Quick Answer
Anthropic cut Fable’s token price 20 percent on 9 November 2026, and the measured cost to complete a benchmark task rose 31 percent, from $67.16 to $88.15. The reason is that the newer model spends far more tokens per task, so the quantity moved more than the price. Cost per point of benchmark score actually fell about 32 percent, which means the bill went up and the value per dollar improved at the same time.
The Two Numbers Everyone Is Comparing Wrong
On 9 November 2026, Anthropic launched Claude Fable 5.2 and cut the list price of a token by 20 percent, from $10 and $50 per million input and output tokens to $8 and $40.
In the same news cycle, SemiAnalysis published its GDPx agentic benchmark results, and Latent Space relayed them with the line that stuck: cost per task “rose 31% ($67.16 → $88.15) despite the 20% cheaper per-token price.”
Both numbers are true. Cheaper tokens, more expensive work. Price per token is the sticker. Cost per task is the invoice, and the invoice is tokens consumed multiplied by price per token. The first factor moves more than the second.
Here is what the GDPx table actually shows.
| Measure | Fable 5.1 | Fable 5.2 |
|---|---|---|
| GDPx score | 8.5% | 16.4% |
| Cost per task | $67.16 | $88.15 |
| Output tokens per task | 418K | 1.23M |
| Cost per point of score | $7.90 | $5.37 |
Output tokens per task went from 418,000 to 1.23 million. That is 2.9 times as many, which is a far bigger move than you would infer from the cost line alone. Work backwards from price and cost and you would guess about 64 percent more tokens. The measured number is nearly three times. Trust the measurement, not the inference, and notice that the gap between the two is where input tokens, cache reads and retries live.
The Part The Headline Leaves Out
The score nearly doubled. 8.5 percent to 16.4 percent on the same benchmark.
Divide cost by score and Fable 5.1 cost $7.90 per point. Fable 5.2 costs $5.37. That is about 32 percent cheaper per unit of measured capability, in the same release where the bill per task went up 31 percent.
So “cheaper tokens, bigger bill” is accurate and incomplete. The honest version is that you are being charged more per task and getting more per dollar. Whether that is good news depends entirely on something no benchmark can tell you: whether the extra capability clears your bar for the task you are actually running.
If you are classifying support tickets, a 93 percent improvement on a hard agentic benchmark buys you nothing, and you just paid 31 percent more for it. If you are running a long multi-step build where a failure costs a day, it may be the cheapest thing you bought this year. Same model, same release, opposite conclusions, and the deciding variable is on your side of the API.
Why This Keeps Happening
This is Jevons, and it is not a metaphor. Make a resource cheaper per unit and total consumption of that resource rises, often by more than the price fell. Every efficiency gain in something you pay for per unit gets spent on more units.
With models the mechanism is specific. Cheaper tokens let a lab spend more tokens per answer. More tokens per answer means longer reasoning chains, more tool calls, more retries, more verification passes. The model gets better by thinking longer, and thinking longer is the thing you are billed for.
Which means the price cut is not a discount you receive. It is a budget the lab spends on your behalf, and the receipt arrives in your usage dashboard. Expect this pattern to repeat at every release. The per-token line will keep falling. Your monthly bill will not.
The operator’s job therefore moves one layer up. You stop negotiating the price of a token, because you cannot, and you start controlling how many of them a task consumes, because you can.
Three Controls That Decide What You Actually Pay
One: Meter The Outcome, Not The Token
Log tokens, retries and tool calls per completed task, and compare that number across model versions before you migrate a workflow. A price page cannot tell you what your task costs. Only your own logs can.
The unit matters more than people think. Cost per token is the vendor’s unit. Cost per thousand tokens is the dashboard’s unit. Cost per completed task is yours, and it is the only one that maps to a business decision, because it is the only one that includes the failures.
A task that succeeds on the first attempt at 40,000 tokens is cheaper than one that succeeds on the third attempt at 18,000 tokens each. The second looks better on every per-call metric and costs more. If your logging stops at the call level, you cannot see that, and you will migrate to the model that makes your dashboard look good.
Three fields are enough to start: task ID, total tokens across every call in that task, and whether it completed without human intervention. Everything else is refinement.
Two: Set Effort By Task Class
The default in most apps is a middle effort setting for a reason. Classification, extraction and formatting belong on low. Maximum is for work where an error costs more than the thinking.
Most workflows have a handful of steps that deserve the expensive setting and a long tail that does not. The mistake is setting effort once at the workflow level, which means either paying maximum for a formatting step or paying minimum for the step that actually needed the reasoning.
The sorting rule I use: ask what a wrong answer costs. If the answer is “a retry”, use low effort. If the answer is “a client sees it”, use high. If the answer is “we do not find out for a month”, use maximum and add a verification step on top.
This is where the GDPx numbers become operationally useful rather than just interesting. A 93 percent improvement on hard agentic work tells you the maximum setting got meaningfully better. It tells you nothing about whether your extraction step should move, and the extraction step is probably 80 percent of your volume.
Three: Respect The Cache
This is the largest single lever and the one most teams leave on the floor.
| Line | Per Million Tokens | Versus Fresh Input |
|---|---|---|
| Fresh input | $8.00 | Baseline |
| Cache read | $0.20 | 2.5 percent |
| Cache write, 5 minute | $10.00 | 125 percent |
| Cache write, 1 hour | $16.00 | 200 percent |
| Output | $40.00 | 500 percent |
| Batch input / output | $4.00 / $20.00 | 50 percent discount |
A cache read costs 2.5 percent of fresh input. That is not an optimisation, it is a different order of magnitude, and it is the difference between a workflow that is viable at scale and one that is not.
The trap is the write cost. A five minute cache write costs more than fresh input, 125 percent of it, and a one hour write costs double. You only come out ahead if you read from that cache enough times before it expires. One write and one read is a loss. One write and ten reads is an enormous win.
Which makes the lifetime a design decision rather than a setting. A workflow that pauses longer than five minutes between calls quietly loses the discount and pays the write premium for nothing. If your steps are minutes apart, either pay for the one hour window deliberately or restructure the workflow so the calls land closer together.
Batch is the other underused line. Half price on input and output, for anything that does not need to be synchronous. Most reporting, enrichment, classification and summarisation work does not need to be synchronous, and gets run synchronously anyway because that is how the first version was built.
What I Would Check This Week
- Pull cost per completed task for your top three workflows. Not cost per call. If you cannot produce that number today, that is the first thing to build, and it is an afternoon of work.
- Find your cache hit rate. If it is low and your prompts share a large stable prefix, you are paying roughly forty times more than you need to on that portion.
- Audit effort settings per step, not per workflow. Look for formatting and extraction steps running on a high setting. They are the easiest savings in the stack.
- Move anything asynchronous to batch. Half price, and most of what you run does not need an answer in two seconds.
- Re-run your migration decision with the task number. A newer model that costs 31 percent more per task and solves 93 percent more of them is a bargain on hard work and a waste on easy work. You cannot tell which you have without your own measurement.
Frequently Asked Questions
Did Anthropic lower Claude’s price in November 2026?
Yes. On 9 November 2026 Anthropic launched Claude Fable 5.2 and Mythos 5.2 and cut the list token price 20 percent, from $10 and $50 per million input and output tokens to $8 and $40. The reduction applies to both input and output on the base rates.
Why did my AI bill go up when token prices went down?
Because newer models spend more tokens per task. On SemiAnalysis’s GDPx agentic benchmark, output tokens per task rose from 418,000 on Fable 5.1 to 1.23 million on Fable 5.2, nearly three times as many. Cost per task rose from $67.16 to $88.15, a 31 percent increase, even with tokens priced 20 percent lower. Your invoice is tokens consumed multiplied by price per token, and the quantity moved more than the price.
Is Fable 5.2 more expensive than Fable 5.1?
Per task on hard agentic work, yes, 31 percent more. Per unit of capability, no. GDPx scores were 8.5 percent for Fable 5.1 and 16.4 percent for Fable 5.2, so cost per point of score fell from about $7.90 to about $5.37, roughly 32 percent cheaper. Which figure matters depends on whether your workload needs the extra capability.
How much does prompt caching save?
A cache read is reported at $0.20 per million tokens against $8 for fresh input, which is 2.5 percent of the cost. The catch is the write: $10 per million for a five minute lifetime and $16 for one hour, both more than fresh input. You profit only if you read from the cache several times before it expires, so workflows with long pauses between calls can pay the write premium and never collect the discount.
What is the Jevons paradox in AI?
It is the observation that making a resource cheaper per unit tends to increase total consumption of it, often enough to raise total spend. With language models, cheaper tokens let labs spend more tokens per answer through longer reasoning, more tool calls and more retries. The per-token price falls and the per-task bill rises, which is exactly what the November 2026 figures show.
The Bottom Line
The per-token price is going to keep falling, and it is going to keep not mattering. It is the one number the vendor controls and the one number that does not appear on your invoice.
What appears on your invoice is quantity, and quantity is a design decision: which model, at what effort, with what cache strategy, retried how many times, for a task defined how tightly. Every one of those is yours.
So stop reading price announcements as good news or bad news. Read them as a change in the inputs to a calculation you should already be running, and make sure you are running it on cost per completed task. If you are not measuring that, a 20 percent price cut and a 31 percent cost increase look identical from where you are standing, which is to say they look like nothing at all.
Sources
- SemiAnalysis, GDPx: The General Agentic Benchmark, for GDPx scores, cost per task and output tokens per task.
- Latent Space, Fable 5.2 retakes world number one LLM spot, for the relayed cost per task comparison.
- ExplainX, Claude Fable 5.2 and Mythos 5.2 launch pricing, for the 9 November 2026 launch date and the $8 and $40 base rates.
- Kingy, How to get Fable 5.2: plans, limits, API pricing, for cache read, cache write and batch rates.
Pricing and benchmark figures were accurate as reported in November 2026 and change often. Verify against Anthropic’s live pricing page and the GDPx leaderboard before making a budget decision. Nothing here is financial advice.
