LLM API prices, tracked as they move.

How this log works

Every day, our pipeline pulls live pricing for every model we track and compares it to the day before. When a model's input, output, or cached-input rate changes, when a new model appears, or when a previously-tracked model disappears from the feed, it's recorded here with the date. Nothing is summarized or sampled — it's a complete log of every change we've observed since we started tracking.

The log covers Jul 13, 2026 onward — reconstructed from the real, dated snapshots our pipeline has committed since it began, then extended by a genuine daily diff from that point forward. We don't estimate or invent anything before that date; that's simply when this data format started existing for us to compare against. The log keeps growing every day; check back as it does.

Why LLM API prices tend to fall over time

Unlike most software pricing, LLM inference costs have a fairly direct line to a physical resource: GPU compute. Three forces push prices down over time, and this log is where you can watch them play out in real numbers rather than take it on faith:

• Model efficiency improves faster than raw scale grows. Newer model architectures increasingly deliver similar or better capability with fewer active parameters per generated token (mixture-of-experts designs being the clearest example), which directly lowers the compute cost per request even before any hardware improves.
• Competition is unusually intense. With a dozen-plus credible providers (see our provider hubs) shipping frontier-adjacent models within months of each other, no single lab can hold a price premium for long without a clear capability edge to justify it.
• Inference-specific hardware and serving software keep maturing. Batching, speculative decoding, quantization, and dedicated inference chips all reduce the real cost of serving a token, and providers pass at least some of that saving through to stay competitive.

None of this means prices only ever go down — new frontier-tier models are routinely launched at a premium before settling, and providers do raise prices when an introductory rate expires or when a model line is repositioned. That's exactly why a real log, not a one-time snapshot, is the useful artifact here. For a deeper dive into what this log actually shows — cuts vs. increases, the biggest single moves, and which providers reprice most — see How LLM API Prices Actually Changed, built entirely from these numbers. Want the whole market as one number instead of an event list? See the Token Price Index, reconstructed from this same log. This log covers models we've already priced — for models scheduled to retire soon, including ones we don't track pricing for, see upcoming deprecations.

How to use this

If you're mid-build and price-sensitive, filtering to price cuts above surfaces models that just got cheaper — worth re-checking against your current choice on the compare page. If you're maintaining a production integration, removed models is worth a periodic glance so a retirement doesn't surprise you. And new models is the fastest way to see who's shipping, before it shows up anywhere else on the site.

Frequently asked questions