Cached Input Token Pricing: Anthropic Cut the Multiplier, OpenAI Held It

September 23, 2026. On Tuesday the two frontier labs cut prices within minutes of each other, and the coverage fixed on the sticker numbers. Anthropic took Claude Opus 5.5 to 4 dollars per million input tokens and 20 dollars per million output. OpenAI halved GPT-6 Sol and Luna against the GPT-5.6 models of the same names. For an agency or a small business running an AI agent in production, neither figure is the one that decides the invoice. The number that decides the invoice is what a cached input token costs as a fraction of a fresh one, and on September 22 that fraction moved on one side of the market and stayed put on the other.

What both labs announced on September 22

1. Anthropic released Claude Opus 5.5 at 4 dollars per million input tokens and 20 dollars per million output, which the company describes as 20 per cent less than Opus 5, and says works out to 40 per cent less on a typical workload at default settings.
2. Anthropic cut cache reads on Opus 5.5 to 20 cents per million tokens, 60 per cent below the 50 cents Opus 5 charged, and stated in the launch post that cache reads make up the majority of agentic and coding work costs.
3. The Claude Platform pricing page now carries a footnote that Opus 5.5 cache hits and refreshes are priced at 0.05 times the base input price, while all other models use the standard 0.1 times multiplier. Claude Fable 5.1 sits lower again at 0.025 times.
4. OpenAI published GPT-6 Sol at 2 dollars per million input and 10 dollars per million output, and GPT-6 Luna at 10 cents and 50 cents, both half of their GPT-5.6 predecessors.
5. OpenAI separately announced improved prompt caching for the GPT-6 family, with higher default cache hit rates and discounts of up to 90 per cent on cached input tokens, and said cache discounts apply to eligible shared prefixes reused within a 30 minute window.
6. OpenAI shipped a Prompt Caching Dashboard and a diagnostics tool, and made it possible to change reasoning effort or toggle tool availability between responses without breaking cache.

The multiplier is the number that moved

Every provider in this market prices a cached input token as a fixed fraction of the fresh input price. That fraction has been 0.1 for years, across most of the industry. It is the reason a long standing system prompt, a tool schema and a knowledge base can be reused cheaply at all.

Anthropic has now broken that convention twice in three weeks. Claude Fable 5.1, released on September 1, prices cache hits at 0.025 times base input, and Opus 5.5 prices them at 0.05 times. Anthropic states the multiplier explicitly in its own pricing documentation rather than leaving it to be inferred. OpenAI, by contrast, holds the 0.1 fraction across the entire GPT-6 family, and the published table proves it: Astra at 10 dollars input and 1 dollar cached, Sol at 2 dollars and 20 cents, Luna at 10 cents and 1 cent. Each pair is exactly ten to one.

So the two cuts are not the same kind of cut. OpenAI moved the whole price line down by half and left the shape of it alone. Anthropic moved the line down 20 per cent and changed the shape, which is why its own headline figure for a typical workload is 40 per cent rather than 20.

What a cache write costs, and when it pays

A cache is not free to create. On Opus 5.5 a five minute cache write costs 5 dollars per million tokens and a one hour cache write costs 8 dollars per million, against 4 dollars for a plain fresh read. On GPT-6 Sol a cache write costs 2 dollars fifty per million against 2 dollars for a fresh read.

That turns caching into a break even question, and the arithmetic is short enough to do in your head. On Opus 5.5 with a one hour write, the cached path costs 8 plus 0.20 for every million tokens read back, and the uncached path costs 4 for every million. Those cross at a little over two reads, so the one hour cache starts paying on the third read inside the hour. The five minute write crosses at about 1.3 reads, so it pays on the second. GPT-6 Sol crosses at about 1.4, so it also pays on the second read.

The practical consequence is blunt. If your agent reuses its standing context at least twice before the cache expires, caching is close to free money. If it does not, you are paying a premium for a cache you never read.

The 30 minute window decides who gets the discount

OpenAI's caching post puts a number on the reuse window that most cost models ignore: cache discounts apply to eligible shared prefixes reused within 30 minutes. Anthropic sells the window as a product choice instead, with separate prices for a five minute cache and a one hour cache.

Either way, the constraint is the same and it is about frequency, not volume. A support agent handling a message every few seconds hits cache constantly. A nightly enrichment job that fires once and stops hits cache once. An after hours intake agent for a professional services firm, the kind we build on AI intake for law firms, may take one call at nine in the evening and the next at midnight, and every one of those calls pays full price on the entire system prompt, the tool schema and the firm's matter intake rules, because the cache expired hours earlier.

That is the part the price cut does not fix. A 50 per cent cut on fresh input tokens helps a bursty agent. A 60 per cent cut on cache reads does almost nothing for it.

A worked example on a 30,000 token standing context

Take a realistic agent: 30,000 tokens of standing context, made up of a system prompt, tool definitions and a small retrieved knowledge base, sent on every call. Price it on Opus 5.5.

1. Fresh, uncached, at 4 dollars per million: 12 cents per call.
2. Served from cache, at 20 cents per million: 0.6 of a cent per call.
3. Writing that context into a one hour cache, at 8 dollars per million: 24 cents, once.

Now run it two ways. A busy agent making 200 calls in an hour pays 24 cents for the write plus 1 dollar 20 for the reads, so 1 dollar 44 against 24 dollars uncached. That is a 94 per cent saving, and it is roughly what the marketing implies. A quiet agent making one call an hour pays 24 cents for the write plus 0.6 of a cent for the read, so about 24.6 cents against 12 cents uncached. Caching doubled the bill.

Same model, same price list, same week. The difference is entirely how often the thing runs.

What it means for operators

Three things follow for anyone buying or running agents for a small business.

First, the quote you were given before Tuesday is probably still roughly right, because most agency builds are not cache heavy enough to collect the headline discount. If a vendor tells you your bill just fell 40 per cent, ask what their cache hit rate is. Anthropic's 40 per cent figure is explicitly for a typical workload at default settings, and an agentic coding workload is not a typical small business workload.

Second, model choice now matters less than call pattern. Batching work so that an agent fires repeatedly inside a short window, rather than being woken at intervals, is worth more than switching providers. We cover the wider arithmetic in what an AI agent costs to run and the per call maths in the Agents API cost breakdown.

Third, watch the long context tier. OpenAI's published table charges a separate, higher rate once a request passes the short context threshold: GPT-6 Sol goes from 2 dollars to 4 dollars input, from 20 cents to 40 cents cached, and from 10 dollars to 15 dollars output. An agent that quietly grows its context window across a long conversation can cross that line and double its input cost without anything else changing.

How to use this

Before you renegotiate anything, measure. OpenAI's new Prompt Caching Dashboard reports what share of your input is served from cache, and the diagnostics tool explains the misses. On the Claude side, the cache read and cache write token counts come back on every response. Pull a week of both and compute the one ratio that matters: cached input tokens divided by total input tokens.

If that ratio is above about 0.8, the September 22 cuts genuinely lowered your bill and you should move to Opus 5.5 or GPT-6 Sol on price alone. If it is below about 0.3, the cuts barely touched you, and the cheaper fix is to change when your agent runs rather than which model it calls. Keep your prompt prefix byte stable, keep tool definitions and their ordering fixed, and put volatile data such as timestamps and customer records at the end of the prompt rather than the start, because a single changed character near the front invalidates everything after it.

If you are scoping a build and want the run cost modelled before anything is written, that is the first thing we do in an AI automation engagement, and it is the question we ask first when a client wants to hire an AI engineer.

Sources: Anthropic, Introducing Claude Opus 5.5, September 22, 2026, and the Claude Platform pricing documentation. OpenAI, Introducing GPT-6 Sol and Luna and Better prompt caching for GPT-6, both September 22, 2026, and the OpenAI API pricing table. Cross checked against SiliconANGLE's report of September 22, 2026.