Gemini 3.8 Flash: Google's Third Budget Model in Six Weeks — Is It Worth Switching?

2026-09-04 · 2 min read

googlegeminimodel-releasecomparison

What's new

Google released Gemini 3.8 Flash on September 2, 2026 — just three weeks after 3.7 Flash. That makes it the third budget-tier model in six weeks (there's also a cybersecurity-focused 3.8 Flash Cyber variant), during a stretch in which Google's frontier Pro models have gone quiet.

Pricing holds at 3.7 Flash's introductory level: $0.75 per million input tokens and $3.75 per million output, rising to $1.50 / $7.50 from January 2027.

The number Google leads with is DeepSWE v1.1, its software-engineering benchmark: 3.8 Flash scores 73.7%, up from 65.3% for 3.7 Flash and within touching distance of Claude Opus 5's 74.0%. The catch is what it costs to get there — cost per task rises from $0.40 to $0.58 (+40%), because the model reasons longer and makes more tool calls. On speed, the high-reasoning setting streams around 300 output tokens per second and averages about 2.5 minutes per task; the low-reasoning setting returns results in roughly 48 seconds.

Where you can try it

Gemini 3.8 Flash is not on AIWITH.CHAT yet — we'd rather say so than pretend. What you can use today is Gemini 3.1 Pro (the current Pro flagship, default version of our Gemini robot), plus Gemini 3.6 Flash and Gemini 3 Pro in the same dropdown — alongside GPT-6 Astra, Claude Fable 5.1, DeepSeek V4.1 and Kimi K3, all under one $9.9/month plan. When 3.8 Flash lands we'll update this post; one subscription means you don't pay extra when a new model arrives.

See what's available now →

Is it worth switching?

If you use 3.7 Flash for code or engineering tasks: yes, the jump from 65.3% to 73.7% is worth having — especially since it puts a budget model at the Opus 5 threshold. But go in with your eyes open on cost: at +40% per task, a batch pipeline's bill will change noticeably.

If you use it for light chat and summaries: the difference is hard to feel, and there's no reason to keep swapping models for a budget tier that iterates every three weeks.

The practical move is to check whether Google's own benchmark actually matches your task type, then — once it's on a platform you already use — run the same prompt through 3.8 Flash and Claude Opus 5 side by side, rather than deciding on a benchmark score alone. For most everyday work, Gemini 3.1 Pro on a multi-model plan already covers the ground 3.8 Flash is aiming at.