
Breakdown / AI pricing · Working with AI
AI is getting cheaper. Why doesn’t it feel that way?
Model prices are falling, but subscriptions and the cost of finished work follow different rules. Here’s why cheaper AI can still feel expensive.
The cost of AI has been falling sharply. That doesn’t necessarily mean a smaller bill, a longer working day with your coding agent, or more useful work from your subscription.
Part of the confusion is that “cheaper AI” can mean several things: a lower price for processing text, less processing needed to solve a problem, or a more generous subscription. A saving in one does not guarantee a saving in the others.
What the falling-price research measures
In September, Epoch AI estimated that the cheapest cost of reaching a given level of performance across five primary benchmarks had fallen about 47% per quarter since 2023, roughly a thirteenfold reduction per year. Its analysis accounts for how much processing models use to answer the benchmark questions. It is more than a comparison of advertised token rates. Epoch’s study
That is a substantial improvement. But the study also explains why it may not describe your experience. Benchmarks imperfectly represent useful work, and customers don’t constantly switch between every available model to find the cheapest one for each task.
A model that reaches yesterday’s benchmark score cheaply may still struggle with the awkward spreadsheet, unfamiliar codebase or ambiguous customer message you need it to handle.
Lower rates are only one way to save
Developers usually pay to use a model through an API, which lets software send it requests. The bill often counts tokens, the pieces of information the model reads and writes.
Anthropic currently lists Haiku 5.5 at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens. Longer prompts have higher rates. That can make narrow jobs such as sorting messages inexpensive, provided the model handles them accurately. It does not establish that Haiku can replace a larger model on your work. Claude pricing
There are savings within a workflow too. Caching allows previously processed input to be reused at a lower rate. Sonnet 5.5 lists cache reads at $0.10 per million tokens, compared with $2 for ordinary input. The saving depends on what can actually be reused, alongside charges for writing the cache.
But a lower rate can be outweighed by more processing. In a simplified example, 100,000 output tokens at $10 per million cost $1. Halve the rate to $5 but generate 300,000 tokens, and the charge becomes $1.50. That example excludes input and tool charges to show the arithmetic.
Count the finished job
Ask an agent to fix a bug and it may read files, change code, run tests and try again. One request can produce many rounds of processing.
Cloudflare’s recent account of its AI code reviewer shows how much the surrounding workflow matters. It reports reducing average review cost from $2.43 to $0.77 through changes to caching, models, effort and the reviewer’s code. Its report describes unnecessary calls and canceled jobs that kept running. These are company-reported results, not an independently replicated saving. Cloudflare’s report
The useful comparison is the total cost of an acceptable result. Include failed attempts and human correction. A cheap answer that needs repeated repairs can cost more than an expensive answer that works first time.
Your subscription follows different rules
A monthly subscription buys access within a provider’s allowances. It does not promise a fixed quantity of work based on API prices.
OpenAI explicitly says API token prices should not be used to estimate included subscription tasks. Model choice, context, reasoning, tools and caching all affect consumption. Several sessions can draw from a shared allowance, so elapsed time alone is a poor measure of usage. OpenAI’s documentation
We felt that gap on October 20. Our Pro 100 account hit its weekly limit after roughly two hours, then again roughly two hours after a banked reset. The first stretch involved three sessions, two on Sol Medium and one on Astra Medium; after the reset, all used Sol. This is one account’s experience, not a general two-hour limit or proof that allowances were cut. We don’t have the consumption history needed to explain the depletion.
Falling model costs and frustrating subscription limits can both be real. The question for someone relying on AI is how much dependable work their money buys, including the time spent checking results and the disruption when access runs out.
A cheaper model is progress. A service you can reliably plan work around is the benefit customers still need to see.
Sources
- The plunging price of thoughtEpoch AI
- Claude API pricingAnthropic
- Cutting our inference bill by 68%Cloudflare
- Codex and Work pricing and usageOpenAI