[Blog](https://brainbaselabs.com/blog) / [Model release](https://brainbaselabs.com/blog?topic=model-release) Sep 2, 2026·[Model release](https://brainbaselabs.com/blog?topic=model-release) # @google/gemini-3.8-flash is live on Brainbase Brainbase Team · 5 min read Gemini 3.8 Flash is now available on Brainbase as `@google/gemini-3.8-flash`. Use it with any supported harness and sandbox through the same Universal Managed Agents API you already use for every other model. Google released Gemini 3.8 Flash on September 2 — its third Flash model in six weeks — alongside a restricted 3.8 Flash Cyber variant for vetted security teams. It is a one-million-token, natively multimodal model that takes text, image, audio, and video input, with three selectable thinking levels and an introductory price of $0.75 per million input tokens through the end of the year. ## Why Gemini 3.8 Flash matters for agents Flash models used to be the tier you routed to when the task was easy. 3.8 Flash is the first one where that framing breaks down in specific verticals: on Google's published set it beats Opus 5 and GPT-5.6 Sol outright on Vals Finance Agent v2, Harvey's Legal Agent, and the hard split of BioMysteryBench, and it ties Opus 5 on Terminal-Bench 2.1 at roughly a seventh of the price per token. The other thing that changed is throughput. Artificial Analysis clocks it at around 300 output tokens per second, with a median task finishing in 2.5 minutes at high thinking and under a minute at low. For agents that fan out across many short tool calls — classification, extraction, document triage, browser steps — that speed is worth more than a few points on any single benchmark. ## Gemini 3.8 Flash benchmarks Google's comparison set puts 3.8 Flash ahead of Opus 5 on nine of the thirteen reported benchmarks, including domain agents, video understanding, and scientific reasoning. It trails on the ones that measure sustained autonomy: DeepSWE v1.1, OSWorld 2.0, GDPVal-AA v2, and — by a wide margin — Terminal-Bench 4.0. Select a metric below to compare the reported results. Gemini 3.8 Flash benchmark results - Gemini 3.8 Flash: 61.4% - Claude Opus 5: 58.6% - GPT-5.6 Sol: 53.8% Source: [Google, Introducing Gemini 3.8 Flash and 3.8 Flash Cyber](https://blog.google/innovation-and-ai/models-and-research/gemini-models/3-8-flash-and-3-8-flash-cyber/) Independent testing from [Artificial Analysis](https://artificialanalysis.ai/articles/gemini-3-8-flash) scores 3.8 Flash at 59 on its Intelligence Index at high thinking — three points over 3.7 Flash, level with GPT-5.6 Sol and Grok 4.6 — with the gain coming almost entirely from agentic evaluations. Its τ³-Banking tool-use score rose twelve points to 45%. Measured cost was $0.58 per Intelligence Index task at high thinking, $0.41 at medium, and $0.24 at low, against $0.84 for Grok 4.6 at a comparable score. ## Gemini 3.8 Flash pricing and context Google is running an introductory rate of half the list price through the end of 2026. Even at list, 3.8 Flash is a third of GPT-5.6 Sol's input rate and well under a third of Opus 5's. Gemini 3.8 Flash Price or limit Input tokens $0.75 / 1M tokens (list $1.50, from 2027) Output tokens $3.75 / 1M tokens (list $7.50, from 2027) Context window 1,000,000 tokens Max output 65,536 tokens Knowledge cutoff March 2026 Thinking levels low, medium, high Input modalities Text, image, audio, video These are model-token prices before sandbox runtime, tool calls, and other infrastructure usage. If you are budgeting past December, plan on the list rate. ## Where the Flash tier still trails The benchmark split is unusually clean, and it is worth being direct about. Where a task is bounded — answer a finance question with tools, reason over a chart, classify a video, run a short terminal session — 3.8 Flash is competitive with or better than the frontier models. Where a task requires the model to stay on course for hundreds of steps, it is not: 19.1% on Terminal-Bench 4.0 against Opus 5's 51.8%, 59% on OSWorld 2.0 against 75.4%, and nearly 280 Elo behind on GDPVal-AA v2. Terminal-Bench 2.1 and 4.0 measure the same skill at very different horizons, and the gap between 3.8 Flash's two scores — 89.4% on 2.1, 19.1% on 4.0 — is the clearest single number for where this model belongs. It is an excellent worker. It is not yet a planner. ## Run Gemini 3.8 Flash on Brainbase Select the model in the agent spec and leave the rest of your stack unchanged. This example runs Gemini 3.8 Flash inside the Cursor harness on a managed sandbox: POST /v2/threadsbash ``` curl https://api.brainbaselabs.com/v2/threads \ -H "Authorization: Bearer $BRAINBASE_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "agent": { "harness": "cursor", "model": "@google/gemini-3.8-flash" }, "input": "Triage the open issues labelled bug, reproduce each one, and reply with a ranked list." }' ``` ## Where Gemini 3.8 Flash fits Gemini 3.8 Flash is a strong default for high-volume, bounded agent work: document and data extraction, financial and legal analysis with tools, video and chart understanding, and the sub-agent roles inside a larger system where a frontier model plans and a cheaper one executes. It is the obvious replacement for 3.7 Flash everywhere, and a credible replacement for a frontier model in the verticals where Google reports it winning. Route it into the leaves of your agent tree first, then evaluate it against your production traces before giving it the root. Brainbase keeps the harness, sandbox, event stream, and observability layer constant so the model is the only variable. ## Related posts [Sep 2, 2026·Model release @meta/muse-spark-1.3 is live on Brainbase Brainbase Team · 5 min read](https://brainbaselabs.com/blog/meta-muse-spark-1-3-live-on-brainbase) [Sep 1, 2026·Model release @anthropic/claude-fable-5.1 is live on Brainbase Brainbase Team · 6 min read](https://brainbaselabs.com/blog/anthropic-claude-fable-5-1-live-on-brainbase) [Aug 26, 2026·Model release @zai/glm-5.3-flash is live on Brainbase Brainbase Team · 5 min read](https://brainbaselabs.com/blog/zai-glm-5-3-flash-live-on-brainbase) [Aug 12, 2026·Announcement Introducing Universal Managed Agents API Gokhan Egri · 7 min read](https://brainbaselabs.com/blog/universal-managed-agents) [Aug 12, 2026·Model release @spacexai/grok-4.6 is live on Brainbase Brainbase Team · 5 min read](https://brainbaselabs.com/blog/spacexai-grok-4-6-live-on-brainbase) [Aug 12, 2026·Model release @meta/muse-spark is live on Brainbase Brainbase Team · 5 min read](https://brainbaselabs.com/blog/meta-muse-spark-live-on-brainbase) [Jul 15, 2026·Model release @thinking-machines/inkling is live on Brainbase Brainbase Team · 4 min read](https://brainbaselabs.com/blog/thinking-machines-inkling-live-on-brainbase) [Jul 7, 2026·Announcement Brainbase is now generally available Gokhan Egri · 7 min read](https://brainbaselabs.com/blog/brainbase-is-now-ga) [View more posts →](https://brainbaselabs.com/blog)