The Gemini 3.8 Flash model card is a document published by Google DeepMind in September 2026, alongside the September 2, 2026 release of Gemini 3.8 Flash, the third Flash point-update in the Gemini 3 series in six weeks (Source: https://deepmind.google/models/model-cards/gemini-3-8-flash/). The card describes the model as "based on Gemini 3.7 Flash" and refers to the Gemini 3.7 Flash model card for architecture, training data, data processing, hardware, software, safety-evaluation approach, safety policies and the Frontier Safety assessment. Its own content is a benchmark table, a table of automated safety-evaluation changes relative to Gemini 3.7 Flash, a statement on human red teaming, and a short frontier-safety statement. A companion page gives the evaluation methodology (Source: https://deepmind.google/models/evals-methodology/gemini-3-8-flash).
Model information and distribution
The card describes Gemini 3.8 Flash as "delivering performance advancements across software engineering and agentic knowledge workflows", with customizable effort levels, text, image, audio and video input with a context window of up to 1M tokens, and text output of up to 64K tokens. It lists six distribution channels: the Gemini app, the Gemini Enterprise Agent Platform, Google AI Studio, the Gemini API, Google AI Mode and Google Antigravity.
Capability evaluation
The results table, labelled "as of September, 2026", compares Gemini 3.8 Flash with Gemini 3.7 Flash, Claude Opus 5, Claude Sonnet 5, GPT-5.6 Sol and GPT-5.6 Terra across DeepSWE v1.1, GDPval-AA v2 (Elo), Vals Finance Agent v2, Harvey's Legal Agent Benchmark (all-pass rate), Terminal-Bench 2.1 and 4.0, GDP.PDF (all-pass rate), CharXiv Reasoning (no tools), LVBench, HLE-Verified, OSWorld-2.0 (partial score, batched tool call enabled), BioMysteryBench (human-solvable and human-difficult subsets) and LABBench2. It also lists prices: $0.75 and $3.75 per million input and output tokens for 3.7 and 3.8 Flash at the introductory rate, $1.50 and $7.50 regular; $5.00 and $25.00 for Claude Opus 5; $2.00 and $10.00 for Claude Sonnet 5; $4.00 and $20.00 for GPT-5.6 Sol; and $2.00 and $12.00 for GPT-5.6 Terra. In the table, Gemini 3.8 Flash has the highest figure on Vals Finance Agent v2, Harvey's Legal Agent Benchmark, Terminal-Bench 2.1, CharXiv Reasoning, LVBench, HLE-Verified, the BioMysteryBench human-difficult subset and LABBench2; Claude Opus 5 is highest on DeepSWE, GDPval-AA v2, Terminal-Bench 4.0, OSWorld-2.0 and the BioMysteryBench human-solvable subset; and GPT-5.6 Sol on GDP.PDF. LVBench is given for Gemini 3.8 Flash in two modes, "agentic" (87.8%) and "static" (87.1%).
The methodology page states that Gemini scores are pass@1 unless noted, that runs used the Gemini API model gemini-3.8-flash with default sampling, that smaller benchmarks are averaged over multiple trials, and that non-Gemini figures are the providers' self-reported numbers at maximum reasoning settings unless stated. Third-party rows are: DeepSWE from the Datacurve leaderboard at each model's highest-scoring thinking level (Gemini 3.8 Flash self-computed with a mini-swe-agent harness at high thinking), Terminal-Bench 2.1 comparison figures from the public leaderboard and Artificial Analysis under the Terminus 2 harness, Terminal-Bench 4.0 from its official leaderboard, GDPval-AA v2 from Artificial Analysis, and Vals Finance Agent v2 and Harvey's benchmark from Vals.AI. GDP.PDF, HLE-Verified and LABBench2 are self-computed for all models; CharXiv is self-computed for the Gemini, GPT-5.6 and Opus 5 rows; LVBench is self-computed without tools at 1,024 frames for Gemini and GPT-5.6 and 300 for Claude. OSWorld-2.0 for Gemini and Sonnet 5 is the maximum of three single-attempt runs with batched tool calls added to the harness for all models, run before the benchmark's August 8 patch; the GPT-5.6 Terra figure is taken from OpenAI's blog and the Opus 5 figure from Anthropic's Claude Fable 5.1 announcement. The methodology page adds that Claude Opus 5's DeepSWE score was "originally incorrectly reported as 74% due to rounding on the Datacurve public leaderboard"; the card table captured on September 21, 2026 still prints 74.0%. On HLE-Verified, Google notes that "a significant proportion of questions were blocked by content policy filters for Sonnet 5 and a small number for Opus 5".
Content safety
The card reports five internal automated safety evaluations as absolute percentage-point changes relative to Gemini 3.7 Flash: text-to-text safety (-0.4pp, lower is better), multilingual safety (+5.4pp, lower is better), image-to-text safety (0.0pp), tone of responses (+0.2pp, higher is better) and unjustified refusals (+1.1pp, lower is better). Google states that Gemini 3.8 Flash "performs similarly to Gemini 3.7 Flash across both safety and tone, with low unjustified refusals", that "safety performance across non-English languages regressed slightly relative to 3.7 Flash", that the results are not comparable with earlier cards because the evaluations were revised, and that manual review found losses were "overwhelmingly either a) false positives or b) not egregious". On human red teaming, the card says the model satisfied required launch thresholds for child safety, showed similar or improved safety performance compared with Gemini 3.7 Flash, and that red teaming beyond the strict policies, compared against Gemini 3.1 Pro, found no egregious concerns. No absolute rates are given.
Frontier safety
The card does not report a new Frontier Safety Framework evaluation. It states that Gemini 3.7 Flash was evaluated under the April 2026 framework and did not reach any tracked or critical capability level, that Google's assessments show Gemini 3.8 Flash "does not have meaningful new capabilities or material increases in performance" in the framework's domains relative to 3.7 Flash, and that on that basis Google is "confident that Gemini 3.8 Flash is also unlikely to reach any T/CCLs". Readers are referred to the 3.7 Flash card.
Reported results
Evaluation results the document states for which there is no separate benchmark page; results on benchmarks that have a page are recorded there and cite this page. Scores and settings are as printed in the document.
| Evaluation | Model | Score | Setting | Section |
|---|---|---|---|---|
| DeepSWE v1.1 | Gemini 3.8 Flash | 73.7% | long-horizon software engineering; Gemini 3.8 Flash self-computed with mini-swe-agent harness, high thinking; other models from the Datacurve public leaderboard at highest-scoring thinking level; developer-reported | Evaluation > Results (as of September 2026) |
| DeepSWE v1.1 | Gemini 3.7 Flash | 65.3% | long-horizon software engineering; Gemini 3.8 Flash self-computed with mini-swe-agent harness, high thinking; other models from the Datacurve public leaderboard at highest-scoring thinking level; developer-reported | Evaluation > Results (as of September 2026) |
| DeepSWE v1.1 | Claude Opus 5 | 74.0% | long-horizon software engineering; Gemini 3.8 Flash self-computed with mini-swe-agent harness, high thinking; other models from the Datacurve public leaderboard at highest-scoring thinking level; Datacurve public leaderboard, as reprinted in the Gemini 3.8 Flash model card (methodology page notes the 74% was a rounding artefact of the leaderboard) | Evaluation > Results (as of September 2026) |
| DeepSWE v1.1 | Claude Sonnet 5 | 53.8% | long-horizon software engineering; Gemini 3.8 Flash self-computed with mini-swe-agent harness, high thinking; other models from the Datacurve public leaderboard at highest-scoring thinking level; Datacurve public leaderboard, as reprinted in the Gemini 3.8 Flash model card | Evaluation > Results (as of September 2026) |
| DeepSWE v1.1 | GPT-5.6 Sol | 72.7% | long-horizon software engineering; Gemini 3.8 Flash self-computed with mini-swe-agent harness, high thinking; other models from the Datacurve public leaderboard at highest-scoring thinking level; Datacurve public leaderboard, as reprinted in the Gemini 3.8 Flash model card | Evaluation > Results (as of September 2026) |
| DeepSWE v1.1 | GPT-5.6 Terra | 69.6% | long-horizon software engineering; Gemini 3.8 Flash self-computed with mini-swe-agent harness, high thinking; other models from the Datacurve public leaderboard at highest-scoring thinking level; Datacurve public leaderboard, as reprinted in the Gemini 3.8 Flash model card | Evaluation > Results (as of September 2026) |
| Harvey's Legal Agent Benchmark | Gemini 3.8 Flash | 10.0% | all-pass rate; complex legal workflows; sourced from Vals.AI; developer-reported | Evaluation > Results (as of September 2026) |
| Harvey's Legal Agent Benchmark | Gemini 3.7 Flash | 8.8% | all-pass rate; complex legal workflows; sourced from Vals.AI; developer-reported | Evaluation > Results (as of September 2026) |
| Harvey's Legal Agent Benchmark | Claude Opus 5 | 6.7% | all-pass rate; complex legal workflows; sourced from Vals.AI; Vals.AI, as reprinted in the Gemini 3.8 Flash model card | Evaluation > Results (as of September 2026) |
| Harvey's Legal Agent Benchmark | Claude Sonnet 5 | 5.0% | all-pass rate; complex legal workflows; sourced from Vals.AI; Vals.AI, as reprinted in the Gemini 3.8 Flash model card | Evaluation > Results (as of September 2026) |
| Harvey's Legal Agent Benchmark | GPT-5.6 Sol | 2.5% | all-pass rate; complex legal workflows; sourced from Vals.AI; Vals.AI, as reprinted in the Gemini 3.8 Flash model card | Evaluation > Results (as of September 2026) |
| Harvey's Legal Agent Benchmark | GPT-5.6 Terra | 0.8% | all-pass rate; complex legal workflows; sourced from Vals.AI; Vals.AI, as reprinted in the Gemini 3.8 Flash model card | Evaluation > Results (as of September 2026) |
| GDP.PDF | Gemini 3.8 Flash | 35.0% | all-pass rate; expert PDF document comprehension; all models self-computed by Google; developer-reported | Evaluation > Results (as of September 2026) |
| GDP.PDF | Gemini 3.7 Flash | 34.0% | all-pass rate; expert PDF document comprehension; all models self-computed by Google; developer-reported | Evaluation > Results (as of September 2026) |
| GDP.PDF | Claude Opus 5 | 37.0% | all-pass rate; expert PDF document comprehension; all models self-computed by Google; self-computed by Google DeepMind, reported in the Gemini 3.8 Flash model card | Evaluation > Results (as of September 2026) |
| GDP.PDF | Claude Sonnet 5 | 28.0% | all-pass rate; expert PDF document comprehension; all models self-computed by Google; self-computed by Google DeepMind, reported in the Gemini 3.8 Flash model card | Evaluation > Results (as of September 2026) |
| GDP.PDF | GPT-5.6 Sol | 40.0% | all-pass rate; expert PDF document comprehension; all models self-computed by Google; self-computed by Google DeepMind, reported in the Gemini 3.8 Flash model card | Evaluation > Results (as of September 2026) |
| GDP.PDF | GPT-5.6 Terra | 29.0% | all-pass rate; expert PDF document comprehension; all models self-computed by Google; self-computed by Google DeepMind, reported in the Gemini 3.8 Flash model card | Evaluation > Results (as of September 2026) |
| LVBench | Gemini 3.8 Flash | 87.8% | agentic; long video understanding; self-computed without tools, 1024 frames; developer-reported | Evaluation > Results (as of September 2026) |
| LVBench | Gemini 3.8 Flash | 87.1% | static; long video understanding; self-computed without tools, 1024 frames; developer-reported | Evaluation > Results (as of September 2026) |
| LVBench | Gemini 3.7 Flash | 85.4% | long video understanding; self-computed by Google without tools; 1024 frames for Gemini and GPT-5.6, 300 frames for Claude (API limitation); developer-reported | Evaluation > Results (as of September 2026) |
| LVBench | Claude Opus 5 | 75.4% | long video understanding; self-computed by Google without tools; 1024 frames for Gemini and GPT-5.6, 300 frames for Claude (API limitation); self-computed by Google DeepMind, reported in the Gemini 3.8 Flash model card | Evaluation > Results (as of September 2026) |
| LVBench | Claude Sonnet 5 | 68.5% | long video understanding; self-computed by Google without tools; 1024 frames for Gemini and GPT-5.6, 300 frames for Claude (API limitation); self-computed by Google DeepMind, reported in the Gemini 3.8 Flash model card | Evaluation > Results (as of September 2026) |
| LVBench | GPT-5.6 Sol | 82.1% | long video understanding; self-computed by Google without tools; 1024 frames for Gemini and GPT-5.6, 300 frames for Claude (API limitation); self-computed by Google DeepMind, reported in the Gemini 3.8 Flash model card | Evaluation > Results (as of September 2026) |
| LVBench | GPT-5.6 Terra | 78.9% | long video understanding; self-computed by Google without tools; 1024 frames for Gemini and GPT-5.6, 300 frames for Claude (API limitation); self-computed by Google DeepMind, reported in the Gemini 3.8 Flash model card | Evaluation > Results (as of September 2026) |
| BioMysteryBench — Human Solvable | Gemini 3.8 Flash | 88.8% | human-solvable subset; Linux terminal with bioinformatics tools, Python, R, and allow-listed internet; Gemini and GPT-5.6 self-computed; developer-reported | Evaluation > Results (as of September 2026) |
| BioMysteryBench — Human Solvable | Gemini 3.7 Flash | 87.1% | human-solvable subset; Linux terminal with bioinformatics tools, Python, R, and allow-listed internet; Gemini and GPT-5.6 self-computed; developer-reported | Evaluation > Results (as of September 2026) |
| BioMysteryBench — Human Solvable | Claude Opus 5 | 90.1% | human-solvable subset; Linux terminal with bioinformatics tools, Python, R, and allow-listed internet; Gemini and GPT-5.6 self-computed; provider self-reported, as reprinted in the Gemini 3.8 Flash model card | Evaluation > Results (as of September 2026) |
| BioMysteryBench — Human Solvable | Claude Sonnet 5 | 87.5% | human-solvable subset; Linux terminal with bioinformatics tools, Python, R, and allow-listed internet; Gemini and GPT-5.6 self-computed; provider self-reported, as reprinted in the Gemini 3.8 Flash model card | Evaluation > Results (as of September 2026) |
| BioMysteryBench — Human Solvable | GPT-5.6 Sol | 79.5% | human-solvable subset; Linux terminal with bioinformatics tools, Python, R, and allow-listed internet; Gemini and GPT-5.6 self-computed; self-computed by Google DeepMind, reported in the Gemini 3.8 Flash model card | Evaluation > Results (as of September 2026) |
| BioMysteryBench — Human Solvable | GPT-5.6 Terra | 83.8% | human-solvable subset; Linux terminal with bioinformatics tools, Python, R, and allow-listed internet; Gemini and GPT-5.6 self-computed; self-computed by Google DeepMind, reported in the Gemini 3.8 Flash model card | Evaluation > Results (as of September 2026) |
| BioMysteryBench — Human Difficult | Gemini 3.8 Flash | 56.5% | human-difficult subset; same terminal/tool setting; Gemini and GPT-5.6 self-computed; developer-reported | Evaluation > Results (as of September 2026) |
| BioMysteryBench — Human Difficult | Gemini 3.7 Flash | 43.5% | human-difficult subset; same terminal/tool setting; Gemini and GPT-5.6 self-computed; developer-reported | Evaluation > Results (as of September 2026) |
| BioMysteryBench — Human Difficult | Claude Opus 5 | 49.4% | human-difficult subset; same terminal/tool setting; Gemini and GPT-5.6 self-computed; provider self-reported, as reprinted in the Gemini 3.8 Flash model card | Evaluation > Results (as of September 2026) |
| BioMysteryBench — Human Difficult | Claude Sonnet 5 | 34.1% | human-difficult subset; same terminal/tool setting; Gemini and GPT-5.6 self-computed; provider self-reported, as reprinted in the Gemini 3.8 Flash model card | Evaluation > Results (as of September 2026) |
| BioMysteryBench — Human Difficult | GPT-5.6 Sol | 44.7% | human-difficult subset; same terminal/tool setting; Gemini and GPT-5.6 self-computed; self-computed by Google DeepMind, reported in the Gemini 3.8 Flash model card | Evaluation > Results (as of September 2026) |
| BioMysteryBench — Human Difficult | GPT-5.6 Terra | 49.4% | human-difficult subset; same terminal/tool setting; Gemini and GPT-5.6 self-computed; self-computed by Google DeepMind, reported in the Gemini 3.8 Flash model card | Evaluation > Results (as of September 2026) |
| Text to Text Safety — vs. Gemini 3.7 Flash | Gemini 3.8 Flash | -0.4pp | absolute change vs. Gemini 3.7 Flash; automated content-safety policy evaluation; lower is better; automated internal evaluation, not human red teaming; not comparable with earlier cards; developer-reported | Ethics and Content Safety > Training and Development Evaluation Results |
| Multilingual Safety — vs. Gemini 3.7 Flash | Gemini 3.8 Flash | +5.4pp | absolute change vs. Gemini 3.7 Flash; automated safety policy evaluation across languages; lower is better (a regression); automated internal evaluation, not human red teaming; not comparable with earlier cards; developer-reported | Ethics and Content Safety > Training and Development Evaluation Results |
| Image to Text Safety — vs. Gemini 3.7 Flash | Gemini 3.8 Flash | 0.0pp | absolute change vs. Gemini 3.7 Flash; automated content-safety policy evaluation; lower is better; automated internal evaluation, not human red teaming; not comparable with earlier cards; developer-reported | Ethics and Content Safety > Training and Development Evaluation Results |
| Tone — vs. Gemini 3.7 Flash | Gemini 3.8 Flash | +0.2pp | absolute change vs. Gemini 3.7 Flash; automated evaluation of objective tone of responses; higher is better; automated internal evaluation, not human red teaming; not comparable with earlier cards; developer-reported | Ethics and Content Safety > Training and Development Evaluation Results |
| Unjustified-refusals — vs. Gemini 3.7 Flash | Gemini 3.8 Flash | +1.1pp | absolute change vs. Gemini 3.7 Flash; automated evaluation on borderline prompts; lower is better; automated internal evaluation, not human red teaming; not comparable with earlier cards; developer-reported | Ethics and Content Safety > Training and Development Evaluation Results |
Provenance
The page summarises the HTML card as captured on September 21, 2026, together with its methodology page. The card links a PDF version (Source: https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-3-8-Flash-Model-Card.pdf). All figures are reported by Google DeepMind; comparison-model figures are mostly the other developers' own published numbers or public leaderboards, selected by Google. Confidence is medium: a single developer-authored source, with fast-decay benchmark claims.
Relationships
- supports: Gemini 3 / Gemini 3 Pro
- depends-on: Gemini 3.7 Flash Model Card (Google DeepMind, August 2026)
- instance-of: AI System Cards
- related: Google DeepMind, AI Safety Cases and Frameworks, Terminal-Bench, GDPval, Finance Agent, OSWorld, Humanity's Last Exam, LAB-Bench, Claude Opus 5, Claude Sonnet 5, GPT-5.6 (Sol, Terra, Luna)