Gemini 3.1 Flash-Lite Officially Released: Input price is only a quarter of Claude 4.5 Haiku, GPQA is nearly 14 percentage points higher

According to Beating Monitoring, Google Gemini 3.1 Flash-Lite has transitioned from a preview in March to an official release (GA). It is the most affordable and fastest model in the Gemini 3 series, now ready for high-concurrency production environments. The model comes standard with four levels of reasoning strength control (minimal, low, medium, high), allowing users to adjust between speed and quality based on the scenario.

Pricing remains at the preview level: $0.25 per million tokens for input, $1.50 per million tokens for output. Compared to competitors in the same tier, the input price is one-quarter of Claude 4.5 Haiku’s ($0.25 vs. $1.00), and the output price is less than one-third ($1.50 vs. $5.00); it is also cheaper than its own previous generation 2.5 Flash, with input dropping from $0.30 to $0.25 and output from $2.50 to $1.50. The context window is 1 million tokens.

Performance across levels: GPQA Diamond (graduate-level scientific reasoning) 86.9%, surpassing Claude 4.5 Haiku’s 73.0% and GPT-5 mini’s 82.3%; MMMU-Pro (multimodal understanding and reasoning) 76.8%, also leading competitors in the same tier. Output speed is 363 tokens/sec, 45% faster than 2.5 Flash, with the first token response speed twice as fast. Arena.ai leaderboard Elo score is 1432.

Several companies are already using it in production environments. Customer service platform Gladly employs Flash-Lite to power text-based AI agents, handling millions of customer interactions weekly, with costs about 60% lower than models of similar reasoning levels, a P95 latency of approximately 1.8 seconds, and a success rate of 99.6%. JetBrains uses it to drive IDE AI assistants and Junie agents. Financial operations platform Ramp utilizes it for high-frequency, latency-sensitive scenarios.

Programming remains a relative weakness for Flash-Lite, with LiveCodeBench at 72.0%, trailing GPT-5 mini’s 80.4%.