Google Releases Gemini 3.1 Flash-Lite with 1M Context and Lower Token Pricing

Google’s new low-cost Gemini 3 tier

Google announced Gemini 3.1 Flash-Lite on March 3, 2026, positioning it as the fastest and most cost-efficient option in the Gemini 3 lineup. The release is explicitly aimed at large-scale workloads where organizations need predictable latency and aggressive cost control without falling back to very old model generations.

The model enters production channels immediately through Google AI Studio and Vertex AI. Google also states that a demo experience in the Gemini app is expected in the coming weeks, which indicates a dual strategy: enterprise API adoption first, broader product exposure second.

Key technical and pricing signals

Google highlights a 1 million-token context window, keeping long-context handling aligned with other Gemini 3 releases. For API users, the company also surfaces a configurable reasoning budget, allowing teams to trade off response depth against speed and spend according to workload profile.

Pricing is framed as a major part of the launch value. Google lists token rates at $0.10 per 1M input text/image/video tokens and $0.40 per 1M output text tokens. At this level, Flash-Lite is positioned for high-volume routing layers such as classification, extraction, constrained generation, and first-pass agent orchestration where per-token economics dominate architecture decisions.

Benchmark direction and deployment implications

Google reports that Gemini 3.1 Flash-Lite outperforms Gemini 2.5 Flash-Lite and Gemini 2.0 Flash-Lite on coding, math, science, and multimodal reasoning evaluations, and notes strong placement on LMArena among a large model set. While external validation remains important, the release message is clear: Google is trying to make the “default model” decision easier for teams that optimize for reliability at scale.

Operationally, this release gives builders a new middle path. Instead of choosing between expensive frontier inference and heavily compressed legacy models, teams can start with Flash-Lite for broad traffic, then route only difficult cases to heavier models. That architecture is often the fastest way to improve quality-adjusted cost in real deployments.