Claude Is 20% Cheaper, Output Still Costs 5x

Anthropic cut the price of its top models by 20%¹ on November 9, 2026. Claude Fable 5.2 and Claude Mythos 5.2 now list at $8 per million input tokens and $40 per million output tokens². Anthropic's earlier list price for the 5.1 generation was $10 and $50. Anthropic also says the new models write up to 30% faster.

A token is a chunk of text, usually part of a word. On paper the cut looks clean. Now look at the ratio: $40 divided by $8 is 5.² Every token the model writes costs five times what every token it reads costs.

Agents write a lot. They plan, draft code, retry and draft again. So the same 20%¹ cut is worth $2.50 on one job and $22 on another, using Anthropic's own rates. Everything below is arithmetic on Anthropic's published prices.

Your token mix sets the saving

The mechanism is output share, meaning the slice of your bill that comes from tokens the model writes. Anthropic's cut applies to both rates, so the percentage saving is 20%¹ everywhere. What moves is the dollars, and output share decides how many of them you keep.

A balanced agent that reads 1 million and writes 200,000 costs $16,² down from $20.³ An output-heavy job that reads 1 million and writes 2 million costs $88, down from $110.

On the same 1 million input tokens, the saving runs from $2.50 to $22. Output share explains the spread. In the retrieval job, output is 20%¹ of the bill ($2 of $10). In the output-heavy job it is about 91% ($80 of $88)².

The day's facts sort into four kinds of efficiency, and the price cut is only one. Price efficiency is dollars per token, cache pricing included. Throughput is tokens per second, which is where the 30%² speed claim and fast mode live. Task efficiency counts tokens per finished job. Workflow efficiency counts retries and human fixes.

We think the last two decide agent economics. Anthropic's model page pitches Fable 5.2 at "agent sessions that run for hours," and those sessions are mostly writing. Mythos 5.2 is the same model at the same price, limited to organizations Anthropic verifies through programs like its Cyber Verification Program. For most builders, the meter runs on Fable.

Where agent dollars leak past the discount

Picture an agent as a contractor who charges one rate to read your files and five times that rate to write anything. You would watch the writing hours closely, because that is where the bill lives. Check it in this order.

Step one is knowing what speed buys. Speed cuts wait time. It does nothing to the bill if the model writes the same number of tokens. An agent that waits on a browser or a human approval only gains on the writing part of its run.

Step two is pricing fast mode honestly. Anthropic lists fast mode at $16 per million input and $80 per million output², double the base rate, for up to 2.5 times faster generation³. That trade pays when a live user is waiting and a quicker answer keeps them. On an overnight job it doubles the bill for speed nobody sees.

Step three is moving repeated context into the cache. A cache read means the model reuses text it already processed, like a long system prompt or a codebase. Anthropic prices cache reads at $0.20 per million tokens², which is 2.5% of the $8 uncached input rate. A pipeline that resends a big codebase every turn can save more here than the headline cut gives it.

Step four is sending non-urgent work to batch. With batch, you submit jobs and collect results later, and nobody waits. Anthropic lists batch at $4 input and $20 output³, half the standard rates. Code migrations and offline research fit here.

Step five is capping output. Fable 5.2 can write up to 128,000 tokens³ in one reply, per Anthropic's model page. At $40 per million², one maxed-out reply costs $5.12. A 100-turn agent writing 20,000 tokens a turn produces 2 million output tokens, or $80 before a single input token is counted.

The quietest leak is longer output. Let output grow 25% and the 20%¹ price cut is gone: 1.25 million tokens at $40 costs $50,² the old price of 1 million.

Long runs compound failure too. If each step of a 20-step agent succeeds 95% of the time with no recovery, the whole run finishes about 36% of the time. Failed runs still bill their tokens. A cheaper token on a run that fails is money spent on nothing.

Output may near $20 within three releases

Anthropic's own price sheet shows the trend. Fable output listed at $50 per million in the 5.1 generation². Fable 5.2, released November 9, 2026,³ lists at $40. On the same model page, Anthropic lists Opus 5.5 at $20 output, Sonnet 5.5 at $10 and Haiku 5.5 from $0.50.¹

If the 20%¹ step repeated each release, three more cuts would take frontier output to about $20.48 ($40 × 0.8 × 0.8 × 0.8)². That is roughly today's Opus 5.5 price. Anthropic has not said when the next cut comes, so the dates stay open. The direction matters more: the top tier keeps sliding toward the tier below it.

Cheaper tokens tend to buy more tokens. Say a team spends $1 million a year on 5.1.² The 20%¹ cut drops that to $800,000, but only if usage holds still. If the lower price lifts usage by 50%, the bill becomes $1.2 million ($1,000,000 × 0.8 × 1.5).

The risk runs one way. A team that tracks cost per finished task turns every future price cut into margin, and that margin compounds with each release. A team that tracks only price per token can ship a cheaper model and a bigger invoice in the same month. Over the next five years we expect the model to keep changing under you, so the lasting advantage sits in the meter you build around it.

Price every agent run by Friday

The build is a small cost meter you can rerun after every model change. It reads your call logs and prints one number: cost per successful task. That means total spend divided by the tasks that passed your check.

First, log five fields on every model call: input tokens, cache-read tokens, output tokens, retry count and whether the task passed. Write them to a CSV file with a task ID so retries group together. A plain spreadsheet is enough to start.

Second, apply Anthropic's rates. Cost per call is input millions times $8,² plus cache-read millions times $0.20, plus output millions times $40. If a job runs through batch, use $4 and $20.³ Add tool and human review costs where you know them.

Third, sum cost by task ID and divide by passed tasks only. Failed runs stay in the total, because you paid for them. That figure is your cost per successful task.

Fourth, compute output share: output cost divided by total model cost. Above 50%, cut output first with tighter reply formats and a hard max_tokens cap. The max_tokens setting limits how long a single reply can run. Below 20%¹, look at your cache hit rate before anything else.

Fifth, set a run budget. Stop any task that passes a fixed turn count or a dollar ceiling, such as $5. Log every stop as a failure so the meter stays honest.

Expect something to break: a small prompt edit kills your cache hits, or a cap truncates a JSON reply halfway. Fix it, rerun and compare the two cost-per-successful-task numbers. That gap is what the 20%¹ cut is actually worth to your pipeline.

Practise it step by step: Diagnose and Fix Bad Output