Gemini 3.8 Flash ships: 90.8% Terminal-Bench, $0.75/1M input
Google's fourth Flash in four months undercuts GPT-5.6 Sol on price, leaps on coding and agentic tool-use, and lands in AI Studio and the API today.
Copy markdown
Cheaper than the frontier, live today
Gemini 3.8 Flash lists at $0.75/$3.75 per 1M input/output tokens (introductory, through Dec 31; $1.50/$7.50 after), with cached inputs 90% off — roughly 6x under GPT-5.6 Sol's $5/$30. It's live now in Google AI Studio, the Gemini API, and the Enterprise Agent Platform as `gemini-3.8-flash`, with a 1M-token context window.
Coding is where it actually moved
Terminal-Bench 2.1 jumps to 90.8% from 3.7 Flash's 81.6%, and SWE-Bench Pro ticks up to 61.6%. On long-horizon software tasks it reportedly beats most larger frontier models — near-flagship coding at Flash prices for your agent loops.
A 12-point jump in agentic tool-use
The single biggest gain was agentic: +12 points on banking/tool-use benchmarks over 3.7 Flash. Its Intelligence Index hits 59 at high reasoning (+3), matching GPT-5.6 Sol's sub-max effort and Grok 4.6's medium — at ~300 tokens/sec and roughly $0.58 per task.
The caveat: raw reasoning stayed flat
Humanity's Last Exam barely moved (45.4% vs 45.7%). This is a coding-and-agents refresh, not a reasoning leap — reach for it in tool-heavy pipelines, not open-ended problem solving.
Elsewhere: a gated cyber sibling
Gemini 3.8 Flash Cyber, restricted to Google's Fairwind Program, claims a 70%+ real-world vulnerability-discovery rate and sits on the CWE-Bench patching frontier. It's defense-first and off the open API — you can't call it today.