The short version: Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on September 1, 2026, its most capable models for coding and knowledge work. The headline API price did not move at all: $10 per million input tokens, $50 per million output, exactly what Fable 5 cost. What moved is a line almost no business owner has ever read, the cache read, which dropped 75 percent from $1.00 to $0.25 per million tokens. That is the line that governs what the agentic AI tools you rent cost to operate, and you will never see the saving on an invoice.
Model launches usually invite the same two questions: is it smarter, and did it get more expensive. Today the answers are yes and no. The interesting answer is to a third question nobody asks, which is where the price actually changed.
What actually changed in Claude Fable 5.1
Open Anthropic’s published pricing table and put Fable 5.1 next to Fable 5. Base input, both $10 per million tokens. Output, both $50. Cache writes, identical at $12.50 for the five-minute version and $20 for the hour. Batch rates, identical at $5 and $25. Five of six lines are unchanged.
The sixth is the cache read, and it fell from $1.00 to $0.25 per million tokens. Anthropic’s own footnote explains why that is stranger than it looks: a cache hit on Fable 5.1 and Mythos 5.1 is priced at 0.025 times the base input price, while every other Claude model uses the standard 0.1 multiplier. The company did not discount a model. It broke its own pricing convention for these two.
Anthropic puts the practical effect at roughly 25 percent cheaper for an average workload, rising toward 45 percent for the heavily cached, agentic kind. On the capability side, the research benchmark more than doubled, with Fable 5.1 scoring 52.6 percent on Terminal-Bench-Science against 24.7 percent for Fable 5.
One more thing happened quietly today. Claude Sonnet 5, the mid-tier model, had been scheduled to rise from $2 and $10 per million tokens to $3 and $15 on September 1. The docs now state that increase “will not occur” and the lower rate is simply the standard price. Two price movements in one day, both downward, neither one advertised.
What is a cache read, and why should a small business care?
When an AI tool works on the same material repeatedly, your policy documents, your product catalogue, your booking history, it does not resend all of that every time. It stores the processed version and reads it back. Writing to that store costs more than normal input. Reading from it costs far less. The read is the operation that repeats.
This matters because of what AI tools have become. A chatbot that answers one question is mostly output tokens. An agent that works through a task over twenty steps, rereading your context at every step, is mostly cache reads. The industry spent this year shipping the second kind. Anthropic just cut the price of the operation that kind of tool performs most.
Here is the part that stings, and the reason this is worth understanding rather than celebrating. You are almost certainly not an Anthropic API customer. You rent tools built on top of it. So this saving lands on your vendor’s cost line, not yours. It becomes their margin, or their higher usage caps, or a feature that was previously too expensive to ship. Waiting for a discount to appear on your bill is the wrong posture. Asking your vendors what got cheaper is the right one, and it is a reasonable question to put to anyone selling you an agentic tool this quarter. We covered the same dynamic when three frontier models launched in four days and not one of them raised its price.
The automation number worth remembering
The benchmark that should interest a small business owner is not the research score. It is AutomationBench, where Fable 5.1 scored 31.4 percent against 17.1 percent for Fable 5.
Read that twice. The best available model nearly doubled its automation score, which is genuinely remarkable progress, and it still fails roughly two attempts in three. Hold that number the next time a vendor tells you their agent can run your back office unattended. The honest reading is that these systems have become good enough to be useful with a person checking the work, and are nowhere near good enough to be left alone with it. That is not a knock on the technology. It is the difference between a tool that gives your team leverage and a promise that quietly assumes nobody needs to be watching. We have written before about how to read a vendor’s own benchmark table without being sold by it.
What to do about it
Nothing urgent, which is itself the point. Three things are worth ten minutes.
Ask any vendor selling you an agent-style tool whether their pricing reflects this month’s model costs, because if their quote was built on last quarter’s economics there is room in it. Treat the 31.4 percent as your planning number for unattended automation rather than the demo you were shown. And if you are drafting customer-facing text with any recent Claude model, note that Fable 5.1 and Mythos 5.1 ship with Anthropic’s invisible text watermark active from launch, which changes nothing about quality but is worth knowing you are doing.
Mythos 5.1 itself is not something you will be using. It is the same model with fewer safeguards, restricted to vetted organizations, and Anthropic lists it as limited availability. It is in the announcement mainly because it shares the pricing table.
Frequently Asked Questions
How much does Claude Fable 5.1 cost?
Claude Fable 5.1 is priced at $10 per million input tokens and $50 per million output tokens, identical to Claude Fable 5. Cache writes are $12.50 per million for the five-minute cache and $20 for the one-hour cache, and batch processing runs at $5 input and $25 output. The one change is the cache read, which fell from $1.00 to $0.25 per million tokens.
Is Claude Fable 5.1 cheaper than Fable 5?
For most real workloads, yes, even though the headline rates are identical. Anthropic puts the saving at roughly 25 percent on an average workload and as much as 45 percent on heavily cached agentic work, all of it coming from the 75 percent cut to cache reads. If your usage involves no prompt caching at all, you will pay exactly what you paid before.
What is the difference between Claude Fable 5.1 and Mythos 5.1?
They are the same underlying model with different safeguards, and they share an identical pricing table. Mythos 5.1 carries fewer safety restrictions and is not generally available, being restricted to vetted organizations in areas such as cybersecurity and life sciences. Ordinary business users will be working with Fable 5.1.
Will the Claude Fable 5.1 price cut lower my software bill?
Probably not directly, because most small businesses do not buy Anthropic API tokens; they subscribe to products built on top of them. The saving lands first on the vendor’s cost of goods, where it can become margin, higher usage limits, or new features rather than a lower subscription price. The practical move is to ask vendors whether their pricing reflects current model costs before signing an annual contract.
Have you ever asked a software vendor to justify their price against what their underlying model actually costs them? We are curious whether anyone gets a straight answer.
