What happened
On 22 September, roughly ninety minutes apart, the two largest model providers shipped new versions and cut prices. Anthropic released Claude Opus 5.5: four dollars per million input tokens and twenty for output, against five and twenty-five for Opus 5. OpenAI answered with the GPT-6 series — Sol at two dollars for input and ten for output, Luna at ten cents and fifty cents. Sol's rates are half those of the previous 5.6 series, and Luna got cheaper still on the output side.
The more important change sits in a different column of the price list. A cache read — reloading context the model has already processed — costs twenty cents per million tokens on Opus 5.5 instead of fifty. Anthropic's documentation says it plainly: until now every model paid one tenth of the input price for a cache read, while Opus 5.5 pays one twentieth. Fable 5.1 went lower still, to one fortieth. The fixed multiplier every calculation rested on is no longer fixed.
On OpenAI's side, a cache read on GPT-6 Sol also costs twenty cents, and one cent on Luna. The company stresses that it improved the default cache hit rate, not just its price. In a separate launch-day post, Anthropic explains why comparing per-token rates is misleading and estimates that a typical workload on Opus 5.5 costs roughly two fifths less than on Opus 5. That is the vendor's claim, not an independent measurement — the post itself ends by advising you to run the numbers on your own tasks.
Our take
Flagged as our interpretation: neither provider calls this a change of strategy. Our read is that the price war has moved from the headline per-token rate to a column most teams never look at.
The reason is simple. An agent working in many steps resends the whole conversation so far to the model at every step. In a long session, the vast majority of input tokens are the same context being loaded yet again. If you pay the full rate for it because your cache doesn't hit, you pay a multiple of what the price list suggests. If your cache is set up well, you pay mostly for output.
A simplified calculation to make it visible. Take an agent session that uses ten million input tokens and three hundred thousand output tokens, and ignore the cost of the first cache write. This is an illustration, not a measurement of any deployment.
- Opus 5.5, nine tenths of input from cache: USD 4 for uncached input, USD 1.80 for cache reads, USD 6 for output — USD 11.80 in total.
- The same session on Opus 5.5 with no cache hits: USD 40 for input and USD 6 for output — USD 46 in total.
- The same session on Opus 5, with cache: USD 17.
- The same session on GPT-6 Sol, with cache: USD 6.80.
The gap between the two providers with a well-configured cache is a few dollars. The gap between the same model with and without cache is almost fourfold. In this example, how you built the agent weighs more than who you buy the model from.
Hence the takeaway for your next cost review. A table comparing providers by the rate per million input and output tokens was already an approximation, and after 22 September it is simply the wrong tool. In August we wrote that the premium for the most expensive model has a ceiling set by buyers. That shift was about choosing a model tier. This one is about architecture: the bill stops depending mainly on which model you pick and starts depending on how often your application loads the same context from scratch.
Why it matters
Private Equity
If a portfolio company's financial model counts AI cost as token volume times the list rate, that line just lost its meaning in both directions. It may be overstated because it ignores caching, or understated because it assumes hits a production agent doesn't achieve. One number is worth asking for at the review: what share of last month's input tokens was billed as cache reads. A company that doesn't know it doesn't know what its AI-based product actually costs either.
Enterprise
The price cuts won't flow through to the invoice automatically. A cache read is cheap only when the start of the request is identical between calls. Any timestamp in the system prompt, any reordering of tools and any dynamic header can invalidate it. Those are decisions made by application teams, not procurement, and negotiating a discount with the provider won't replace order in the request structure. Then there's the model change itself: migrating from Opus 5 to Opus 5.5 is not just a new identifier but a different cache multiplier, so the post-migration cost forecast has to be recalculated on the new rules.
SMB / mid-market
For a smaller company this is simply good news: the same agent costs less from this week, and noticeably less with a well-configured cache. You don't need to switch providers to benefit. If you use an off-the-shelf agent, ask whoever built it whether they've moved it to the new model and whether caching is switched on at all. For the patterns in our product library, this is a line in the specification, checked at every model change, not a default someone once left in place.
One move for this week
Pull the last thirty days of usage from your provider's console and split input tokens into two columns: billed as cache reads and billed at the full rate. If the second column is bigger, you have a better savings opportunity than any change of provider — and you'll find it in your own application's code, not in the price list. Then recalculate one typical task at the new rates instead of taking the announced cut on faith. Describe your case: mailto:[email protected]?subject=Rozmowa%20z%20Aurora%20AI