Google shipped Gemini 3.8 Flash on September 2, 2026, and the pitch is narrow and deliberate: this is the most capable Flash-tier model Google has built for mid-difficulty, long-horizon software engineering and autonomous agents, without giving up Flash speed or raising the price. It succeeds Gemini 3.7 Flash and lands at the same introductory rate, which reframes it less as a new frontier model and more as a cheaper default for builders who run a lot of agent steps.
What Google shipped
Gemini 3.8 Flash is a multimodal model that accepts text, images, audio, and video, carries a context window of up to 1M tokens, and produces up to 64K tokens of output. Its knowledge cutoff is March 2026. Google positions it squarely at software engineering and agentic knowledge workflows rather than at raw reasoning records, and it is available immediately across the Gemini API, Google AI Studio, the Gemini app, AI Mode, the Enterprise Agent Platform, and Google Antigravity.
For anyone building with it, the fastest way in is Google AI Studio, where you can point it at a repository, hand it a multi-file task, and watch how it plans across many tool calls. The model exposes customizable effort levels, so you can dial reasoning depth up for a hard refactor or down for a quick edit and pay accordingly.
Gemini 3.8 Flash vs the coding field
The honest read on the benchmarks is that this is a refinement, not a leap. On DeepSWE v1, a hard long-horizon software-engineering test, Gemini 3.8 Flash scored 71.0%, behind Anthropic Opus 5 at 74.0% and OpenAI GPT-5.6 Sol at 72.7%. What makes that number interesting is the price it is attached to: a frontier-adjacent coding score at a fraction of frontier cost. The table below sets it against its predecessor and two current coding rivals. Benchmarks are drawn from different tests and are not directly comparable cell to cell, so read each row on its own terms.
| Model | Context | Input $/M | Output $/M | Notable coding result |
|---|---|---|---|---|
| Gemini 3.8 Flash | 1M | $0.75 | $3.75 | DeepSWE v1 71.0% |
| Gemini 3.7 Flash | 1M | $0.75 | $3.75 | Terminal-bench 4.0 11.2% |
| Qwen3.8-Max-0902 | 1M | $2.00 | $6.00 | CodeArena WebDev #1 (1,691) |
| Opus 5 (reference) | varies | higher | higher | DeepSWE v1 74.0% |
| GPT-5.6 Sol (reference) | varies | higher | higher | DeepSWE v1 72.7% |
If your work already runs on open weights, the same week gives you a local-friendly point of comparison in GLM-5.3 Flash and the DeepSeek and Qwen coding tier, where the tradeoff is self-hosting and control against Google-hosted convenience and the 1M window. Alibaba's Qwen3.8-Max-0902, which topped CodeArena's web-development track the same day, is the closer hosted rival, though it costs more than double per token and does not match Flash speed.
The multimodal input matters more than the coding scores suggest. Because 3.8 Flash reads images, audio, and video natively, a build agent can take a screenshot of a broken UI, a Figma export, or a screen recording of a bug and act on it inside the same loop that writes the fix. For creators building apps and sites, that collapses the usual round trip of describing a visual problem in text before an agent can touch it.
What actually improved over 3.7 Flash
The gains are concentrated in agent execution rather than one-shot answers. On Terminal-bench 4.0, a test of running real terminal tasks to completion, 3.8 Flash reached 19.1% versus 11.2% for 3.7 Flash. On OSWorld-2.0, which measures agents driving a desktop environment, it moved to 59.0% from 50.6%. Those are the numbers that matter if you are chaining dozens of tool calls, because small per-step reliability gains compound across a long run.
Early testers also flag a quality-of-life change: 3.8 Flash is noticeably less verbose than earlier Flash models, a persistent complaint that inflated token bills and cluttered agent transcripts. Tighter output is not a headline benchmark, but in an agent loop it is the difference between a clean plan and a wall of hedging.
The pricing angle is the real story
Gemini 3.8 Flash keeps the same introductory pricing as 3.7 Flash: $0.75 per million input tokens and $3.75 per million output tokens, a rate Google says holds through the end of 2026 before reverting to $1.50 and $7.50. For an agent that reads a large codebase and emits many steps, that intro rate against a 71% DeepSWE score is the sharpest cost-to-capability ratio in the current coding tier. The catch is the expiry date, so any budget you model on this price should assume the doubled rate in January.
How to put it to work this week
Start by swapping 3.8 Flash into an existing agent config where 3.7 Flash or a pricier frontier model was doing routine coding work: bug triage, test generation, dependency bumps, and small refactors. Point it at a real repository through AI Studio or the API, give it a multi-step task, and set the effort level low first to see how far the cheaper setting gets you. Reserve high effort and a frontier model for the genuinely hard, ambiguous tasks where the DeepSWE gap actually shows.
The pattern most builders will land on is a router: 3.8 Flash as the cheap, fast default for the long tail of agent steps, with an escalation to Opus 5 or GPT-5.6 Sol only when a step fails or a task is flagged high-stakes. At this price, running 3.8 Flash as the workhorse and paying frontier rates only on exceptions is the move that keeps a coding-agent bill sane.
A concrete starting workflow: wire 3.8 Flash into an agent that scaffolds a small web app end to end, generating the component tree, styling it, writing tests, and iterating against terminal output. The OSWorld and Terminal-bench gains are exactly the reliability that makes this kind of unattended, multi-step build finish cleanly rather than stall halfway. Log the token spend on a real project for a week, then compare it against what the same run would cost on a frontier model to see where the escalation threshold should sit for your workload.
Frequently asked questions
Is Gemini 3.8 Flash available now?
Yes. It went live on September 2, 2026 across the Gemini API, Google AI Studio, the Gemini app, AI Mode, the Enterprise Agent Platform, and Google Antigravity.
How much does Gemini 3.8 Flash cost?
The introductory rate is $0.75 per million input tokens and $3.75 per million output tokens, held through the end of 2026. Regular pricing after that is $1.50 and $7.50 per million.
Is it better than Gemini 3.7 Flash for coding?
Yes, especially for agents. It improved from 11.2% to 19.1% on Terminal-bench 4.0 and from 50.6% to 59.0% on OSWorld-2.0, and it produces less verbose output at the same price.
How does it compare to Opus 5 and GPT-5.6 Sol?
On DeepSWE v1 it scores 71.0%, behind Opus 5 at 74.0% and GPT-5.6 Sol at 72.7%. It trades a few points of hard-task accuracy for a much lower price and Flash-tier speed.
What context window and output does it support?
Up to 1M tokens of context and up to 64K tokens of output, with a March 2026 knowledge cutoff and multimodal input across text, images, audio, and video.
Who should switch to it?
Builders running high-volume coding agents where per-step cost dominates. It is best deployed as a cheap default with an escalation path to a frontier model for the hardest tasks.
