A 75% Price Cut That Leaves the Chat Bill Unchanged

On 25 October Max Woolf published a tally of frontier API price changes since June, under the question “OpenAI and Anthropic cut their LLM prices by up to 80% within a few months. Why?” His table is a useful record. His explanation is that “no one cuts prices by 80% unless there is some externality in play,” and he points to open-weight competition and the coming IPOs.

Before asking why prices fell, I wanted to know by how much a bill actually falls, because the cuts are not all on the same line of the bill. A token price has at least three parts: fresh input, output, and cache reads, the cheap rate for input the provider has already seen in the same context. A cut to one of them moves a bill only as far as that part’s share of the bill.

Where each cut landed

Anthropic’s own pages make this concrete. Claude Opus 5.5, released on 22 September, costs $4 per million input tokens and $20 per million output tokens, against $5 and $25 for Opus 5. Its cache reads fell from $0.50 to $0.20. Anthropic says “Opus 5.5 requires less compute to serve than Opus 5, and its pricing reflects that.” That is a 20% cut on input and output and a 60% cut on cache reads.

Claude Fable 5.1 is the sharper case. Its announcement says “Cache reads now cost 75% less, or $0.25 per million tokens,” and that the rest of the pricing is the same as Fable 5’s: $10 per million input tokens and $50 per million output. The 80% in Woolf’s headline is OpenAI’s cut to GPT-5.6 Luna on 30 July, which by his table fell from $1, $6 and $0.10 to $0.20, $1.20 and $0.02, an 80% cut on all three lines.

Size of each price cut by part of the bill Three groups of bars on a 0 to 100 percent scale. Claude Opus 5 to 5.5: input 20 percent, output 20 percent, cache reads 60 percent. Claude Fable 5 to 5.1: input 0, output 0, cache reads 75 percent. GPT-5.6 Luna, 30 July: input, output and cache reads each 80 percent. PRICE CUT BY PART OF THE BILL · PER MILLION TOKENS Opus 5 to Opus 5.5 (22 Sep) Fable 5 to Fable 5.1 (Sep) GPT-5.6 Luna (30 Jul, per Woolf) InputOutputCache reads InputOutputCache reads InputOutputCache reads 20% — Anthropic, Opus 5.5 page 20% — Anthropic, Opus 5.5 page 60% — Anthropic, Opus 5.5 page 75% — Anthropic, Fable 5.1 announcement 80% — Woolf, 25 Oct 2026 80% — Woolf, 25 Oct 2026 80% — Woolf, 25 Oct 2026 20%20%60% 0%0%75% 80%80%80%
The biggest Anthropic cuts are on cache reads. Fable 5.1's 75% is entirely there, with input and output untouched.

What a bill does

To see what this means for a bill, I priced two workloads per million tokens, using only list prices and ignoring cache-write charges. They are my own illustrations, not anyone’s measured traffic. A chat workload has no caching: 75% of tokens are input and 25% output. An agent loop rereads its context every turn: 85% of tokens are cache reads, 10% fresh input and 5% output.

For chat, Opus 5.5 is 20% cheaper than Opus 5 ($8.00 against $10.00 per million tokens), and Fable 5.1 costs exactly what Fable 5 did ($20.00). For the agent loop, Opus 5.5 is about 28% cheaper ($1.57 against $2.18) and Fable 5.1 about 15% cheaper ($3.71 against $4.35). None of the four lands on the cut a reader takes from a headline. Anthropic’s own note for Fable 5.1 makes the same point in its own numbers: “For typical workloads, costs are reduced by around 25% relative to Fable 5,” and for highly agentic work “the savings could be up to around 45%.” Which of those a customer sees depends on how much of their traffic is cached.

Blended cost cut for two illustrative workloads Horizontal bars on a 0 to 100 percent scale. Chat workload, 75 percent input and 25 percent output: Opus 5 to 5.5 cut 20 percent, Fable 5 to 5.1 cut 0 percent. Agent workload, 85 percent cache reads, 10 percent input and 5 percent output: Opus cut 27.8 percent, Fable cut 14.7 percent. Author's arithmetic on list prices, cache writes ignored. BLENDED COST CUT · TWO ILLUSTRATIVE WORKLOADS · LIST PRICES Chat: 75% input, 25% output, no caching Agent loop: 85% cache reads, 10% input, 5% output Opus 5 to 5.5 Fable 5 to 5.1 Opus 5 to 5.5 Fable 5 to 5.1 20% — author's arithmetic, $10.00 to $8.00 27.8% — author's arithmetic, $2.18 to $1.57 14.7% — author's arithmetic, $4.35 to $3.71 20% 0% (no change) 27.8% 14.7% Per million tokens, list prices from Anthropic's pages, cache-write charges left out. The workload mixes are illustrations, not measured traffic.
A 60% or 75% cut on cache reads becomes a 15% to 28% cut on an agent bill, and nothing at all on a chat bill when it is the only cut.

Output is most of the bill

The reason is that output tokens are expensive and cache reads are cheap even before a cut. In my agent loop on Fable 5.1, output is 5% of the tokens and about two-thirds of the cost: $2.50 of $3.71. Cache reads are 85% of the tokens and, after the cut, about 6% of the cost. Cutting the cache price by 75% takes $0.64 off a bill that is still dominated by the line nobody cut.

Where the agent-loop bill goes, Fable 5 and Fable 5.1 Two stacked bars in dollars per million tokens. Fable 5: cache reads 0.85, input 1.00, output 2.50, total 4.35. Fable 5.1: cache reads 0.21, input 1.00, output 2.50, total 3.71. Output is about two-thirds of the Fable 5.1 bill. AGENT LOOP · DOLLARS PER MILLION TOKENS · FABLE 5 VS 5.1 Fable 5 Fable 5.1 Cache reads $0.85 — author's arithmetic Input $1.00 — author's arithmetic Output $2.50 — author's arithmetic Cache reads $0.21 — author's arithmetic Input $1.00 — author's arithmetic Output $2.50 — author's arithmetic $4.35 $3.71 Outlined white: cache reads Dashed: fresh input Dark: output Output is 5% of the tokens and about two-thirds of the Fable 5.1 bill.
The 75% cut shrinks the thin white slice. The dark slice, which nobody cut, is most of the bill.

This is an old point about service pricing that the headlines keep skipping. A 2024 IEEE paper on adopting LLM-as-a-service compared the offerings through the inference cost of a real chatbot case study rather than through their price sheets, which is the comparison a price sheet alone cannot give.

The question to ask a price cut

Woolf’s question is why prices fell, and his answers may be right. The question I would add comes first: which line of the bill fell, and how much of your bill is on that line? For a chat product on Fable, the answer to September’s cut is none. For an agent product that rereads long contexts, it is somewhere between a seventh and a quarter. The 80% belongs to one smaller model, cut on every line at once on 30 July.

References

  1. Woolf, M. (2026). OpenAI and Anthropic cut their LLM prices by up to 80% within a few months. Why?. 25 October 2026.
  2. Anthropic (2026). Claude Opus 5.5. 22 September 2026. Pricing for Opus 5.5 and Opus 5.
  3. Anthropic (2026). Claude Fable and Mythos 5.1. September 2026. Fable 5.1 pricing and cache-read change.
  4. Liagkou, V., Filiopoulou, E., Fragiadakis, G., Nikolaidou, M. & Michalakelis, C. (2024). The cost perspective of adopting Large Language Model-as-a-Service. 2024 IEEE International Conference on Joint Cloud Computing (JCC), 80–83.