Google releases Gemini 3.8 Flash at $0.75/$3.75 per million tokens

• Gemini 3.8 Flash ships as Google's new Flash workhorse at $0.75 per million input tokens and $3.75 per million output tokens, the rate in Google's Gemini 3.8 announcement and Google's pricing post.
• Google's comparison card shows meaningful long-horizon gains over 3.7 Flash, including DeepSWE, Terminal-Bench 2.1, and OSWorld-2.0, in _philschmid's benchmark card.
• Flat token pricing masks a larger high-effort bill: ArtificialAnlys's cost analysis puts completed-task cost at $0.58, about 40% above 3.7 Flash, after more output tokens and agent turns.
• Gemini 3.8 Flash Cyber arrives as a separate defender-only model, with Google's Flash Cyber introduction reporting 86.2% on CyberGym discovery and 47.2% on CWE-Bench patching.

Google's launch post pairs a 1M-token general model with a gated cyber variant. The Cloud migration guide makes the trade explicit: 3.8 can spend more tokens than 3.7 to raise reliability, while its effort control can reduce that overhead.

• Gemini 3.8 Flash is Google's third Flash release in six weeks, positioned for software engineering, agents, and multi-step reasoning in lmthang's release post.
• The introductory $0.75 input and $3.75 output rates run through December 31, 2026, according to Google's availability list.
• Developers get it through Antigravity, Google AI Studio, Android Studio, and the Gemini API, while Gemini Enterprise is the enterprise surface in Google's availability list.
• Google AI Pro and Ultra subscribers get it in Gemini App, AI Mode in Search, and Gemini in Sheets, as GeminiApp's availability post confirms.
• Gemini 3.8 Flash Cyber is limited to trusted defenders through Fairwind, a companion launch from GoogleDeepMind's release post.

Google describes Flash Cyber as a specialized vulnerability discovery and automated-patching model, rather than a public API sibling of the general Flash model.

• Google reports 86.2% on CyberGym vulnerability discovery, 47.2% on CWE-Bench patching, and more than 70% success on its internal discovery benchmark across 20 programming languages in Google's benchmark post.
• Google's internal results say Chrome Security saw 2.6 times more correct vulnerability patches than from larger commercial models, while Wiz reported 7.5% to 9.7% higher recall at 2.3 to 5.2 times lower cost in Google's partner-results post.
• The Fairwind Program is a limited-access channel for Google Cloud customers, government agencies, and cybersecurity partners. Google says it initially targets national cyber authorities and operators of essential services.

First-party

• DeepSWE v1.1: 65.3% → 73.7%, +8.4 points, in _philschmid's comparison card.
• Terminal-Bench 2.1: 85.8% → 89.4%, +3.6 points, in _philschmid's comparison card.
• OSWorld-2.0: 50.6% → 59.0%, +8.4 points, in _philschmid's comparison card.

Third-party evaluators

• Artificial Analysis Intelligence Index, high reasoning: 56 → 59, +3 points, according to ArtificialAnlys's release analysis.
• Agent Arena net improvement: +0.84% → +5.94%, +5.10 points, in arena's five-signal breakdown.
• Vals Index: 59.3% → 62.3%, +3.0 points, in ValsAI's lineage chart.

Customer-reported

• Chrome vulnerability fixes: best commercial-model baseline → 2.6 times as many correct patches, +160%, per GoogleDeepMind's Chrome result.
• Wiz penetration-test recall: leading-model baseline → 7.5% to 9.7% higher recall, +7.5% to +9.7%, in Google's report of Wiz's result.

Higher reasoning effort changes the unit economics. Artificial Analysis measured high-effort 3.8 Flash at $0.58 per Intelligence Index task, versus $0.40 for 3.7 Flash, and attributed the roughly 40% increase to a 30% rise in output tokens and more agent turns in its cost analysis. Its time per task rose from 2.2 minutes to 2.5 minutes, despite output speed near 300 tokens per second in its latency comparison.

ValsAI found an 8.3-point Vibe Code Bench improvement, with one zero-score app out of 50 instead of five. The same analysis reported a 14.2-point SAGE decline and a 10-point ProofBench decline, with harsher math grading and 10 fewer Lean proofs, respectively.

Agent Arena also still identifies steerability as a relative weakness: it improved from -5.38% to -1.43%, but remained below its baseline in arena's comparison. One same-day run from theo's tool-call test reported more than 100 tool calls and over 10 minutes for a four-line integration change.

The Cloud model reference lists the production ID as gemini-3.8-flash, with a 1,048,576-token context window, 65,536 maximum output tokens, text output, and text, image, audio, and video inputs. It also lists structured output, system instructions, context caching, and thinking support.

Google frames the capability change as longer-running agentic loops. At high effort, the model takes smaller steps and verifies more often, according to _philschmid's launch notes; low, medium, and high reasoning levels are separately measured in ArtificialAnlys's token-usage comparison.

The promotional price expires at year-end. Artificial Analysis reports standard list rates of $1.50 per million input tokens and $7.50 per million output tokens from January 1, 2027, while cached input retains a 90% discount.

Claim: Unchanged token rates preserve Flash's cost advantage for agentic work. Cited by: Google's pricing announcement says the model keeps 3.7 Flash's rate, while Agent Arena measured $0.22 per task with a 44% to 50% lower price than comparable Grok 4.5 and GLM 5.2 Max results. Counter: ArtificialAnlys's task-cost analysis measured 3.8 Flash High at $0.58 per task, about 40% above 3.7 Flash High. Evidence so far: the two results use different task suites and effort settings, but both report that completed-task cost changes even when the token price does not.

Early hands-on reports split between fast, capable general use and agents that overwork trivial changes.

• osanseviero described several days of testing as well-rounded across finance, legal, and multimodal work in an early hands-on.
• gabriberton wrote that a week of use brought one "silly mistake," several surprises, and fast responses in a week-long test.
• Simon Willison reported a 13-second, 1.8-cent HTML experiment and used the model with his basic coding-agent plugin to add sandboxed HTML rendering to a Markdown SVG tool in his release notes.
• Google's m__dehghani solicited concrete reports of odd behavior in a feedback request, while theo's integration test documented the opposite end of the spectrum.

• Cursor added Gemini 3.8 Flash on launch day in Cursor's update.
• Nous Portal added it for Hermes Agent, according to Teknium's portal announcement.
• LiteLLM's day-zero support routes the model through existing Google AI Studio and Vertex AI configurations, without a Docker image upgrade.
• Google's 3.7 Flash build gallery already spans real-time website generators, 3D physics simulators, interactive webcam tools, and personalized field guides in Google's app gallery, all across surfaces that now list 3.8 Flash as available.

Google's agentic-video API release is a separate capability with a concrete tool loop: a model can inspect transcripts, fetch selected frame ranges at an adaptive frame rate, pull audio, then iterate before answering.

The API example uses processing="agentic" on an individual video input, as

shows. Google's release names Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite as the available models for that feature, while its benchmark claims compare 3.7 Flash static processing with agentic processing.

The published comparison reports up to 88% fewer tokens, 66% lower cost, and up to 7% better accuracy for the agentic mode in GoogleAIStudio's launch post. Those figures are a result for the earlier Flash models named in that announcement, not a published 3.8 Flash API measurement.