Gemini 3.8 Flash
ReleasedModel Gemini 3.8 Flash is a model by Google, released on 2 Sept 2026 (43 days ago).
In the Frontier SetGoogle positions this model in its top tier — role: Work & coding. “Our most intelligent Flash model, engineered for long-horizon software engineering, autonomous agents” Source: ai.google.dev · checked on 4 Oct 2026
- Where can it be used?
- Evidenced as usable in: Gemini, GitHub CopilotFor 12 more recorded products there is no statement; that does not mean “not available”.Available in the provider’s API
- Did we test it?
- No lab run published yet
- Last change
- 43 days ago
- Evidence
- 4 sources from 3 sitesRegistry entry:
checked 3 Oct 2026
What backs this up?
Three kinds of evidence, kept apart. None of them proves AGI on its own; where one is missing, the page says so.
- In real productsEvidenced as usable in: Gemini, GitHub Copilot
- agi.jetzt PrüfstandNo published run yet.
- Public benchmarks18 reported results on 15 benchmarks — each with its source
A model and the product around it are different things: a result for one does not carry over to the other.
Public benchmark results
Results reported by benchmark maintainers or providers, each with its source. They are not averaged and not compared across benchmarks; a provider result is self-reported. Some sources are data files of the benchmark’s own site rather than pages; they are marked as machine-readable.
- Gemini 3.8 Flash40.9 %Chartography · v1 · accuracy (expert-range graded), no toolssource as of 03/10/2026 · benchmark maintainer · mode not stated · judge: google/gemini-3.5-flash · tools: no · setting: Inspect AI (surge-ai/chartography) · attempts: 10 · web: no · reasoning: High · computer use: no · reported as „Gemini 3.8 Flash (High reasoning)“Source: surgehq.ai — Chartography, Gemini 3.8 Flash
- Gemini 3.8 Flash23.2 %GDP.pdf · v1 · pass rate, all criteria (strict)source as of 03/10/2026 · benchmark maintainer · mode not stated · judge: google/gemini-3.5-flash · tools: no · setting: Inspect AI (surge-ai/gdp-pdf) · attempts: 5 · web: no · reasoning: High · computer use: no · reported as „Gemini 3.8 Flash (High reasoning)“Source: surgehq.ai — GDP.pdf, Gemini 3.8 Flash
- Gemini 3.8 Flash19.09 %Terminal-Bench · 4.0 · resolution rate (mean of 5 runs)source as of 03/10/2026 · benchmark maintainer · model + harness · mini-SWE-agent · tools: yes · attempts: 5 · pass@k: 1 · reasoning: highSource: www.tbench.ai — Terminal-Bench, Gemini 3.8 Flash
- Gemini 3.8 Flash$5,093.79Vending-Bench 2 · 2 · final bank balancesource as of 03/10/2026 · benchmark maintainer · model + harness · Andon Labs Vending-Bench 2 agent environment (email, web search, inventory tools, notes/reminders) · tools: yes · budget: ~1 simulated year; 3000-6000 messages, 60-100M output tokens per run · attempts: 6 · web: yesMachine-readable source: andonlabs.com · data bundle of the website, not linked
- Gemini 3.8 Flash89.17 %ARC-AGI-2 · semi-private · accuracy (task pass rate)source as of 02/10/2026 · benchmark maintainer · mode not stated · reasoning: high (top tested) · reported as „Gemini 3.8 Flash (High)“Machine-readable source: arcprize.org · JSON — ARC-AGI-2, Gemini 3.8 Flash
- Gemini 3.8 Flash10.37 %ARC-AGI-3 · semi-private · scoresource as of 02/10/2026 · benchmark maintainer · model + harness · ARC Prize standard harness · reasoning: high (top tested) · reported as „Gemini 3.8 Flash (High)“Machine-readable source: arcprize.org · JSON — ARC-AGI-3, Gemini 3.8 Flash
- Gemini 3.8 Flash35 %ARC-AGI-3 · semi-private · score · provider adaptersource as of 02/10/2026 · benchmark maintainer · model + harness · ARC Prize provider adapter · reasoning: high (top tested) · reported as „Gemini 3.8 Flash - Provider Adapter (High)“Machine-readable source: arcprize.org · JSON — ARC-AGI-3, Gemini 3.8 Flash
- Gemini 3.8 Flash73.83 %DeepSWE · 1.1 · pass@1 (mean of 4 runs)source as of 22/09/2026 · benchmark maintainer · model + harness · mini-swe-agent · tools: yes · attempts: 4 · pass@k: 1 · reasoning: high · reported as „gemini-3-8-flash“Source: deepswe.datacurve.ai — DeepSWE, Gemini 3.8 Flash
- Gemini 3.8 Flash33.3 %HLE-Diamond · 2026-09-22 · accuracysource as of 22/09/2026 · benchmark maintainer · model alone · web: no · reasoning: highSource: lastexam.ai — HLE-Diamond, Gemini 3.8 Flash
- Gemini 3.8 Flash36.6 %HLE-Diamond · 2026-09-22 · accuracy · reasoning halfsource as of 22/09/2026 · benchmark maintainer · model alone · subset: reasoning partition (500 questions) · web: no · reasoning: highSource: lastexam.ai — HLE-Diamond, Gemini 3.8 Flash
- Gemini 3.8 Flash60.6 %HLE-Diamond · 2026-09-22 · accuracy · toolssource as of 22/09/2026 · benchmark maintainer · model + harness · Gemini CLI 0.60.0 · tools: yes · web: yes · reasoning: highSource: lastexam.ai — HLE-Diamond, Gemini 3.8 Flash
- Gemini 3.8 Flash44.52 %Humanity's Last Exam · 2025-04 · accuracyevaluated 09/09/2026 · benchmark maintainer · mode not stated · judge: o3-mini-2025-01-31 · temperature: 0.0 when configurableSource: labs.scale.com — Humanity's Last Exam, Gemini 3.8 Flash
Show 6 more
- Gemini 3.8 Flash82.4 %SimpleBench · avg5 · accuracy (average of 5 runs)evaluated 03/09/2026 · benchmark maintainer · model alone · pass@k: AVG@5 · temperature: 0.7 (top-p 0.95)Machine-readable source: simple-bench.com · JavaScript file — SimpleBench, Gemini 3.8 Flash
- Gemini 3.8 Flash86.2 %CharXiv (reasoning) · reasoning · accuracy, no toolssource as of 02/09/2026 · self-reported by the provider · mode not stated · tools: no · setting: Google DeepMind internal (self computed) · reasoning: Gemini default API samplingSource: deepmind.google — CharXiv (reasoning), Gemini 3.8 Flash
- Gemini 3.8 Flash21.95 %FrontierMath Tier 4 · v2 · accuracy (mean score)evaluated 02/09/2026 · benchmark maintainer · mode not stated · tools: yes · budget: 1,000,000 tokens hard limit (forced submission at 660,000) · setting: Epoch AI FrontierMath implementation (Inspect): python tool + submit_answer tool · web: no · reasoning: high · reported as „gemini-3.8-flash_high“Machine-readable source: epoch.ai · ZIP download — FrontierMath Tier 4, Gemini 3.8 Flash
- Gemini 3.8 Flash68.42 %FrontierMath Tiers 1–3 · v2 · accuracy (mean score)evaluated 02/09/2026 · benchmark maintainer · mode not stated · tools: yes · budget: 1,000,000 tokens hard limit (forced submission at 660,000) · setting: Epoch AI FrontierMath implementation (Inspect): python tool + submit_answer tool · web: no · reasoning: high · reported as „gemini-3.8-flash_high“Machine-readable source: epoch.ai · ZIP download — FrontierMath Tiers 1–3, Gemini 3.8 Flash
- Gemini 3.8 Flash87.1 %LVBench · v1 · accuracy (static), no toolssource as of 02/09/2026 · self-reported by the provider · mode not stated · tools: no · context: 1024 frames · setting: Google DeepMind internal (self computed)Source: deepmind.google — LVBench, Gemini 3.8 Flash
- Gemini 3.8 Flash41.19 %FrontierCode · 1.1-main · overall score (mergeability)historicalbenchmark maintainer · model + harness · chisel · subset: main · reasoning: mediumSource: cognition.com — FrontierCode, Gemini 3.8 Flash
Where can it be used?
Only recorded availability statements are shown. A missing plan or surface means unknown, not unavailable. Where it applies only to a plan or a region, the row says so. “Rolling out” is not “available”. “Rollout reported” means a rollout was reported, but nothing has supported it for more than 90 days.
| Product | Status | since | Checked | Sources |
|---|---|---|---|---|
| GeminiApp · AI Pro | Limited availability | not documented | ||
| GeminiApp · AI Ultra | Limited availability | not documented | ||
| GitHub CopilotCopilot · Business | Limited availability | not documented | checked 11 Oct 2026 | 1docs.github.com |
| GitHub CopilotCopilot · Enterprise | Limited availability | not documented | checked 11 Oct 2026 | 1docs.github.com |
| GitHub CopilotCopilot · Max | Limited availability | not documented | checked 11 Oct 2026 | 1docs.github.com |
| GitHub CopilotCopilot · Pro | Limited availability | not documented | checked 11 Oct 2026 | 1docs.github.com |
| GitHub CopilotCopilot · Pro Plus | Limited availability | not documented | checked 11 Oct 2026 | 1docs.github.com |
| GitHub Copilot | Limited availability | not documented | checked 11 Oct 2026 | 1docs.github.com |
Where a start is “not documented”, a source check found the state in place; the source names no start date.
No statement on record for 12 further products: ChatGPT, ChatGPT Work, Claude, Claude Code, Claude Enterprise, Claude Team, Codex, Gemini for Google Workspace, Grok, Langdock, Le Chat Enterprise, Microsoft 365 Copilot. No statement is not the same as not available.
Developer APIs
Developer access — not a statement about any work product.
| Product | Status | since | Checked | Sources |
|---|---|---|---|---|
| Gemini APIAPI | Available | not documented | checked 3 Oct 2026 | 1ai.google.dev |
What did we test?
No lab run of Gemini 3.8 Flash is published yet.
Nothing is shown until a real run exists. See what the lab tests:
Lifecycle and history
The dates the provider confirmed, and every recorded period of availability. A current period shows its checked status; one without a source check is marked not verified.
What changed?
Sourced changes concerning this model, newest first.
- 43 days agoadded to the record on 30 Sept 2026
1ai.google.dev
Evidence
Every source document behind the facts on this page, by domain.
| Domain | Docs | Documents |
|---|---|---|
| ai.google.dev | 2 | |
| blog.google | 1 | |
| docs.github.com | 1 |