The AI price crash has a catch.

The posted price of model input collapsed after GPT-4. But the frontier price has rebounded, reasoning can multiply token use, and AI cost management has become nearly universal. The sticker price is now only the start of the bill.

24× lower
GPT-5's launch input price versus GPT-4's, within OpenAI's flagship line
4× rebound
GPT-5.6 Sol's input price versus the GPT-5 low one year earlier
69% hidden
of GPT-5-nano output tokens were reasoning in one tool-use benchmark
31% → 98%
FinOps respondents managing AI spend, 2024 to 2026
01
The split

One price fell. The full cost became harder to read.

Headline API rates are dramatically lower than they were at GPT-4's launch. But “AI got cheaper” compresses several different things into one sentence: model tier, input and output rates, reasoning effort, prompt length, and how many calls an application makes.

The floor

OpenAI's flagship input price fell 24-fold — then quadrupled

Selected launch prices within one provider show the arc clearly: a steep fall from GPT-4 to GPT-5, followed by a rebound as the flagship tier became more capable and more expensive per token.

Standard API input price at launch · US dollars per million tokens · log scale
Selected OpenAI flagship model input prices at launch GPT-4 cost 30 dollars per million input tokens in 2023, GPT-4o cost 5 dollars in 2024, GPT-5 cost 1.25 dollars in 2025, and GPT-5.6 Sol cost 5 dollars in 2026.

Today's menu spans an eighteen-fold gap

Standard, uncached input prices from five current endpoints range from $0.14 to $2.50 per million tokens. They are not equivalent in capability, context, latency or output price — which is precisely why the cheapest list rate is not automatically the cheapest workload.

Frontier tier
Budget tier
02
Why the bill bends

The reasoning tax

Reasoning models can generate internal tokens before returning an answer. Those tokens are billed at the output rate but are not shown to the user. The amount depends on the model, task and effort setting, so a low input rate alone cannot predict the final cost.

Turning up "thinking" costs far more than it earns back

In one 260-question ophthalmology benchmark, the same GPT-5 model answered at three reasoning-effort settings. More effort raised both token use and accuracy — but on this already high-scoring task, the cost moved much more.

Cost per question (bars) vs. accuracy (line) · same model, same questions, three effort levels
GPT-5 cost and accuracy by reasoning effort Across low, medium and high reasoning effort, mean cost per ophthalmology question rose from 0.002 to 0.012 US dollars while accuracy rose from 95.0 to 96.5 percent.

Going from low to high effort multiplied mean reasoning tokens more than six-fold (172 → 1,087) and mean cost six-fold ($0.002 → $0.012) while accuracy rose 1.5 percentage points (95.0% → 96.5%). This is one closed, medical multiple-choice benchmark — not a universal return curve. Source: Antaki et al., arXiv:2508.09956, Supplemental Table 5.

Some models hide most of their bill in tokens you never read

Across eight models on the same 12,267-example validation set of tool-use tasks, the share of output tokens spent on internal reasoning ranged from zero to nearly 70%.

Share of output tokens that are invisible reasoning, by model

GPT-5-nano averaged 170 output tokens per query, including 118 reasoning tokens: 69% of its output tokens. Kimi-K2.5 and Qwen-3.5-9B reported none on this benchmark. Reasoning tokens are billed at the output rate; total cost also includes input tokens. Source: Agarwal et al., arXiv:2605.07112, Table 26.

03
What it did to operations

The management problem

Lower per-token rates did not make cost management disappear. As AI moved into more products and workflows, almost every FinOps respondent said managing AI spend had entered their remit.

AI cost management went from minority task to standard practice

Share of FinOps respondents who said they manage AI spend as part of their work.

% of FinOps respondents managing AI spend
FinOps respondents managing AI spend The share rose from 31 percent in 2024 to 63 percent in 2025 and 98 percent in 2026.

Source: FinOps Foundation, State of FinOps 2026. The overall survey had 1,192 respondents representing more than $83 billion in annual cloud spend; the 2026 AI-management item reports N=693.

The sticker price and the workload can move in opposite directions

The list rate can fall while a particular query becomes more expensive. Model selection, prompt length, output length and reasoning effort determine how many priced tokens a workload consumes.

OpenAI flagship input rate
2023 → 2026
$30.00
2023
$5.00
2026
83% lower per input token
vs.
GPT-5 cost per benchmark question
Low → high reasoning effort
$0.002
Low
$0.012
High
6× higher on this task
69%
of GPT-5-nano output tokens were internal reasoning in one tool-use benchmark
84%
lower inference cost for Switchcraft routing versus the best single model at matched accuracy
98%
of 2026 FinOps respondents said they manage AI spend
Reading the divergence

The unit-price crash is real. It simply does not settle the budget question. A workload's cost depends on which model handles each request and how many input, visible-output and reasoning tokens it uses. In the Switchcraft tool-use study, routing queries among models cut inference cost 84% while matching the best single model's accuracy — evidence that architecture can matter as much as the menu price.

The complete arc

Three chapters, one market.

Chapter 1 · The split

The launch-price crash was real.

Within OpenAI's flagship line, standard input pricing fell from $30 per million tokens for GPT-4 to $1.25 for GPT-5 — before rebounding to $5 for GPT-5.6 Sol.

Chapter 2 · The reasoning tax

Tokens are not all visible.

Reasoning can improve performance, but it also consumes billed output tokens. How much depends on the model, task and effort setting.

Chapter 3 · Managing the bill

The bill became a systems problem.

Routing, context length, output verbosity and usage determine whether lower rates become lower costs. FinOps teams increasingly have to manage the difference.