The Real Cost of AI APIs: A Token-by-Token Breakdown of Claude, GPT-4, and Gemini

Ask a finance team what changed about their AI spend this year and you get a strange answer. Nothing on the pricing page went up. Several things went down, sharply. Opus-class intelligence that cost $15 per million input tokens in early 2025 costs $5 today. OpenAI cut the price of its cheapest current-generation model by 80% in a single announcement at the end of July. Google's Flash tier is running at an introductory rate that undercuts what mid-tier models cost a year ago.And the invoice still went up.

That gap is the most interesting thing happening in AI infrastructure right now, and it is not really a pricing story. It is a story about what a token has come to represent. In 2024, one user question meant one API call. In 2026, one user question routed through an agent means a planning pass, four tool calls, a failed test, a retry with the whole context re-sent, and a synthesis step — and somewhere in the middle of that, several thousand tokens of reasoning the model did silently and charged you for at the output rate.

The rate fell. The number of tokens per unit of useful work rose faster. Analysis across billions of enterprise API calls this year put the year-on-year token price drop at roughly 67%, with the majority of enterprises still overshooting their AI budgets over the same period.

Here is what the three major providers actually charge, and where the money goes.

The rate cards, as of this week

All figures are US dollars per million tokens, input then output, standard tier, short context. These move constantly, and three of the numbers below have expiry dates attached.

Anthropic

Sonnet 5 launched in June at $2/$10 as an introductory rate with a standard rate of $3/$15 scheduled behind it. The introductory rate held. That makes it the most aggressively priced model in its capability class, and it is the reason a lot of teams that were running Opus a year ago are no longer running Opus.

Opus 5 arrived on 24 July at the same $5/$25 its predecessor charged, with a 1M-token context window as both default and maximum and 128K output. Fable 5.1 sits above it at double the price, and its cache reads were cut to $0.25 per million — a meaningful change for long-running agent work, where the same context gets re-read dozens of times. There is also a Mythos tier sharing Fable's model and pricing, but it is not self-serve; access runs through approved programmes rather than a credit card.

The detail worth knowing: Opus 5 ships with five reasoning effort levels, and the default is high, not medium. Two teams paying identical rates can produce very different bills without ever touching a pricing page.

OpenAI

The GPT-5.6 family went generally available on 9 July with a 1.05M-token context window across all three tiers. Three weeks later OpenAI cut Terra by 20% and Luna by 80%, which rewrote the bottom of the market — Luna now costs roughly four times less than the previous generation's mini tier while sitting inside the flagship family. Astra landed on 3 September at the top.

Sol's $4/$20 is explicitly promotional and guaranteed only through 21 November 2026. Anyone building a budget on it should schedule a review for that week rather than treating it as a fixed cost.

And there is a harder deadline. On 23 October 2026, gpt-4-turbo, gpt-4, gpt-4o-2024-05-13 and o1 all shut down, with GPT-5.6 Sol listed as the replacement for each. A service still pointing at any of those model IDs does not degrade that day. It stops.

Google

Google runs the most complicated pricing surface of the three, and the most scheduled change. The $0.10 Flash-Lite rate is the cheapest number in this entire article, and it disappears on 16 October when the 2.5 family retires; the successor costs two and a half times as much. The Flash rate is introductory and doubles to $1.50/$7.50 on 1 January 2027. And Gemini 3.1 Pro carries a context threshold — prompts above 200K tokens bill at $4/$18 instead of $2/$12, which means a workload that grows gradually can double its own input cost without anyone changing a line of code.

Google also runs four service tiers against the same model, with Batch and Flex at half rate and Priority at 1.8x. Which tier a request lands on can move the bill more than which model handles it. Search grounding is metered separately again: the Gemini 3 family includes 5,000 free grounded prompts a month, then $14 per thousand search queries, and one request can fire several queries.

Where the money actually goes

A request does not go to a model. It goes through a pipeline, and each stage has a cost consequence that the per-token rate does not express.

It starts with tokenization. Your text is broken into the numeric IDs the network consumes, and every provider uses a different tokeniser — which is why comparing headline rates across providers is less informative than it looks. The same 10,000-word contract is a different number of billable tokens depending on who reads it. Providers also change tokenizers between generations. Anthropic's newer one expanded token counts by up to 1.35x relative to the previous Sonnet for identical text. That is a silent price increase of up to 35% that never appeared on any rate card, and teams migrating between model versions have to account for it.

Then comes inference. The model runs a forward pass for every token it generates, each one conditioned on everything before it. Reading your input can be parallelized; writing the answer cannot. That asymmetry is the entire reason output costs three to five times input, and it holds across all three providers.

The part that catches people is what happens between those two steps. On current-generation models, extended thinking is on by default, and those internal reasoning tokens bill at the output rate whether or not you ever see them. A two-sentence answer to a hard question can sit on top of several thousand billed tokens of deliberation. In a coding agent session, the actual generated code is often only 5% to 15% of total tokens consumed — the rest is context overhead and reasoning. Effort parameters are the only real lever, and, again, the defaults are not the cheap setting.

Caching is the one genuinely large discount available, and it is underused. Cache reads bill at roughly 10% of fresh input across all three providers, and batching halves everything on top. Stack both and a cache-heavy asynchronous workload can run at a fraction of list price. The catch is storage: explicit caching charges per token per hour, and on Gemini that runs between $1 and $4.50 per million tokens per hour depending on the model. A cache left sitting overnight can cost more than the inference it was meant to save.

The decision that decides the bill

Everything above matters less than one architectural choice, usually made in the first week of a project and then never looked at again.

That Q1 analysis of enterprise API traffic found that organizations running tiered model architectures — cheap models for routine work, expensive models only on escalation — landed at a median blended cost of about $2.31 per million tokens. Organizations that routed everything to frontier models paid around $18.40. Same category of work. An 87% difference.

It is easy to see why it happens. Picking the best model is the safe default, it is what you did during the prototype, and nobody gets fired for it in month one. But classification, intent detection, extraction, routing and short summarization make up the bulk of most production workloads, and none of them need frontier capability.

Consider a support system handling 10,000 tickets a day, averaging 2,000 input and 500 output tokens per conversation. That is 600 million input and 150 million output tokens a month. Before any caching or batching, the monthly cost looks like this: