Gemini 3.7 Flash — adjustable thinking and its benchmarks

Google has released Gemini 3.7 Flash, built on 3.6 Flash. Everything below comes from Google's own model card, published 13 August 2026, rather than from third-party estimates.

Adjustable thinking

The main practical change is what Google calls customisable thinking configurations: you choose the mix of quality, cost and speed for a task. A quick translation or summary does not need the same effort as a long piece of code, and you can set the model accordingly rather than switching to a different one.

Google does not publish a specific thinking-token range or a response-time figure for this, so we do not quote one.

Context and price

• Context: up to 1 million tokens of input and 64,000 tokens of output.
• API price: $0.75 per million input tokens and $3.75 per million output tokens.
• Where it runs: the Gemini app, Gemini Enterprise, AI Studio, the Gemini API and Antigravity.

The benchmark results

These are the numbers Google publishes on the model card:

• DeepSWE v1.1 (long-horizon software engineering): 65.3%
• Terminal-bench 2.1 (command-line tasks): 85.8%
• GDM-MRCR v2 (finding 8 items across 128k tokens of text): 97.0%
• LVBench (long video understanding): 85.4%
• Humanity's Last Exam, verified set: 53.6%

Google does not publish a SWE-bench Verified or an AIME score for this model, so we do not quote one. Figures for it on other sites are worth checking against the model card before you rely on them — we had two here ourselves until we did.

Getting it in Ethiopia

For consumers, Gemini comes with Google's AI plans — Google AI Pro is the one most people want. Google bills those to an international card, which many birr cards are not enabled for. We set up Gemini Pro on your own Google account with payment in birr by Telebirr or bank transfer.

Written by the Bezuale Solutions team in Addis Ababa. Questions? Message us on Telegram.