Google's Gemini 3.6 Flash, announced July 21, 2026, dropped the floor on the workhorse tier again — and it did it in a way that breaks the pricing table you'd use to compare it. So the real shortlist for a cost-sensitive agent is three models, and they do not rank the way their sticker prices say they do.

If you read one line: the cheapest per token is GPT-5 mini ($0.25/$2.00 per 1M), but the cheapest per completed task depends on your workload — Gemini 3.6 Flash wins on long-horizon coding where it emits fewer tokens, Claude Haiku 4.5 wins on cacheable multi-agent fan-out, and GPT-5 mini wins on short, high-volume tool-calling. Measure cost per finished task before you switch a default.

The sticker prices (and why they lie)#

Per 1M tokens, input / output:

On this table alone you'd never pick Gemini: it costs about 3x GPT-5 mini per token. But price per token is the wrong unit. The bill on a real agent run is set by two things the table doesn't show — how many output tokens the model emits to finish, and how many times it has to retry. That's where Gemini's pitch lives.

Gemini 3.6 Flash's real product is fewer tokens#

Google's headline isn't the price cut (though output did drop from 3.5 Flash's $9.00/1M to $7.50/1M). It's token efficiency: Gemini 3.6 Flash uses roughly 17% fewer output tokens than 3.5 Flash on average (per the Artificial Analysis Index), and up to 65% fewer on long-horizon engineering tasks, measured on the DeepSWE agentic benchmark (VentureBeat).

On a long agentic run — where a single coding task can burn 1–3.5M tokens including retries — output volume dominates the invoice. A model that finishes the same task with 40% fewer tokens can be cheaper overall even at a higher per-token rate. That's the whole trick, and it only shows up if you measure the finished task, not the price list.

Claude Haiku 4.5's real product is the cache#

Haiku 4.5's sticker ($1.00/$5.00) sits in the middle, but its cost lever is prompt caching: cache reads at $0.10 per 1M — about 90% off input — plus a 50% batch discount. If you run many parallel sub-agents that share one large system prompt or tool schema, the second call onward reads that context from cache instead of paying full input price. For multi-agent fan-out over many small tasks, that structure, not the sticker, is what you're buying.

The tradeoff is the smallest context window of the three (200K) and a 64K output ceiling — fine for sub-tasks, tight for a single long document.

GPT-5 mini's real product is the input price#

At $0.25 per 1M input, GPT-5 mini is genuinely the cheapest to feed, with a roomy 400K context. For high-volume, latency-sensitive tool-calling where each call is short and you control verbosity, it's hard to beat. The risk is the flip side of "mini": on harder tasks it can get chatty or need an extra retry, and either erases the input-price lead. Cap max_output_tokens, force structured output, and watch the retry rate.

The cheapest model on the pricing page and the cheapest model on your invoice are frequently not the same model. The gap between them is retries and output verbosity — and both are invisible until you measure the completed task.

How to actually decide#

Not from the table above. Pick the two or three tasks your agent runs most, fix a representative input set, and run each model through them recording **dollars per successful task** — including the tokens burned on retries and failed attempts left in context. This is the same discipline behind why one tokens-per-second number is lying to you: the vanity number and the number you pay diverge exactly where it matters.

We wrote the step-by-step for it: how to measure cost per completed task for your agent. Run it once and the choice between these three usually makes itself — and it won't be the one the sticker prices predicted.

Related on the desk: Gemini 3.6 Flash: the cheaper workhorse, explained · The frontier price war: picking an agent model on cost per run · This week's Wire on the Flash price cut.