Gemini 3.8 Flash
Gemini 3.8 Flash is a multimodal model in Google's Gemini family. Google DeepMind released it on September 2, 2026 as a generally available model for software engineering, tool-using agents, and knowledge work. Its stable Gemini API identifier is gemini-3.8-flash.[1][2][3]
The model accepts text, images, video, audio, and PDF files and produces text. It has a 1,048,576-token input limit and a 65,536-token output limit. Google describes it as an iteration of Gemini 3.7 Flash, rather than disclosing a separate architecture for the release.[2][6]
Release and position in the Gemini family
Gemini 3.8 Flash followed Gemini 3.7 Flash by three weeks. Google described it as the company's third Flash release within six weeks. The September 2026 launch introduced two variants that share a common model foundation: the generally available Gemini 3.8 Flash and Gemini 3.8 Flash Cyber, a cybersecurity variant restricted to vetted defenders through Google DeepMind's Fairwind program.[1]
The standard model was released as a stable API model, not a preview. Google retained Gemini 3.7 Flash for workloads where lower token use was more important than the additional reasoning and tool calls that 3.8 Flash may use on difficult tasks.[1][3]
Google's model card says Gemini 3.8 Flash is based on Gemini 3.7 Flash. It refers readers to the earlier model's documentation for architecture, training data, data processing, hardware, and software details. The 3.8 card does not state a parameter count, training-corpus size, training-compute figure, or new architecture specific to this release.[6]
Model specification
| Property | Documented value at release |
|---|---|
| Model code | gemini-3.8-flash |
| API status | Stable and generally available |
| Inputs | Text, image, video, audio, and PDF |
| Output | Text |
| Input-token limit | 1,048,576 |
| Output-token limit | 65,536 |
| Thinking levels | Low, medium, and high |
| Default thinking level | Medium |
| Knowledge cutoff | March 2026, with some domains potentially limited to January 2025 |
The exact token limits come from the Gemini API model page. The model card rounds them to a one-million-token context window and a 64K-token output limit. The card's knowledge-cutoff statement is qualified: it says users can expect updated information in some areas, while knowledge in other areas may be limited to January 2025.[2][6]
API capabilities
Inputs, tools, and output controls
The Gemini API lists context caching, code execution, file search, function calling, Google Search grounding, Google Maps grounding, structured outputs, and URL context as supported. Computer use is supported in preview. The model can also run through the Batch API, Flex inference, and Priority inference.[2]
Gemini 3.8 Flash does not generate audio or images through the documented model endpoint, and it does not support the Gemini Live API. Its ability to interpret audio, images, and video is an input capability rather than a media-generation capability.[2]
Thinking controls and token use
Gemini 3.8 Flash uses dynamic thinking by default at the medium level. Developers can choose low, medium, or high with the API's thinking_level control. The minimal level is not supported and returns an error. Google recommends lower effort for latency-sensitive or routine tasks and higher effort for requests that benefit from more reasoning and tool use.[2][5]
Thought summaries can expose a summary of the model's reasoning, but billing is based on all thought tokens the model generates, not only the summary returned to the caller. Google also warns that the model may use more tokens on complex tasks, particularly at higher effort levels.[1][5][6]
Availability
At launch, developers could access Gemini 3.8 Flash through the Gemini API and Google AI Studio, as well as through Android Studio, Google Antigravity, and Stitch. Google listed Gemini Enterprise for organizational access. Its model card also names the Gemini Enterprise Agent Platform as a distribution channel.[1][6]
For consumers, Google made the model available to Google AI Pro and Ultra subscribers in the Gemini app, AI Mode in Google Search, and Gemini in Google Sheets. Availability through one product does not imply that every API feature or tool is exposed through that product's interface.[1]
API pricing
Google introduced Gemini 3.8 Flash at a temporary paid rate matching Gemini 3.7 Flash. The following prices apply through December 31, 2026. Token figures are prices per one million tokens in US dollars; cache storage is priced per one million tokens per hour.[1][4]
| Processing mode | Input | Output, including thinking tokens | Cached input | Cache storage |
|---|---|---|---|---|
| Standard | $0.75 | $3.75 | $0.075 | $0.50 |
| Batch | $0.375 | $1.875 | $0.0375 | $0.50 |
| Flex | $0.375 | $1.875 | $0.0375 | $0.50 |
| Priority | $1.35 | $6.75 | $0.135 | $0.50 |
Google's published schedule raises those rates on January 1, 2027:[4]
| Processing mode | Input | Output, including thinking tokens | Cached input | Cache storage |
|---|---|---|---|---|
| Standard | $1.50 | $7.50 | $0.15 | $1.00 |
| Batch | $0.75 | $3.75 | $0.075 | $1.00 |
| Flex | $0.75 | $3.75 | $0.075 | $1.00 |
| Priority | $2.70 | $13.50 | $0.27 | $1.00 |
The pricing page lists standard usage on the free tier as free of token charges, subject to its limits, while Batch and Flex are unavailable there. It also states that free-tier content may be used to improve Google's products and paid-tier content is not used for that purpose. Search and Maps grounding can add separate query charges after the shared Gemini 3.x monthly allowances.[4]
Evaluation
Methodology and comparability
Google published an evaluation-methodology document alongside the model card. It says Gemini results were obtained through the Gemini API with default sampling unless a benchmark note specified otherwise. Results are pass@1 unless noted, and Google averaged multiple trials on smaller benchmarks. Several results for competing models came from their providers rather than a single common test run, so cross-model comparisons do not all share one evaluation environment.[7]
DeepSWE v1.1 consists of 113 original, long-horizon software-engineering tasks drawn from 91 active open-source repositories and five programming languages. Each task uses a purpose-built verifier rather than tests inherited from a previously merged change.[9] Google said Gemini 3.8 Flash outperformed most larger frontier models on the benchmark, but the launch article did not state a numerical score in its text. Google's result was self-computed with the mini-swe-agent harness and high thinking, while comparator results came from the public leaderboard at their highest reported thinking levels.[1][7]
Reported results
| Evaluation | Gemini 3.8 Flash result | Source and qualification |
|---|---|---|
| HLE-Verified | 54.9% | Google-computed accuracy on the full 1,811-item verified set[1][7] |
| Vals Finance Agent v2 | 61.4% | Vals.AI result reproduced by Google; Gemini 3.7 Flash was shown at 59.0%[7][8] |
| DeepSWE v1.1 | No numerical score stated in Google's launch text | Google said it exceeded most larger frontier models; its 3.8 Flash run was self-computed[1][7] |
HLE-Verified is a revised version of Humanity's Last Exam whose authors audited the questions and answers and separated uncertain items from the verified set.[10] Google's methodology says its reported result used all 1,811 verified or revised items and excluded 689 items identified as uncertain. The comparison should not be treated as identical to results calculated on the original HLE set.[7][10]
The launch material also presented results for legal work, terminal tasks, multimodal reasoning, long context, computer use, and scientific reasoning. Many were first-party runs, differently configured public-leaderboard results, or provider-reported comparator scores. They are not included here without the full table and benchmark-specific setup needed to interpret them.[6][7]
Safety evaluation
Automated comparisons
Google reported internal automated safety comparisons against Gemini 3.7 Flash. These are changes in percentage points, not absolute rates. Lower values are better for the three safety and refusal measures, while higher is better for tone.[6]
| Evaluation | Change from Gemini 3.7 Flash | Direction in Google's metric |
|---|---|---|
| Text-to-text safety | -0.4 percentage points | Lower is better |
| Multilingual safety | +5.4 percentage points | Lower is better |
| Image-to-text safety | 0.0 percentage points | Lower is better |
| Tone | +0.2 percentage points | Higher is better |
| Unjustified refusals | +1.1 percentage points | Lower is better |
The largest reported regression was multilingual safety at 5.4 percentage points. Google said manual review found that the losses it inspected were overwhelmingly false positives or not egregious. Because Google also changed parts of its automated evaluations, the card cautions that these results are not directly comparable with figures in earlier Gemini model cards.[6]
Red teaming and frontier assessment
Google said specialist teams outside the model-development group conducted manual red teaming. The company reported that Gemini 3.8 Flash met its child-safety launch thresholds and that the broader review found no egregious concerns.[6]
The model card does not report a new full Tracked or Critical Capability Level assessment for 3.8 Flash. Google instead cited the April 2026 evaluation of Gemini 3.7 Flash, which did not reach a tracked or critical level, and stated that 3.8 Flash showed no meaningful new capability or material increase in the Frontier Safety Framework domains. On that basis, Google judged 3.8 Flash unlikely to reach those thresholds.[6]
Limitations
The model card lists hallucinations, occasional slowness, and timeouts among the model's known limitations. Higher thinking levels can increase both time to the first non-thinking output token and billed token use. A large context limit therefore does not guarantee that every long-document or long-running task will be accurate, fast, or inexpensive.[5][6]
The published evaluation evidence also has limits. Google computed several results itself, used different harnesses across tasks, and relied on provider-reported numbers for many competing models. Benchmark scores measure performance under those stated conditions and do not establish reliability for every deployment.[7]
Gemini 3.8 Flash Cyber should not be treated as an interchangeable endpoint. It uses more permissive cybersecurity mitigations and is restricted to approved defenders, while the standard Gemini 3.8 Flash retains Google's ordinary safeguards against cyber misuse and chemical, biological, radiological, and nuclear misuse.[1]
References
- ^Google. "Introducing Gemini 3.8 Flash and 3.8 Flash Cyber." September 2, 2026. blog.google/...3-8-flash-and-3-8-flash-cyber
- ^Google AI for Developers. "Gemini 3.8 Flash." Updated September 2, 2026. ai.google.dev/...gemini-3.8-flash
- ^Google AI for Developers. "Gemini API release notes." September 2, 2026 entry. ai.google.dev/...changelog
- ^Google AI for Developers. "Gemini Developer API pricing." Accessed September 3, 2026. ai.google.dev/...pricing
- ^Google AI for Developers. "Gemini thinking." Accessed September 3, 2026. ai.google.dev/...thinking
- ^Google DeepMind. "Gemini 3.8 Flash - Model Card." Published September 2, 2026. deepmind.google/...gemini-3-8-flash
- ^Google DeepMind. "Gemini 3.8 Flash: Model evaluation approach, methodology and results." September 2026. deepmind.google/...gemini-3-8-flash
- ^Google DeepMind. "Gemini 3.8 Flash." Accessed September 3, 2026. deepmind.google/...flash
- ^Wenqi Huang, Charley Lee, Leonard Tng, and Serena Ge. "DeepSWE: Measuring Frontier Coding Agents on Original, Long-Horizon Engineering Tasks." arXiv, July 8, 2026. arxiv.org/...2607.07946
- ^Weiqi Zhai et al. "HLE-Verified: A Systematic Verification and Structured Revision of Humanity's Last Exam." arXiv, February 15, 2026. arxiv.org/...2602.13964
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation. Every suggestion is reviewed for sourcing before it goes live.
v1 · 1,875 words · full history
Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify
Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here
Reviewer note: Independently checked against cited and current primary sources on 2026-09-03.
Cite this page: AI Wiki. "Gemini 3.8 Flash." aiwiki.ai, updated 3 Sept 2026, fact-checked 3 Sept 2026. CC BY 4.0. https://aiwiki.ai/wiki/gemini_3_8_flash