The discount is paid in retries
The price that matters is per solved task, not per million tokens.
On 21 September, Artificial Analysis published version 4.3.2 of its model index. Grok 4.7 costs 2 dollars per million input tokens and 6 per million output, roughly a third of what Claude Fable 5.1 or GPT-6 charge for the same volume. Its score on the index's own Intelligence benchmark is 46, against 53 for both of the pricier models. Read as a menu, that looks like a discount. Read as a budget for a team running agents against real business tasks, it is something else: fewer tasks finished per dollar, with the difference spent somewhere the invoice never shows.
Two dollars a million, forty-six on the index
The sticker price on a model is what you pay for a fixed amount of text going in and coming out. It says nothing about how much text a task actually needs, because a model that gets the answer wrong the first time does not stop, it produces a second answer, and often a third. Nobody bills that separately. It shows up as more tokens burned on the same task, or as a human opening the output and starting over.
That is the gap an index score is trying to point at without quite saying it. Forty-six against fifty-three is an eight-point difference on a scale most operators never look at closely, because the number that actually determines cost is downstream of it: how often does this model finish the task correctly on the first pass.
Twenty-six percent in the terminal, and what that tier is actually for
Terminal-Bench 4.0 measures exactly that, on agentic coding tasks that run for multiple steps. Grok 4.7 scores 26 percent. GPT-6 Astra scores 60, Fable 5.1 scores 55. DeepSeek V4.1 Flash, a model built to be cheap and fast rather than capable, scores 27.
That places the discount model next to a flash-tier model, not near the frontier pair it is priced to compete against on the marketing page. On multi-step agent runs the difference is not eight points, it is more than double. That is not an argument against using it. It is an argument for using it somewhere specific: routing, classification, bulk labeling, anything where a wrong output is cheap and visible on the spot. It is a bad target for a chain of tool calls where one wrong step three calls in produces a plausible-looking result that nobody re-checks before it feeds the fourth call.
Fifty-seven percent wrong, paid or not
A separate study, reported 21 September and covered by the Financial Times, ran 18 chatbots against 121 finance questions, asking each one five times. Averaged across everything, the answers were wrong 57 percent of the time. Free-tier models were wrong 63 percent of the time, paid models 49 percent. Paying moved the error rate by 14 points. On the complex questions in the set, both tiers were wrong 88 to 99 percent of the time. The single best-performing model, Claude Opus 5 with reasoning turned on, was still wrong 39 percent of the time.
Paying for a better tier moved the error rate fourteen points. A source to check the answer against would move it further, and it does not care which tier produced the answer.
What buys reliability, and what doesn't
Put next to each other, these two results say something an index score alone does not. The spread between a cheap tier and an expensive one is real, but it is smaller than the spread between an answer nobody checked and an answer checked against a source. Retrieval plus a verification step is what buys reliability. A tier upgrade mostly buys a smaller number of the same kind of mistake.
That is why an adversarial verification pass belongs before anything gets marked done, as a standing step, not as something reached for once a result looks suspicious. A result that looks fine is exactly the case the pass exists for, because the 39 percent wrong figure from the best model in that finance study includes plenty of confident, fluent, wrong answers. The routing decision follows from the same logic: a cheap model is a reasonable target for classification and bulk work, where the failure is cheap and somebody sees it immediately. It is a poor target for a multi-step run, where a wrong intermediate step compounds silently and the person downstream has no reason to suspect it.
The quota is the real budget
One more point, worth stating as what it is: a single user's own measurement, not a benchmark. That user tracked a large Claude subscription tier delivering roughly 2.2 times the weekly quota of the smaller tier a month ago, and about 1.5 times now, at an unchanged price. If that holds, quota is the actual constraint, more than the sticker price per token. Under a fixed quota, sending routine bulk work to a small, cheap model is not only a call about output quality. It is a decision about which tasks get to spend the scarce budget at all.
The per-token price is a menu. The per-task price is the budget, and it includes every retry, every re-read, and every wrong answer nobody caught before it shipped.

