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

If you've been putting off trying AI tools because the API costs looked steep, the timing has quietly gotten a lot better. Model pricing has dropped sharply across the board in the past few weeks, and it's changing what's practical to build — not just what's cheap to experiment with.

OpenAI cut the price of its Luna model by roughly 80%, bringing input tokens down to around $0.20 per million — a level that makes high-volume workloads like customer support bots, document processing, or always-on automations dramatically cheaper to run than they were even a couple of months ago. DeepSeek pushed in the same direction with its V4 Flash release, pricing it low enough to undercut its own larger model while still beating it on agent benchmarks. Anthropic's Claude Sonnet 5, meanwhile, has an introductory pricing window that's about to close — the per-token cost steps up on September 1, alongside a tokenizer change that alters how many tokens the same piece of text now counts as. If you're mid-way through evaluating it for a project, that's worth factoring into your timeline.

It's not just about price

Three trends are moving together, and they compound. Alongside falling costs, context windows keep growing — providers are pushing toward million-token ranges, which changes what's realistic for whole-codebase reasoning or feeding in long documents without chunking them first. And "agents" — AI that can take multi-step actions rather than just answer questions — are moving out of demos and into mainstream products from the major labs, rather than staying a niche, DIY thing.

Put together, the practical effect is that things which felt like a stretch six months ago — automating a genuinely multi-step workflow, running AI over your entire codebase, processing large batches of documents — are now closer to routine, and cheaper to get wrong while you figure it out.

What to actually do with this

• If you shelved an AI-powered idea because the token costs didn't pencil out, it's worth re-running the math with current pricing.
• If you're on a model with a pricing deadline approaching, check whether locking in current rates (or migrating early) makes sense for your use case.
• 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. But if cost was the specific thing holding a project back, that specific objection has gotten a lot weaker in the last few weeks.