Share with your CTO
Google is betting that “effort-configurable” inference, where a model performs more reasoning steps and tool calls on harder tasks, is the right architecture for agentic workloads, and Gemini 3.8 Flash is the first public test of that bet. The new model keeps 3.7 Flash’s per-token pricing at $0.75 input and $3.75 output, but Artificial Analysis clocked a 40% effective cost increase in practice, driven by 30% more output tokens per task and longer agentic runs. It tops the DeepSWE v1.1 software engineering benchmark and leads on finance and legal agent evals.
What this means for your business
The teams most exposed here aren’t the ones building chat interfaces. They’re the ones already running agentic pipelines, where Gemini 3.7 Flash is handling multi-step coding reviews, document processing, or automated QA. Those teams likely scoped their infrastructure budgets around 3.7 Flash’s behavior, and 3.8 Flash’s token expansion isn’t a config flag they can ignore. A 40% cost increase on a high-volume agent loop isn’t rounding error; it’s a line item that resurfaces in a quarterly cloud bill review.
The benchmark results deserve scrutiny, but they’re pointing at something real. Topping DeepSWE v1.1 ahead of Anthropic’s Fable 5 matters specifically because software engineering agents, the kind autonomously writing, testing, and debugging code, are where token count inflates fastest. More reasoning steps mean more tokens mean more cost, but also, potentially, fewer failed runs and less human intervention downstream. The engineering question is whether the reduced failure rate offsets the higher per-run cost, and that ratio will vary sharply by task complexity. Google is essentially asking teams to rerun their own benchmarks before they assume the answer.
The pricing structure Google chose is a slow-burn pressure tactic. Holding the per-token rate flat while the model consumes more tokens per task keeps the headline number clean while shifting real cost to usage behavior, a pattern that rewards teams with tight observability over their inference spend and punishes those running on autopilot. The falsification condition for Google’s bet is straightforward: if autonomous agents actually reduce total cost-per-outcome despite higher token counts, 3.8 Flash wins; if they just burn more tokens without proportional quality gains on real production workloads, developers route back to 3.7 Flash exactly as Google suggested they could.
Concept deep-dive: Effort-configurable inference
Effort-configurable inference lets a model dynamically adjust how many reasoning steps and tool calls it makes before returning an answer, similar to asking a contractor to do a quick pass versus a full audit on the same job. The model “works harder” on complex requests and shorter on simple ones. For agentic systems handling varied task difficulty, this sounds efficient, but because token consumption scales with effort level, cost becomes a function of task complexity rather than a predictable flat rate.
Based on reporting from Google says its new Gemini 3.8 Flash model ‘works harder’ but might cost more, originally published 2026-09-02 16:11:00.
