Google ships Gemini 3.7 Flash, its third new Flash model in three months, with a 4-point independent score jump

[Google](#/company/google) shipped Gemini 3.7 Flash on Aug. 13, the company's third new Flash-tier model in roughly three months and, per Google's own framing, its "most intelligent workhorse model yet for coding and agents." [Artificial Analysis](#/scoreboard), the independent benchmarking group this newsroom's Scoreboard relies on, scored the high-reasoning-effort variant at 56 on its Intelligence Index — four points above Gemini 3.6 Flash and enough to land the model on what the firm calls the Intelligence-vs-Time Pareto frontier, meaning nothing scoring higher finishes tasks faster.

56 (Artificial Analysis Intelligence Index score, Gemini 3.7 Flash (high)) That score places the model just behind the two current leaders in its class, GPT-5.6 Terra (max) and Muse Spark 1.2 (xhigh), both at 57 — a one-point gap Google is closing with price rather than a bigger jump in reasoning score. Introductory pricing runs $0.75 per million input tokens and $3.75 per million output tokens, half the standard rate, through Dec. 31, 2026; list price then doubles to $1.50/$7.50 on Jan. 1, 2027. Artificial Analysis separately measured output speed at roughly 340 tokens per second, against a same-tier median of 74.8 — the basis for its second claim, that discounted Gemini 3.7 Flash also sits on the Intelligence-vs-Cost frontier.

Availability is broad from day one: the model is live in Google AI Studio and the Gemini API, inside Android Studio and Google Antigravity for developers, on the Gemini Enterprise Agent Platform, and in Gemini Spark for Google AI Pro and Ultra subscribers across more than 160 countries. That's a wider simultaneous rollout than a typical preview-first launch, and it matches the positioning: Google is not framing 3.7 Flash as an experiment to be validated slowly, but as the default coding/agent tier it wants developers on immediately, price included.

[Google](#/company/google) ships Gemini 3.7 Flash in three reasoning-effort tiers rather than one — low, medium, and high — each independently scored by Artificial Analysis at 51, 53, and 56 respectively, all sharing the same list price. That's a different trade than most labs offer at this tier: instead of a single fixed model, a developer picks how much reasoning effort (and therefore latency) a given task actually needs, without paying a separate per-tier price for the privilege. A high-volume, low-complexity workload can run at "low" and still land close to the field's median score for comparably priced models; a harder task can call "high" and get the full four-point gain over Gemini 3.6 Flash.

Where Google says the gains actually are

Google's own release numbers show the jump concentrated in coding and business-agent tasks rather than general knowledge. On FrontierCode 1.1, a coding benchmark, the model moved from 34.4% to 43.6%; on DeepSWE v1.1, an agentic software-engineering test, from 49.0% to 65.3%. AutomationBench, which grades multi-step business-workflow completion, nearly doubled from 17.0% to 30.4%. Document comprehension (GDP.pdf) rose from 22.0% to 34.0%, and the model's WebDev Arena Elo — a head-to-head ranking of generated web apps — moved from 1538 to 1588. Those are Google's figures, not independently reproduced numbers — Artificial Analysis's own agent benchmarks show a smaller, though still real, gain: 60% on AA-AnalystAgent and 62.7% on AutomationBench-AA.

The pattern is consistent across every metric Google published: the widest gains are all on agentic and coding tasks, the exact workloads Google's own positioning names. General-knowledge and reasoning gains, the kind the independent Intelligence Index aggregates across coding, reasoning, and knowledge evaluations together, moved by a comparatively modest four points. Read together, the two sets of numbers tell a consistent story rather than a conflicting one: a model tuned hardest for the coding-agent workload it's priced to win, with a smaller, but independently confirmed, gain in general capability.

Artificial Analysis frames Gemini 3.7 Flash (high) as sitting on two frontiers at once — intelligence-vs-time and, with the discounted rate applied, intelligence-vs-cost — meaning no model both scores higher and costs less per task. That's a stronger claim than a plain benchmark win: it says Google isn't just catching up on the Intelligence Index, it's catching up while also being cheap and fast enough that a buyer doesn't have to trade one for the others. The one-point gap to GPT-5.6 Terra and Muse Spark 1.2 is real, but it now has to be weighed against a list price roughly a third of what a max-effort frontier model typically commands — the kind of trade a cost-sensitive coding-agent deployment is built to make.

This is the third Gemini Flash release in roughly three months, following 3.5 Flash and 3.6 Flash earlier in the cycle — a cadence Artificial Analysis itself flagged when it published the score. Google is not the only lab iterating this fast at the low-cost tier, but the pattern of shipping a new Flash model roughly monthly, each time with a fresh independent score rather than a vendor claim standing alone, is now distinct enough to be its own story: at this pace, the number that matters for a buyer isn't just where a given model lands, but how quickly its replacement arrives.