AI costs are falling fast — here's what that means for you

In late July 2026, lower prices for specific models made it worth revisiting AI projects whose API costs had looked too high. That was a reason to recalculate a budget, not evidence that prices were falling across every provider or that every workflow had become economical.

On 30 July, OpenAI announced an 80% API price reduction for GPT‑5.6 Luna, with new rates of US$0.20 per million input tokens and US$1.20 per million output tokens. The comparison is the same model’s pricing before and after that announcement, not GPT‑6 Luna or an 80% reduction in a customer’s total AI bill. OpenAI said subscription prices and quota budgets were unchanged.

DeepSeek’s 31 July changelog reported stronger results for V4-Flash on specified agent benchmarks than V4-Pro-Preview. Those are vendor-reported results for particular model versions and test conditions, not proof that a smaller model will outperform a larger one on your work.

At the time of this article’s original 5 August publication, Anthropic had scheduled Sonnet 5’s introductory rates to rise on 1 September. That plan was subsequently cancelled, as explained above. The separate tokenizer change still matters: Anthropic says the same input can count as more tokens than with Sonnet 4.6, depending on the content. A token price alone does not tell you the cost of processing a document.

It's not just about price

Context limits and the ability to use tools also affect what you can build. A model that accepts a long document may still need its answer checked; an agent that takes several steps may make several billable calls. Lower unit prices do not remove the need to measure whether the workflow does the job.

What to actually do with this

• If you shelved an AI-powered idea because the token costs didn't pencil out, re-run the maths with the exact model’s current pricing.
• Check input and output rates, caching, processing tier and tool charges. Compare the cost of a completed task, including retries and human checking.
• Falling cost per token doesn't mean falling total spend if usage scales up to match — keep an eye on volume, not just the unit price.

None of this is a reason to rush into a tool you don't need. If cost was the specific barrier, use a small representative workload to find out whether the numbers now work for you.