If you build agents, the last week reset your cost base twice. On July 24, Anthropic shipped Claude Opus 5 at Opus 4.8's price — $5 per million input tokens, $25 per million output — putting near-frontier reasoning at what used to be a mid-tier rate (Axios). Three days earlier, on July 21, Google shipped Gemini 3.6 Flash at roughly $1.50 / $7.50 per million, a workhorse tier that undercuts almost everything (TechCrunch). So the obvious question — which one do I make my agent's default? — is now worth a real answer.

Here it is up front: use Gemini 3.6 Flash for the high-volume, low-stakes steps, and Opus 5 for the small number of steps where being wrong is expensive. The price gap is real, but it is not the decision. The decision is which task tier each model wins.

The per-token gap is the wrong number#

The spread is about 3x: Flash is roughly a third of Opus 5's price on both input and output. If your agent's cost were a straight function of tokens, this would be over — run Flash, pocket the difference.

But an agent's bill is not cost-per-token, it's cost-per-completed-task. A cheaper model that needs a second pass, or produces code that a human has to fix, can cost more per finished job than the "expensive" model that closes the task on the first try. This is the same trap we flagged when one tokens-per-second number lies to you: the headline metric measures the input, and you get billed for the outcome.

So the useful move isn't to compare prices. It's to sort your agent's steps into two tiers and price each tier on both models.

Tier one: the forgiving steps → Gemini 3.6 Flash#

Most of what an agent does is not hard. Classifying an intent, extracting fields from a document, deciding which tool to call, formatting a retrieval result, summarizing a page — these are high-volume and forgiving. A wrong answer is cheap to catch and cheap to retry, and the sheer call count means the per-token rate dominates.

This is Flash's tier. Google's own pitch is that 3.6 Flash uses about 17% fewer output tokens than 3.5 Flash (Google; VentureBeat) — and on long agent runs, output volume, not the per-token rate, is where the money goes. Cheaper tokens and fewer of them compounds fast.

Tier two: the expensive-to-be-wrong steps → Opus 5#

Then there's the handful of steps where a mistake is costly: the actual code change, the multi-step plan the rest of the run depends on, the irreversible action, the reasoning-heavy analysis a customer sees. Here a retry isn't cheap — it burns a whole run, or ships a bug, or loses trust.

This is Opus 5's tier. Frontier reasoning that closes the task in one pass is worth a 3x token premium when the alternative is two Flash attempts and a human correction. Anthropic positions Opus 5 as near-top-tier intelligence at an everyday price — and "everyday price" is exactly what makes it viable as the escalation target inside an otherwise cheap agent, rather than a model you can only afford to touch occasionally.

The tie-breaker: Opus 5's effort dial#

There's a wrinkle that complicates the clean two-provider split. Opus 5 exposes a five-level effort setting plus adaptive thinking, so you can run the same model cheap-and-fast on easy calls and crank it to maximum reasoning on hard ones (Axios). That means "route by task tier" can partly happen inside one model instead of across two backends — and it's the knob we walk through in cutting your Opus 5 bill with the effort parameter.

So the real fork is operational:

If your agent is open-weight-curious, the third pan on the balance is Kimi K3 versus Opus 5, whose weights land this week — but for a hosted-API default, the Opus 5 / Gemini 3.6 Flash split above is the one most teams will actually run.

How to decide in an afternoon#

Don't argue it in the abstract. Take ten representative jobs from your agent, run each on both models, and record: first-pass success, retries, human fixes, and real token counts. Multiply out to a total cost per finished job. You will almost always land on the same shape — Flash for the easy tier, Opus 5 for the hard tier — but now you'll have the number that proves it for your workload, which is the only version of this comparison that pays your bill.