Google ships Gemini 3.8 Flash: 73.7% on DeepSWE, but 'working harder' burns tokens

Google is shipping Gemini 3.8 Flash just three weeks after 3.7 Flash landed, making this the third budget-model release in six weeks. The pace is either impressive or a sign that the frontier tier is stalled, depending on who you ask. DeepMind’s new chief, Koray Kavukcuoglu, insists Google still wants to lead on raw capability, not just price-performance. But the flagship Gemini 3.5 Pro and Gemini 4 remain conspicuously absent, so the Flash churn is doing a lot of narrative work right now.

The numbers are genuinely strong for a budget model. Google reports 73.7 percent on DeepSWE v1.1, a long-horizon software engineering benchmark. That lands just behind Claude Opus 5 at 74.0 percent and ahead of GPT-5.6 Sol at 72.7 percent. The previous 3.7 Flash scored 65.3 percent. Pricing stays at $0.75 per million input tokens and $3.75 per million output tokens until January 2027, when it doubles. Even then, it undercuts Opus 5 and GPT-5.6 Sol by a wide margin on per-token cost.

But per-token pricing is only half the story. Google says 3.8 Flash “works harder” on complex tasks, running extra reasoning steps and calling tools iteratively. That means higher token consumption. Artificial Analysis puts cost per task at $0.58, up about 40 percent from 3.7 Flash’s $0.40, even though the per-token price didn’t move. For efficiency-focused workloads, Google recommends sticking with 3.7 Flash or dialing down the reasoning level. The model also produces about 300 output tokens per second at high reasoning, with an average task time of 2.5 minutes—slower than Claude Fable 5.1 and 3.7 Flash.

There’s also a cybersecurity variant, 3.8 Flash Cyber, distributed only through the Fairwind Program to vetted defenders. Google claims 86.2 percent on CyberGym for vulnerability detection and 47.2 percent Pass@1 on CWE-Bench for automated patching. More notable is the prompt injection resistance: a 5.5 percent attack success rate on Gray Swan’s IPI benchmark, compared to 60.1 percent for DeepSeek V4 Pro and 51.8 percent for Grok 4.6. Only Claude Opus 5 does better at 4.8 percent. The security model is real progress, but it’s locked behind an access program most developers will never touch. Meanwhile, the gap between Flash releases keeps shrinking while the frontier models stay missing. That’s the part worth watching.