE v1.1, against 74.0% for Claude Opus 5, which cost US$ 5 and US$ 25 in the same Google table.
- Legal and finance agents: leads Vals Finance Agent v2 (61.4%) and Harvey's Legal Agent Benchmark (10.0%) in the launch comparison.
- Charts and video: 86.2% on CharXiv Reasoning without tools and 87.1% on LVBench, a long-video test, leading both rows.
- Large context without a surcharge: 1,048,576 input tokens at a single price. On GPT-6 Luna, prompts above 272 thousand tokens pay double for input.
- A jump over 3.7 Flash at the same price: Terminal-Bench 4.0 from 11.2% to 19.1% and OSWorld 2.0 from 50.6% to 59.0%.
Where it falls behind
- Long terminal work: 19.1% on Terminal-Bench 4.0, against 51.8% for Claude Opus 5 and 37.3% for GPT-5.6 Sol.
- Computer use: 59.0% on OSWorld 2.0, against 75.4% for Claude Opus 5.
- Safety outside English: the model card admits a 5.4 percentage point regression on the automated multilingual safety test, compared with 3.7 Flash.
- Spends more tokens: according to Google, 3.8 Flash works harder per task, with extra reasoning steps and tool calls, and reasoning is billed as output.
- Price doubles in 2027: from Jan 1, 2027, US$ 1.50 and US$ 7.50, above Claude Haiku 4.5 (US$ 1 and US$ 5).
Who it makes sense for
- High-volume coding agents and business workflows, where 3.7 Flash fell short and a frontier model weighs too much on the bill.
- Long, mixed material in one request: PDFs, contracts, meeting recordings and video, within 1 million tokens.
- Financial and legal research with an agent, two professional-work tests it leads in Google's table.
- Prototyping on the Google AI Studio free tier before paying.
When to pick another model
- Agents that spend hours in the terminal or operate screens: Terminal-Bench 4.0 and OSWorld 2.0 are where it trails the most.
- Huge volumes of simple tasks, such as classifying and extracting: GPT-6 Luna costs US$ 0.10 and US$ 0.50.
- Public-facing support in Portuguese or other non-English languages on sensitive topics without your own moderation, because of the safety regression outside English.
- Instant answers with no reasoning: the lowest level accepted is low. GPT-6 Luna lets you turn reasoning off.
What Google published at launch
3.8 Flash came out on September 2, 2026, three weeks after 3.7 Flash and as Google's third Flash release in six weeks, at the same price and speed as its predecessor. The same day saw 3.8 Flash Cyber, a version trained for security and released only to approved defenders.
The table below is the one from the announcement, repeated in the model card. It compares 3.8 Flash with models on the market on September 2. Since then, Claude Opus 5.5, GPT-6 Sol and GPT-6 Luna came out on Sep 22, and GPT-6.1 Sol on Sep 29; none of them appears here.
Two caveats from Google itself. The methodology PDF admits that the 74.0% for Claude Opus 5 on DeepSWE was reported incorrectly because of rounding, without publishing the correct value. And LVBench shows 87.8% in agentic mode, a feature the Google Cloud documentation marks as not supported on 3.8 Flash; the static mode, 87.1%, already leads the row.
| 3.8 Flash | 3.7 Flash | Opus 5 Anthropic | Sonnet 5 Anthropic | GPT-5.6 Sol OpenAI | GPT-5.6 Terra OpenAI | |
|---|---|---|---|---|---|---|
| Input price US$ per million, no caching |
0.75 (1.50 later) | 0.75 (1.50 later) | 5.00 | 2.00 | 4.00 | 2.00 |
| Output price US$ per million |
3.75 (7.50 later) | 3.75 (7.50 later) | 25.00 | 10.00 | 20.00 | 12.00 |
| DeepSWE v1.1 long-horizon coding |
73.7% | 65.3% | 74.0% | 53.8% | 72.7% | 69.6% |
| GDPVal-AA v2 office work, Elo |
1,545 | 1,482 | 1,824 | 1,584 | 1,710 | 1,528 |
| Vals Finance Agent v2 financial research |
61.4% | 59.0% | 58.6% | 53.9% | 53.8% | 54.4% |
| Harvey's Legal Agent Benchmark legal work |
10.0% | 8.8% | 6.7% | 5.0% | 2.5% | 0.8% |
| Terminal-Bench 2.1 terminal |
89.4% | 85.8% | 89.1% | 80.4% | 88.8% | 87.4% |
| Terminal-Bench 4.0 long terminal work |
19.1% | 11.2% | 51.8% | 12.4% | 37.3% | 23.6% |
| GDP.PDF all pass rate |
35.0% | 34.0% | 37.0% | 28.0% | 40.0% | 29.0% |
| CharXiv Reasoning chart reading, no tools |
86.2% | 84.5% | 83.7% | 70.1% | 85.8% | 85.9% |
| LVBench long video; for 3.8, static and agentic |
87.1% and 87.8% | 85.4% | 75.4% | 68.5% | 82.1% | 78.9% |
| HLE-Verified hard knowledge |
54.9% | 53.6% | 54.4% | 31.0% | 54.5% | 51.1% |
| OSWorld 2.0 computer use |
59.0% | 50.6% | 75.4% | 42.6% | 62.6% | 50.2% |
| BioMysteryBench biology, human solvable |
88.8% | 87.1% | 90.1% | 87.5% | 79.5% | 83.8% |
| BioMysteryBench biology, human difficult |
56.5% | 43.5% | 49.4% | 34.1% | 44.7% | 49.4% |
| LABBench2 science |
86.2% | 82.1% | 84.2% | 80.1% | 82.1% | 81.2% |
Figures from the Gemini 3.8 Flash announcement and model card, with the prices Google published in the same table; the 3.8 and 3.7 Flash price is introductory until Dec 31, 2026. The dot marks the best in each row.
3.8 Flash, GPT-6 Luna or Claude Haiku 4.5
No maker has put the three in the same table, so the comparison here is about price and specs; the independent scores are at the top of this page. GPT-6 Luna, from OpenAI, is the cheapest per token: US$ 0.10 and US$ 0.50. Claude Haiku 4.5, from Anthropic, costs US$ 1 and US$ 5. 3.8 Flash sits between the two at the introductory price and passes Haiku 4.5 from January. Against Luna, Flash takes video, audio and PDF as input, where Luna takes text and images; Luna answers up to 128 thousand tokens, against 65,536, and has a more recent knowledge cutoff.
On performance, with caution: OpenAI published 66.6% for Luna on DeepSWE v1.1 at max effort, and Google published 73.7% for 3.8 Flash on the same test, run by Google with high reasoning. These are measurements by different makers, under different conditions, and they cannot settle a winner.
| 3.8 Flash until Dec 31, 2026 | 3.8 Flash from 2027 | GPT-6 Luna OpenAI | Haiku 4.5 Anthropic | |
|---|---|---|---|---|
| Input | US$ 0.75 | US$ 1.50 | US$ 0.10 | US$ 1 |
| Output | US$ 3.75 | US$ 7.50 | US$ 0.50 | US$ 5 |
| Cache read | US$ 0.075 | US$ 0.15 | US$ 0.01 | US$ 0.10 |
| Context window | 1,048,576 | 1,048,576 | 1,050,000 (input up to 922,000) | 200 thousand |
| Max output | 65,536 | 65,536 | 128,000 | 64 thousand |
| Knowledge through | March 2026 | March 2026 | May 18, 2026 | February 2025 |
Standard pricing. GPT-6 Luna charges double for input above 272 thousand tokens; 3.8 Flash has a single price. Google, OpenAI and Anthropic documentation checked on Oct 1, 2026.
Before switching from 3.7 Flash to 3.8
The price per token is the same, but cost per task can go up, because 3.8 Flash uses more reasoning steps. When efficiency is the main constraint, Google itself recommends lowering the effort level or staying on 3.7 Flash, which remains fully supported.
The migration touches code: the Gemini API guide says to replace thinking_budget with thinking_level and remove temperature, top_p and top_k. Google Cloud still lists default sampling values for the same model; check your platform.
The test that decides: ten real tasks on 3.7 and 3.8 Flash at medium effort, comparing accuracy and cost per task, already at the 2027 price. If your app talks to the public in a language other than English, include sensitive cases: the safety regression Google admitted is outside English, although the company says its manual review found mostly false positives or non-egregious cases.
Examples of using Gemini 3.8 Flash
An agent fixes bugs and opens pull requests across dozens of repositories a day. On DeepSWE v1.1, 3.8 Flash scores 73.7%, and Google says it beats most larger frontier models at a fraction of the cost.
Contracts and balance sheets in PDF go into the same request, within 1 million tokens; on Google Cloud, PDFs of up to 50 MB and 3,000 pages. Vals Finance Agent v2 and Harvey's Legal Agent Benchmark, which measure this work, are among the tests it leads.
A recording of about 45 minutes with audio, the limit documented on Google Cloud, becomes minutes with decisions and owners. LVBench, a long-video test, is one of the tests 3.8 Flash leads.
Where Gemini 3.8 Flash stands out and where it falls behind, by the numbers
Gemini 3.8 Flash has published results on 30 tests compiled by Epoch AI. The yardstick here is its generation: its position among the models released between Dec 6, 2025 and Dec 1, 2026 that took the same test. Against the whole archive, which includes 2023 models, a recent release lands near the top of almost everything, and that says nothing.
Where it stands out
Tests where it sits in the top third of its own generation.
Where it falls behind
Tests where it sits in the bottom half of its own generation.
See all 30 tests, one by one
| Test | Score | In its generation | Overall | Test record |
|---|---|---|---|---|
| GPQA Diamond | 95.4% | #4 / 76 (tied with 1) | #4 / 173 | 96.4% Gemini 4 Argon |
| AIME 2024 and 2025 (OTIS) | 98.9% | #17 / 78 (tied with 3) | #17 / 168 | 100.0% GPT-5.5 |
| Dtbench | 95.7% | #16 / 66 | #16 / 148 | 98.9% Claude Opus 5.5 |
| Lmca | 52.90 | #16 / 66 | #16 / 121 | 68.23 Claude Opus 5.5 |
| CritPt | 18.3% | #29 / 77 | #29 / 112 | 32.3% GPT-5.6 Sol |
| WeirdML | 84.8% | #7 / 50 | #7 / 112 | 93.6% GPT-6 Astra |
| SciCode | 56.6% | #19 / 75 (tied with 1) | #19 / 104 | 66.9% Claude Opus 5.5 |
| Frontiermath Tiers 1 3 V2 | 68.4% | #29 / 67 | #29 / 86 | 93.7% GPT-6 Astra |
| ARC-AGI 2 | 89.2% | #8 / 46 (tied with 1) | #8 / 83 | 95.0% GPT-6 Astra |
| ALE-Bench | 1,270 | #24 / 60 | #24 / 82 | 2,951 GPT-6 Astra |
| WebDev Arena | 1,568 | #26 / 70 | #26 / 82 | 1,820 Claude Opus 5.5 |
| ARC-AGI | 98.5% | #1 / 46 (tied with 4) | #1 / 81 | 98.5% Claude Fable 5 |
| SimpleQA Verified | 69.7% | #8 / 62 (tied with 1) | #8 / 81 | 76.0% Gemini 4 Argon |
| Chess puzzles | 58.0% | #12 / 62 | #12 / 80 | 76.0% GPT-6 Astra |
| Game puzzles | 47.0% | #13 / 60 | #13 / 77 | 84.0% GPT-6 Astra |
| ProofBench | 48.0% | #33 / 67 | #33 / 74 | 100.0% Claude Fable 5.1 |
| Frontiermath Tier 4 V2 | 22.0% | #43 / 57 (tied with 1) | #43 / 67 | 100.0% GPT-6.1 Sol |
| Vending Bench 2 | 5,094 | #29 / 46 | #30 / 57 | 15,515 GPT-6 Astra |
| Apex Agents | 64.3% | #9 / 48 | #9 / 50 | 82.4% Gemini 4 Argon |
| FrontierCode | 41.2% | #24 / 43 | #24 / 43 | 54.6% Claude Opus 5.5 |
| Hle | 44.5% | #5 / 14 | #5 / 39 | 54.8% GPT-6 Astra |
| Gdp Pdf | 23.4% | #17 / 37 | #17 / 37 | 35.0% GPT-6.1 Sol |
| Furniture Assembly | 31.7% | #24 / 33 | #24 / 34 | 83.3% Claude Opus 5.5 |
| Blueprint Bench 2 | 38.6% | #5 / 30 (tied with 1) | #5 / 31 | 54.4% Gemini 4 Argon |
| DeepSWE | 73.8% | #2 / 26 | #2 / 26 | 74.1% GPT-6 Astra |
| Surface Evolver Bench | 76.9% | #7 / 23 | #7 / 24 | 95.0% Claude Fable 5 |
| FrontierSWE | 19.6% | #16 / 20 | #16 / 20 | 65.5% GPT-6 Astra |
| CursorBench | 39.6% | #13 / 16 | #13 / 16 | 57.8% Claude Opus 5.5 |
| Rli | 5.8% | #7 / 10 | #7 / 15 | 21.3% Claude Opus 5.5 |
| Weirdml V3 | 7.4% | #8 / 10 | #8 / 11 | 42.2% GPT-6 Astra |
For each test, the model's best effort level counts, as in the ranking. Source of the scores: Epoch AI, under CC BY 4.0.
What to compare Gemini 3.8 Flash with
The models from other labs closest in capability and released in the same period, at most two per lab. Click any of them to see its page.
| Model | Lab | ECI | ELO | LiveBench | Release |
|---|---|---|---|---|---|
| Kimi K3 | Moonshot | 157.42 | 1,528 | 79.2 | Jul 16, 2026 |
| Gemini 3.8 Flash | Google DeepMind | 157.20 | 1,551 | 75.8 | Sep 2, 2026 |
| Qwen 3.8 Max | Alibaba | 156.88 | 1,536 | 78.5 | Aug 2, 2026 |
| Muse Spark 1.3 | Meta AI | 156.76 | 1,528 | 81.6 | Sep 2, 2026 |
| GPT-5.3 Codex | OpenAI | 156.76 | n/a | n/a | Feb 5, 2026 |
| GPT-5.4 | OpenAI | 156.72 | 1,519 | 78.0 | Mar 5, 2026 |
| Claude Sonnet 5 | Anthropic | 156.36 | 1,491 | 76.0 | Jun 30, 2026 |
Gemini 3.8 Flash or Kimi K3?
Kimi K3 is the model from another lab closest to Gemini 3.8 Flash in capability. Across the 26 tests both took, this is the score:
| Test | Gemini 3.8 Flash | Kimi K3 |
|---|---|---|
| ProofBench | 48.0% | 87.0% |
| Game puzzles | 47.0% | 26.0% |
| Frontiermath Tier 4 V2 | 22.0% | 39.0% |
| Chess puzzles | 58.0% | 36.0% |
| ARC-AGI 2 | 89.2% | 60.4% |
| SimpleQA Verified | 69.7% | 50.6% |
| FrontierSWE | 19.6% | 25.9% |
| Blueprint Bench 2 | 38.6% | 29.5% |
| CritPt | 18.3% | 23.4% |
| Apex Agents | 64.3% | 50.6% |
Unweighted score: each test counts one point, whether math or coding. The dot marks who did better on each row.
What Gemini 3.8 Flash costs to deliver each task
Price per token is what is on the price sheet; cost per task is what lands on the invoice, because it depends on how many tokens the model spends. Each point is an effort level, published by whoever ran the test and compiled by Epoch AI. The further up and to the left, the more score per dollar.
CursorBench
The score scale does not start at zero.
DeepSWE
The score scale does not start at zero.
Other models from Google DeepMind
The models from the same lab closest in time, newest first.
| Model | Release | ECI | ELO | LiveBench | Access |
|---|---|---|---|---|---|
| Gemini 4 Argon | Sep 30, 2026 | 164.56 | 1,608 | n/a | Limited access |
| Gemini 3.8 Flash | Sep 2, 2026 | 157.20 | 1,551 | 75.8 | API access |
| Gemini 3.7 Flash | Aug 13, 2026 | 157.22 | 1,551 | 78.8 | API access |
| Gemini 3.6 Flash | Jul 21, 2026 | 154.24 | 1,530 | 73.6 | API access |
| Gemini 3.5 Flash-Lite | Jul 21, 2026 | 145.15 | 1,467 | 63.9 | API access |
| Gemini 3.5 Flash | May 19, 2026 | 154.38 | 1,526 | 74.6 | API access |
| Gemma 4 31B IT | Apr 2, 2026 | 142.86 | n/a | n/a | Open weights (restricted use) |
| Gemma 4 26B A4B | Apr 2, 2026 | 141.81 | 1,495 | n/a | Open weights (unrestricted) |
| Gemini 3.1 Pro | Feb 19, 2026 | 154.67 | 1,528 | 77.0 | API access |
Have you used Gemini 3.8 Flash? Rate it
Users' opinion, not measurement: the rating is kept apart from the ranking and does not move anyone's position. One vote per browser, no sign-up, and you can change your rating later.
What people ask about Gemini 3.8 Flash
How much does Gemini 3.8 Flash cost?
Until Dec 31, 2026, US$ 0.75 per million input tokens and US$ 3.75 per million output, an introductory price. From Jan 1, 2027, US$ 1.50 and US$ 7.50. Reasoning tokens count as output, Batch and Flex cost half, and the Gemini API has a free tier for the model.
What is the difference between Gemini 3.8 Flash and 3.7 Flash?
Price, context and max output are the same. In Google's table, 3.8 Flash scores higher on all 14 tests, with bigger jumps in long-horizon coding (73.7% against 65.3% on DeepSWE v1.1) and computer use (59.0% against 50.6% on OSWorld 2.0). In exchange, it spends more tokens per task. 3.7 Flash remains fully supported.
Is Gemini 3.8 Flash safe for customer support in languages other than English?
The model card says safety in non-English languages regressed slightly from 3.7 Flash, with 5.4 percentage points more on the automated multilingual test, where lower is better. That does not make the model unsafe, but it calls for testing sensitive cases in your language and adding your own moderation before putting it in front of the public.
Gemini 3.8 Flash or GPT-6 Luna?
GPT-6 Luna costs much less per token (US$ 0.10 and US$ 0.50) and answers up to 128 thousand tokens. 3.8 Flash takes video, audio and PDF as input, while Luna takes text and images, and it does not charge more for long prompts. No maker compared the two in the same table; the independent scores are at the top of this page.
Where does Gemini 3.8 Flash rank among AI models?
In the ROO3 AI Benchmark, ordered by the Epoch AI ECI, Gemini 3.8 Flash ranks #19 out of 267 models, with an ECI of 157.20. The ECI is a composite index that combines several hard benchmarks into a single scale. The position changes when a new model comes out, and the page is refreshed six times a day.
Is Gemini 3.8 Flash better than Kimi K3?
Across the 26 tests both models took and Epoch AI compiled, Gemini 3.8 Flash comes out ahead on 14, behind on 12 and tied on 0. It is a test-by-test count with no weighting: a win in math counts the same as a win in coding. To choose between the two, look at the tests that resemble your own use.
What is Gemini 3.8 Flash best at?
Against models of the same generation (released up to 9 months before and up to 3 months after it), the best relative results of Gemini 3.8 Flash are: DeepSWE (#2 of 26); ARC-AGI (#1 of 46); GPQA Diamond (#4 of 76). The yardstick is its position among the models of that period that took the same test.
Where do the numbers on this page come from?
From the same three sources as the ranking, all openly licensed: Epoch AI (ECI and individual tests, CC BY 4.0), Arena (human-preference ELO, CC BY 4.0) and LiveBench (overall average, CC BY-SA 4.0). ROO3 does not measure any model. When the page shows a figure published by the maker itself, it appears in the written analysis, labeled as the maker's, and never enters the ranking.
Where each number comes from
- Google, Gemini 3.8 Flash and 3.8 Flash Cyber announcement
- Google DeepMind, Gemini 3.8 Flash model card
- Google DeepMind, Gemini 3.8 Flash evaluation methodology (PDF)
- Gemini API, Gemini 3.8 Flash page
- Gemini API, latest model guide (reasoning and migration)
- Gemini API, pricing
- Google Cloud, Gemini 3.8 Flash page
- OpenAI, GPT-6 Luna page (documentation)
- OpenAI, API pricing
- OpenAI, GPT-6 Sol and GPT-6 Luna announcement
- OpenAI, API changelog
- Anthropic, Claude Opus 5.5 announcement
- Anthropic, model comparison in the Fable 5.1 documentation (Haiku 4.5)
- Anthropic, pricing (documentation)
Epoch AI, "AI Benchmarking Hub". Published at epoch.ai. Data under the CC BY 4.0 licence, reordered by roo3.co.
The Arena (LMArena) leaderboard, from the public lmarena-ai/leaderboard-dataset on Hugging Face, under the CC BY 4.0 licence. We are not affiliated with the Arena.
LiveBench, by the LiveBench team, published at livebench.ai under the CC BY-SA 4.0 license. The overall average is recomputed by roo3.co with the same formula as the website (average of the categories), and those numbers stay under CC BY-SA 4.0. We are not affiliated with LiveBench.
← See the full ranking