Gemini 3.8 Flash
- Provider
- Status
- Current
- Context
- 1,000,000 tok
- Price
- $0.75 / $3.75 /MTok
Gemini 3.8 Flash is Google’s fourth Flash-tier release in under four months, shipped on 2 September 2026 alongside a gated cybersecurity variant, Gemini 3.8 Flash Cyber. It is the cheapest model near the frontier: $0.75/$3.75 per million tokens on an introductory rate that runs to 31 December 2026 ($1.50/$7.50 thereafter), for a model Artificial Analysis scores 59 on its Intelligence Index at high reasoning — three points up on Gemini 3.7 Flash, above Claude Opus 4.8 (56), and one point under the Kimi K3 and GLM-5.3 pair (60).
The startling number is on agentic coding: on DeepSWE v1.1, OpenAI’s own GPT-6 Astra launch table prints Gemini 3.8 Flash at 73.8% — within half a point of Astra (74.1%) and Claude Opus 5 (73.7%) — from a model priced at roughly one-thirteenth of Astra and one-seventh of Opus 5. If it holds up in practice, Flash-class agentic coding at frontier-adjacent quality is the week’s most economically significant release, more than either flagship.
Quick specs
| Provider | |
| Tier | Flash — value line, flagship-adjacent scores |
| Released | 2 September 2026 |
| Status | Available — Gemini API, AI Studio, Vertex AI |
| Context window | 1,000,000 tokens |
| Modalities | Text, image, video and speech in; text out |
| Input price | $0.75 / MTok intro (list $1.50, from 1 Jan 2027) |
| Output price | $3.75 / MTok intro (list $7.50) |
| AA Intelligence Index | 59 high reasoning (57 medium, 52 low) |
| AA cost per task | $0.58 high / $0.41 medium / $0.24 low |
| Output speed | ~300 tokens/sec (AA) |
| SWE-bench | Not published |
| Best for | Agentic coding and tool use on a budget; high-volume workloads |
| Limitations | No SWE-bench; intro price expires 31 Dec; Cyber variant gated |
What actually changed
The independent score crossed into flagship territory. AA’s 59 at high reasoning is a three-point gain over 3.7 Flash and lands above models that cost five to seven times more per token. AA measures ~300 output tokens per second and $0.58 per index task at high reasoning — against $2.34 for Opus 5.
Agentic work is where Google aimed it. The launch post claims wins over 3.7 Flash and “other frontier models” on Vals Finance Agent V2 and Harvey’s Legal Agent Benchmark, and AA’s largest measured gain is on τ³-Banking, up 12 points to 45%. The DeepSWE 73.8% — printed by OpenAI, not Google — is the strongest third-party-adjacent corroboration a vendor claim gets in launch week.
Reasoning level is a price dial. Low (52 on AA’s index, $0.24/task), medium (57, $0.41), high (59, $0.58) — the spread means the same model spans value and flagship use cases depending on a parameter.
Flash Cyber joins the gated-cyber club. Google reports frontier-level vulnerability detection — a real-world discovery success rate above 70% across 20 programming languages, and 47.2% pass@1 on CWE-Bench patching against a leading frontier model’s 47.8% — available only to vetted defenders through the new Fairwind Program. It launched the same week OpenAI rated GPT-6 Astra Critical for cyber and gated the capability behind Daybreak Blue: both frontier labs now decide who is allowed a cyber model, not whether to build one.
How Gemini 3.8 Flash compares
vs GPT-6 Astra and the frontier
On measured intelligence Astra leads 61 to 59; on DeepSWE they are within half a point on OpenAI’s own table; on price they are separated by a factor of thirteen. For agentic coding workloads that do not need Astra’s computer-use or long-context strengths, 3.8 Flash is the value proposition of the week. See best AI models for placement.
vs Kimi K3 and GLM-5.3
The 59-vs-60 gap to the leading Chinese pair is inside the index’s noise, and all three compete on price: GLM-5.3 at $1.40/$4.40, Kimi K3 at $3/$15, Flash at $0.75/$3.75 intro. Flash is the cheapest and the only one with video and speech input; the other two publish open weights (K3) or fuller benchmark tables.
vs Gemini 3.5 Flash
Three Flash generations in four months separate them — 3.6 and 3.7 shipped in between without a board placement here. The 3.8 jump is the one that matters: AA scores 3.7 Flash at 56 and 3.8 at 59, and 3.8 adds the agentic gains and the Cyber variant.
Known limitations
No SWE-bench figure of any kind. The DeepSWE and agent-benchmark numbers are vendor-reported (or printed by a competitor); nothing independent on coding beyond the AA index components exists yet.
The intro price is a deadline. $0.75/$3.75 doubles to $1.50/$7.50 on 1 January 2027. Cost-per-task comparisons made today overstate its January economics.
Cyber is gated. Flash Cyber goes through Fairwind vetting; the general model does not carry those capabilities.
Speed claims meet reasoning reality. ~300 tokens/sec is fast, but AA’s time-per-task at high reasoning is still 2.5 minutes — reasoning depth, not throughput, dominates latency on hard tasks.
Frequently asked questions
Is Gemini 3.8 Flash better than GPT-6 Astra?
No on measured intelligence — Artificial Analysis scores Astra 61 and 3.8 Flash 59 — and Astra leads clearly on computer use and long context. But on DeepSWE v1.1, OpenAI’s own launch table separates them by half a point (74.1% vs 73.8%), and 3.8 Flash costs roughly one-thirteenth as much. For budget agentic coding it is the stronger buy; we rank Astra #6 and 3.8 Flash #11 on our board.
How much does Gemini 3.8 Flash cost?
$0.75 per million input tokens and $3.75 per million output on an introductory rate that expires 31 December 2026; the standard rate is $1.50/$7.50. Cached input is discounted 90%. Artificial Analysis measures $0.58 per completed task at high reasoning, $0.24 at low.
What is Gemini 3.8 Flash Cyber?
A cybersecurity variant Google reports at frontier level for vulnerability detection (over 70% real-world discovery success across 20 programming languages) and automated patching (47.2% pass@1 on CWE-Bench). It is not generally available: access runs through Google’s new Fairwind Program for vetted defenders, mirroring OpenAI’s Daybreak gating of GPT-6 Astra’s cyber capability the same week.
What is the context window?
1,000,000 tokens, unchanged from Gemini 3.7 Flash, with text, image, video and speech input and text output.
Does Gemini 3.8 Flash have a SWE-bench score?
No — Google published none. The nearest coding evidence is DeepSWE v1.1 at 73.8% as printed in OpenAI’s GPT-6 Astra launch table, and the coding components inside Artificial Analysis’s Intelligence Index.
Last verified 4 September 2026. Headline figures are Google-reported from the 2 September 2026 launch post and vendor-run unless marked independent; Intelligence Index, speed and cost-per-task figures are Artificial Analysis measurements; the DeepSWE figure is as printed in OpenAI’s GPT-6 Astra launch table. Pricing and availability subject to change.