Claude Fable 5.2 Is 20% Cheaper Than 5.1. Against Opus 5.5, Your Cache Decides the Gap

What changed on November 9

Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on November 9, 2026. The announcement says they come at a lower price but does not say by how much. The pricing page does: every Fable 5.2 rate is 20% below Fable 5.1.

We added both models to our pricing dataset the next morning, along with the Amazon Bedrock and Google Vertex AI listings. This post uses those records to answer the question I actually care about: what happens to a bill. All prices are per million tokens, standard processing, global endpoint.

Two releases, two different cuts. Fable 5.1 left input and output alone and cut cache reads to a quarter. Fable 5.2 cut everything by 20%. Mythos 5.2 has the same rates as Fable 5.2. Live rates are on the Claude Fable 5.2 pricing page, and every Claude change since 2024 is in the LLM API price history.

The four workloads I priced

The token counts below are assumptions, not measurements from a real Fable run. I kept them identical across models, so the only thing that changes between columns is the rate card.

• Coding agent, caching on: 30,000 tokens of system prompt, tools and repo context, then 40 steps. Each step writes 800 output tokens and gets back a 3,000-token tool result, and every step resends the whole conversation. Best case: each step reads the previous request from the 5-minute cache and writes only the new tokens. That is 4,164,000 input tokens and 32,000 output tokens per session.
• The same agent with no cache hits, which is what you get when something at the top of the prompt changes every call or the cache expires between steps.
• A long written answer: 6,000 tokens in, 10,000 out.
• Batch extraction: 20,000 tokens in, 800 out, through the Batch API.

From 5.1, the cut is exactly 20%, whatever you run

Every row drops by 20.0% from Fable 5.1 to 5.2. That is not a coincidence of the shapes I picked. Input, cache writes, cache reads, output and Batch all moved by the same factor, 0.8, so any mix of them moves by 0.8 too. Cache-heavy or output-heavy, you save the same share.

That makes the upgrade easy to budget. If Fable 5.1 cost you $10,000 last month and your token counts stay the same, Fable 5.2 costs $8,000. The tokenizer did not change between 5.1 and 5.2, so the same prompt is the same number of tokens. The only open question is whether the new model produces the same number of thinking and output tokens, which I come back to below.

From Fable 5, a cached agent session costs half

If you skipped 5.1, the picture is different. The cold agent, the long answer and Batch all drop 20%, because Fable 5 and 5.1 had the same input and output rates. The cached agent drops 50.6%, because the two cuts stack: 5.1 took cache reads from $1.00 to $0.25, and 5.2 took 20% off everything.

On Fable 5, cache reads were 51% of this session's bill. Now they are 21%. The biggest line is cache writes.

Writes cost 50 times as much as reads on Fable 5.2, $10 against $0.20. So the cheapest thing you can do is to stop paying for the same tokens to be written twice. Keep the system prompt and tool list byte-identical between calls, put anything that changes after them, and do not use the 1-hour cache ($16 per million to write) unless your steps really are more than 5 minutes apart.

If you are not sure how often your agent actually hits the cache, we priced what a 70% vs 90% cache hit rate does to an agent bill across 72 models.

Fable 5.2 vs Opus 5.5: 2x on paper, 1.66x with a working cache

Opus 5.5 lists $4 input and $20 output, exactly half of Fable 5.2. On the cold agent, the long answer and Batch, Fable 5.2 costs exactly twice as much. But cache reads cost $0.20 per million on both models. Anthropic prices Fable cache reads at 2.5% of input and Opus 5.5 at 5%, which lands on the same number. In the cached agent session, Fable 5.2 costs $3.86 against $2.33, 1.66 times.

The last column is the break-even. If Fable 5.2 finishes your cached agent task with about 40% fewer tokens than Opus 5.5 (fewer steps, shorter thinking), it costs the same. Anthropic claims it does better than that on some benchmarks: about half Opus 5.5's cost per task on FrontierCode 2.0 at high effort, and Opus 5.5's top score on WANDR for about 40% of the cost. Those are Anthropic's numbers, run on Anthropic's harness, often against Opus at a higher effort level. I have not reproduced them.

One trap if you move from Opus 5.5: its default effort is medium, and Fable 5.2's is high. A request that does not set effort runs one level higher after the switch, with no error, only a different bill. Set effort explicitly on both before you compare.

Opus 5.5's rates are on the Claude Opus 5.5 pricing page. For a cross-vendor comparison, see GPT-6.1 Sol vs Claude Fable 5.2.

Fast mode brings back the June price

Fable 5.2 has a fast mode at $16 input and $80 output, double the standard rates. Anthropic says it generates tokens up to 2.5 times faster. The cache multipliers apply on top, so fast-mode cache reads are $0.40 and 5-minute writes $20. Fast mode does not work with the Batch API.

Our cached agent session costs $7.72 in fast mode, almost exactly the $7.81 it cost on plain Fable 5 in June. The cold agent costs $69.18 per session. If your agent waits on people, fast mode can make sense for the interactive turns. For background jobs it is the most expensive way to run this model.

What can eat the 20%

The rate cut is certain. Your token counts are not. These are the changes in Anthropic's Fable 5.2 migration guide that move tokens, not prices:

• Thinking is always on, as on 5.1, and lines like "answer directly without deliberating" no longer stop it at medium effort and above. If you relied on that to keep 5.1 short, expect more output tokens.
• At xhigh and max effort Anthropic saw Fable 5.2 start subagents to review its own work. Each subagent is more tokens.
• If your code edits earlier turns, the thinking blocks no longer match. The default returns an error; the drop_block option keeps going but can raise output tokens.
• A switch rebuilds the prompt cache. The first requests on the new model pay write prices for prefixes that were being read at $0.25 on 5.1.
• The API default effort is still high. Anthropic's token-saving claim is for medium.

None of this changes the arithmetic of the price table. It changes the token counts you feed into it. Take 200 real requests, run them on both models at the effort you plan to use, and compare tokens per finished task before you move all traffic.

Mythos 5.2, Bedrock, Vertex and the price tables

Mythos 5.2 is the same model as Fable 5.2 at the same prices, but only for organizations approved through Anthropic's verification programs. Anthropic names a Cyber Verification Program and a Life Sciences Verification Program. Claude Enterprise customers get Mythos 5.2 inside Claude Security without joining a program. If you do not have that approval, plan on Fable 5.2.

Anthropic lists Fable 5.2 on the Claude API, Amazon Bedrock, Claude models on Google Cloud and Microsoft Foundry. Cloud list prices and their quirks, such as regional endpoint premiums, are on the Claude on Bedrock and Vertex pricing page.

If your cost tracking runs through an open-source price table, it already knows the new model. I checked the commit history of the four tables we track. LiteLLM added claude-fable-5-2 at 19:35 UTC on November 9 and models.dev at 19:40. pydantic's genai-prices had it by 23:21 the same day, and Portkey's model repository at 11:15 UTC on November 10. All four list $8 input, $0.20 cache reads and $40 output.

That matches what we found for Anthropic launches in general: listed within a day or so, while smaller labs can take weeks. The full comparison of 31 launches is in how long LLM price tables take to list new models.

Check your own bill

new_cost  = 0.8 × fable_5_1_cost                       # same token counts

agent step k (cached):
  input  = previous_context × cache_read + new_tokens × cache_write_5m
  cost   = (input + output_tokens × output) / 1e6

tokens Fable 5.2 must save to tie Opus 5.5 = 1 - opus_cost / fable_cost

Every number in this post comes from a short script that reads the prices from our dataset and runs these lines, so it can be re-checked when a price changes. For your own traffic, the model switch calculator prices your token counts on any two models, and the prompt caching savings calculator shows what your cache hit rate is worth. Anthropic's own rates are on its pricing page and in the Fable 5.2 migration guide.

FAQ