Gemini 3.7 Flash: a cheap agent model with a real long-horizon jump
DeepSWE agent scores jump from 49% to 65%; the introductory price holds through Dec 31 before rates double, and two rival Flash models still cost less.
Copy markdown
The leap is in long-horizon agent work
Gemini 3.7 Flash posts 65.3% on DeepSWE v1.1 (up from 49%), 43.6% on FrontierCode 1.1, and nearly doubles AutomationBench to 30.4% — the multi-step, keep-your-place work budget models usually choke on. It keeps the 1M-token context and 64K-token output ceiling.
Intro pricing holds until Dec 31, then doubles
Standard rates are $0.75/M input and $3.75/M output through year-end; on Jan 1, 2027 they double to $1.50/$7.50. Cache reads stay cheap at $0.075/M. If you're building on it now, budget for the January step-up.
Batch and Flex halve it again
Async jobs on the Batch or Flex tier pay $0.375/M input and $1.875/M output — half the standard rate. That's the lane for overnight agent runs and bulk processing where latency doesn't matter.
Two Flash rivals still win on raw price
Picking on cost alone, GPT-5.6 Luna ($0.20/$1.20) and DeepSeek V4-Flash ($0.14/$0.28) undercut it; Claude Haiku 4.5 ($1.00/$5.00) sits above. Gemini's pitch is finishing long chains reliably, not the lowest sticker.
The API is global; the Spark app isn't
Developer access via the Gemini API and Google AI Studio is available everywhere, but the Spark consumer app excludes the EEA, UK, Switzerland and Nigeria. Shipping a consumer front-end in those regions means routing through the API.