Google shipped Gemini 3.7 Flash on August 13, just three weeks after its predecessor, and the model’s combination of low price and improved tool-calling is drawing attention well beyond typical AI circles. For the growing field of on-chain trading bots and autonomous Web3 agents — software that increasingly needs to call smart contract functions, parse block explorer data and execute multi-step transactions without a human in the loop — pricing and reliable tool use matter more than poetic prose.
The model is priced at $0.75 per million input tokens and $3.75 per million output tokens through the end of the year, roughly half of what Gemini 3.6 Flash launched at, before the rate doubles to $1.50/$7.50 on January 1, 2027. At that introductory rate, it undercuts Anthropic’s Claude Sonnet 5 ($2/$10 per million) and OpenAI’s GPT-5.6 Terra ($2/$12 per million) by a wide margin — the kind of gap that matters for agent frameworks making thousands of tool calls per trading session.
What Actually Improved
Google says the model was built specifically around coding and agent workflows, and Decrypt’s own testing backs that up on structured tasks: Gemini 3.7 Flash produced a playable browser game in 2 minutes and 13 seconds, on the first attempt, with clean syntax — a task the prior Flash model failed at entirely. On Google’s own benchmarks, the model scored 1,588 Elo on Code Arena’s web development leaderboard and 30.4% on AutomationBench, with the company claiming wins over Claude Sonnet 5 and GPT-5.6 Terra in 11 of 18 tested categories.
According to Google’s own announcement, the model shows measurable gains on coding benchmarks like FrontierCode 1.1 Main (43.6% versus 34.4% for the prior version) and DeepSWE v1.1 (65.3% versus 49.0%), alongside better multi-step planning and more diligent tool calls — precisely the capabilities an automated agent executing a DeFi strategy or monitoring a wallet needs to get right.
Where It Still Falls Short
The gains are narrower once tasks move away from structured execution. In testing that involved genuine reasoning rather than pattern completion, the model stumbled: it solved a degree-19 polynomial problem correctly up to the final step, then returned the answer as an unevaluated expression instead of a number, and it produced the same wrong answer to a classic logic puzzle that an earlier Claude model also got wrong months ago. On a creative-writing task with a hard structural constraint, it lost out to a free 27-billion-parameter community model.
Why the Price Matters More Than the Prose
None of that weaker reasoning performance is disqualifying for the use case crypto builders care about most. An agent executing a pre-defined DeFi yield strategy, monitoring on-chain events, or drafting boilerplate smart contract code doesn’t need to win a creative-writing contest — it needs to call the right function with the right parameters, cheaply, at scale. On that narrower but commercially important axis, Gemini 3.7 Flash’s combination of aggressive pricing and improved tool-calling gives Web3 agent developers a meaningfully cheaper option than the two most established alternatives, at least until the introductory pricing expires at the end of the year.
