Tokens Got 214x Cheaper. AI Bills Went Up Anyway

On 30 July 2026 OpenAI cut the price of GPT-5.6 Luna by 80 percent. Input dropped to 20 cents per million tokens, output to $1.20. Terra came down by a fifth. Six days earlier Anthropic had shipped Opus 5 at the same price as the previous generation while claiming more than double the task performance. Google's Gemini 3.6 Flash, released 21 July, spends 17 percent fewer output tokens on comparable work.

Read only the price lists and July looks like the best month ever to be a company paying for AI.

The accounting says otherwise. The State of FinOps 2026 survey, covering 1,192 practitioners who manage more than $83 billion in annual cloud spend, found that 73 percent of organisations blew past their AI budget. Over the same stretch the average enterprise AI budget went from $1.2 million a year in 2024 to $7 million in 2026.

Prices are falling. Bills are rising. Both are true, and the reason is worth understanding before you sign off on your next AI project.

July 2026 was the densest month in the history of language models

Four major labs shipped flagship or near-flagship models inside thirty days, and the largest open weights ever published landed alongside them.

Moonshot's Kimi K3 is the largest openly available model at 2.8 trillion parameters. The full weight download runs to roughly 1.56 TB, so almost nobody is hosting it in their own rack. The number that matters is a different one. Models that sat at the top of the price ladder a year ago now sit in the middle of the table.

The price of a token stopped being the interesting price

When GPT-4 launched in March 2023, a million input tokens cost $30. By late July 2026 comparable quality was available at 14 cents. That is a 214-fold compression in forty months, a deflation curve with no precedent in enterprise software.

Budgets still burst. The explanation is not in the rate card, it is in how models get used now.

A chatbot takes a question and returns an answer. One call, one billed batch of tokens. An agentic workflow that handles an incoming invoice will fetch the document, classify it, extract line items, match them against a purchase order, flag the mismatch and draft a reply. That is easily twenty calls for one document. Same token price, twenty times the consumption.

Reasoning models add a second layer. They generate internal thinking before answering, which you never see and always pay for at output rates. Five hundred visible words can sit on top of tens of thousands of billed tokens.

Analyst Jacob Bourne put it plainly to VentureBeat. “The era of tokenmaxxing is over. Enterprises have figured out how easy it is to burn tokens without getting value back.”

The metric worth adopting

Stop comparing price per million tokens. Compare price per finished task.

The gap between those two numbers is wider than it looks, and published SWE-Bench Pro measurements show it cleanly. Grok 4.5 resolves an average task in 15,954 output tokens. Opus 4.8 in maximum effort mode needs 67,020.

The rate card differs by 4.2x. Consumption also differs by roughly 4.2x. Those two multipliers do not add up, they multiply, which puts the real per-task gap near seventeen times.

Treat that as an illustration of the mechanism rather than a verdict on either model. These are published benchmark figures, not our own measurements, and a different workload reorders them. The principle is what carries over. Token consumption is the other half of the price and most companies never track it.

The labs already know. Google led its Gemini 3.6 Flash announcement with 17 percent lower output token spend rather than higher intelligence. Less talking, not more thinking.

Effort became a line item

The second July shift is quieter and matters more in practice. A model is no longer one thing at one price.

Opus 5 ships with a selectable effort level of low, medium or high. Same question, same model, different consumption and a different answer. GPT-5.6 Sol gained a fast mode running about 2.5 times quicker at double the price. Grok 4.5 bills one way below 200K tokens of context and another way above it, and crossing the line reprices the whole request.

So "we use this model" is no longer a single decision. It is a model, an effort level, a context size and a caching strategy. A company that configures this once and forgets it is paying contract-analysis rates to sort email.

What to actually do about it

Measure cost per process, not per API

Your provider invoice tells you what you spent. It does not tell you which workflow spent it. Tag calls by the process that triggered them and track cost per processed document, per resolved ticket, per prepared quote. That is the only number you can put next to a person's hourly rate.

Do not hardcode model names

Rankings changed three times in July alone. If the model name is a constant scattered across twenty files, every switch is a project. If it sits behind one interface as a single config value, it is a ten-minute change. This is the cheapest insurance available and it costs you one afternoon.

Build the eval set before you start switching

Without thirty to fifty real cases with expected outcomes, moving to a cheaper model is a gamble. With them it is a single measurement run. That same set is your only defence against silent quality drift when a provider updates a model underneath you.

Cap spend at the workflow level

An alert on total monthly spend tells you once it is too late. A ceiling on calls and tokens within a single agent run stops a runaway loop the same day. Most expensive surprises are not slow growth, they are one bug that ran over a weekend.

Use caching and cheap models for triage

Cached input costs a fraction of the standard rate at most providers. Grok 4.5 bills cached input at $0.50 instead of $2. For an agent that resends the same system context at every step, that is not a rounding error. Classification, routing and field extraction run fine on the cheapest tier, which is now close to free.

What this means for a smaller company

For a twenty-person firm, the difference between $5 and $0.20 per million tokens is not what decides whether a project pays off. At ordinary volumes that is tens of euros a month.

The weight sits elsewhere. If swapping a model means rewriting twenty files, every price cut on the market passes you by, because you cannot reach it. Then there is visibility. A company that can say on request what it costs to process one invoice can also push that number down. A company that cannot will hear about rising consumption from the invoice.

The market will keep moving. In one month the middle tier fell by a fifth, the bottom tier by four fifths, and flagship pricing went up. Companies that treated the model as a replaceable part will collect that. Companies that picked one vendor and poured it into the code will keep paying for last season's decision.

If you are working out which model fits which process, we covered that separately in our piece on multi-model strategy and in the GPT-5.6 versus Claude Fable 5 comparison. For how token cost stacks up against a developer hour, start here.

At Rise.sk we build AI automation and workflows so the model can be swapped without touching the architecture. We measure cost per processed unit from day one. If you want to know what one process costs you today and whether it can be made cheaper, get in touch.

Sources

• Anthropic, Introducing Claude Opus 5
• OpenAI, Advancing the price-performance frontier with GPT-5.6
• VentureBeat, AI price wars, OpenAI cuts GPT-5.6 Luna prices by 80 percent
• CNBC, OpenAI cuts prices for two of its GPT-5.6 AI models
• FinOps Foundation, State of FinOps 2026
• Analytics Vidhya, July 2026 AI releases timeline
• DataCamp, Grok 4.5 features, benchmarks and pricing
• Axis Intelligence, AI inference cost statistics 2026
• OpenRouter, Kimi K3 model card