Gemini 4 Hydrogen costs 76% less per task than 3.8 Flash

With a new Gemini model in the family, the first number being compared is not a benchmark score but a bill. An analysis by FourWeekMBA concludes that Gemini 4 Hydrogen costs 76% less than Gemini 3.8 Flash to complete the same task.

The metric deserves a definition. Cost per task bundles the price of a single call with the number of calls a task requires, plus the extra spend on retries and failures. For chat, a per-million-token price is a decent proxy; for agents it is not.

Seventy-six percent is a large gap, roughly one-fifth the cost. If it holds on real workloads, features previously confined to a small model, or forced to run fewer calls, could move to a stronger model within the same budget.

For agentic workloads, the number that sets a budget is not the sticker price per million tokens but the total cost of getting one thing done. A single task can involve a dozen tool calls, reading files, editing code and then verifying the result, so a small per-call difference compounds into an order-of-magnitude one.

The caveats are equally clear. This is a single third-party calculation, and the result moves a lot depending on which workloads are sampled and whether cache hits and batch discounts are counted. Published price sheets and independent measurements frequently disagree.

The more interesting signal is where price competition has moved. For the past year the comparison point was the per-token number; now it is the cost of finishing a job, which is closer to how enterprises actually buy.

For developers that means redoing the tiering math. Cascade designs that use a cheap model for triage and an expensive one to finish may invert once the cost gap reaches 76%.

What to watch: whether fuller official pricing and rate limits are published, whether independent benchmarks reproduce the ratio, and whether rivals start competing on cost per task rather than token price.