Source: 9to5Google — 2026-08-13
Summary
Google shipped Gemini 3.7 Flash on August 13, just three weeks after Gemini 3.6 Flash, describing it as an algorithmic-improvement refresh on the same base model rather than a from-scratch retrain, aimed at coding, web development, and agent workflows. DeepSWE v1.1 rose from 49.0% to 65.3%, AutomationBench climbed from 17.0% to 30.4%, and WebDev Arena Elo increased from 1538 to 1588. Launch pricing is $0.75 per million input tokens and $3.75 per million output tokens — half of 3.6 Flash's launch price. The refresh lands while Google's flagship Gemini 3.5 Pro remains undelivered past its original June 2026 target, leaving the Flash tier to carry the company's near-term coding and agent story.
Key Takeaways
- Gemini 3.7 Flash shipped only three weeks after 3.6 Flash; Google frames the gains as coming from algorithmic/post-training improvements on the same base model, not a new pretraining run.
- DeepSWE v1.1, a software-engineering task benchmark, jumped from 49.0% to 65.3% — a 16.3-point gain.
- AutomationBench, which measures agentic workflow automation, nearly doubled from 17.0% to 30.4%.
- WebDev Arena Elo rating rose from 1538 to 1588, a 50-point gain in head-to-head web-dev preference matchups.
- Launch pricing of $0.75 / $3.75 per million input/output tokens is exactly half of 3.6 Flash's launch price, so the newer, more capable model is also the cheaper one to call.
- The release comes as Gemini 3.5 Pro, Google's flagship model, remains undelivered past its original June 2026 promise — Flash-tier refreshes are filling the gap in the meantime.
Reel Script
Hook: Google just nearly doubled a key coding-agent benchmark and cut the API price in half — in three weeks flat. If your team built cost estimates around last month's Gemini Flash pricing, that spreadsheet is already out of date, and not in your favor.
Core Concept: Here's the part that matters: Google says this isn't a new model trained from zero. It's the same underlying Gemini 3.6 Flash base, refreshed through what Google calls algorithmic improvements — think of it as retuning the engine's software rather than building a new engine. That distinction is why it shipped in three weeks instead of the usual months-long training cycle: no new pretraining run, just better post-training and inference-time techniques layered on top of an existing model. It's the same playbook other labs have used this year to squeeze big benchmark jumps out of models they already had sitting around, and it's cheaper to execute than a full retrain — which is likely part of why the price dropped instead of rising. The bigger context here is what this release is covering for: Gemini 3.5 Pro, the actual flagship model, was promised for June 2026 and still hasn't shipped. So Google is iterating hard on the Flash tier — the smaller, cheaper model line — to keep its coding and agent story competitive while the bigger model stays in the oven.
Hands-On: Three numbers tell the story. DeepSWE v1.1, which scores a model on realistic software-engineering tasks, went from 49.0% to 65.3%. AutomationBench, which tests whether a model can carry out multi-step automated workflows, nearly doubled from 17.0% to 30.4% — that's the one to watch if you're building agents that click through UIs or chain tool calls. And WebDev Arena Elo, a head-to-head preference score for web-development output, climbed from 1538 to 1588, a meaningful 50-point jump in a ranking system where small differences separate the top models. Then there's the pricing table: 3.6 Flash launched at roughly double today's rate, and 3.7 Flash now costs $0.75 per million input tokens and $3.75 per million output tokens — so you're paying half as much for a model that's clearing agent and coding benchmarks by double-digit margins.
Takeaway: A three-week refresh that nearly doubles agent benchmark scores while halving the price is a genuinely good deal if you're running Flash-tier workloads today — but it's also a tell that Google is leaning hard on iteration speed at the cheap tier because the expensive flagship model is late. If you're on 3.6 Flash, re-run your own eval suite against 3.7 before you migrate production traffic — vendor benchmark deltas and your workload's deltas aren't always the same number.