Anthropic just made its top-tier Claude model cheaper to run, but not in the way most headlines are describing it. Nobody cut the price per token.
What actually got cheaper is something most people never think to check: how much you pay when the model reuses something it already read.
Get that one detail right and you can genuinely cut your Claude costs by close to half on the right kind of work. Miss it, and the update barely touches your bill at all.
Key Takeaway
Claude Fable 5.1 launched September 1, 2026, with the same base price as Fable 5, $10 per million input tokens, $50 per million output. The real savings come from cheaper cache reads, up to 45 percent lower on highly agentic, cache-heavy work. It also gained the ability to find software vulnerabilities, though not exploit them. Mythos 5.1, the same underlying model with looser safety filters, stays restricted to vetted cybersecurity and life sciences programs. Here is what to actually do with all of that.
Where Fable Sits in Claude’s Lineup
Claude comes in a few sizes. Haiku is small and fast. Sonnet is the everyday middle option. Opus is the large, capable model most serious work runs on. Fable sits above all three, Anthropic’s most capable tier for coding, research, and long, complicated problem-solving.
Mythos is not a separate size. It is the same underlying model as Fable, just with different safety filtering.
Fable 5.1 is available to anyone with a paid Claude plan or an API account. Mythos 5.1 is not. It only opens up through Anthropic’s trusted access programs, built specifically for vetted cybersecurity defenders and life sciences researchers who need the filters loosened for legitimate work. If you are not enrolled in one of those programs, Fable is the version you will actually use.
What is the Real Cost Savings
Base pricing did not move. Fable 5.1 costs exactly what Fable 5 cost, $10 per million input tokens, $50 per million output.
What changed is the price of a cache read, the cost of Claude reusing context it already processed earlier in a session instead of reading it fresh.
Every time you send a long coding session, a research thread, or an agent workflow back to Claude, a big chunk of what you are paying for is Claude re-reading context it has already seen.
Cutting that price is what delivers Anthropic’s 25 percent typical savings, and up to 45 percent for work that leans heavily on repeated context, long coding sessions, multi-step agent tasks, anything with a lot of back-and-forth built on the same foundation.
Here is what that means practically. If you use Claude for short, one-off questions with little repeated context, you will barely notice this update in your bill.
Again if you run long coding sessions, multi-turn agent workflows, or anything where Claude keeps referring back to the same codebase or document, the savings are real and worth restructuring your workflow around.
The practical move: keep related work inside the same conversation or session instead of starting fresh each time, when that makes sense for the task.
Every time you start over, you lose the cache benefit and pay full price to re-establish context. Long, continuous sessions on the same project now cost meaningfully less than starting a new one for every small step.
What Changed With the Safety Filters
Anthropic tightened the filters without making them stricter, they made them more accurate. False positives, cases where the system flags something completely harmless, dropped 60 percent in cybersecurity contexts and 85 percent on ordinary medical questions.
If you have ever had Claude decline a completely reasonable security or health-related question because it misread the intent, this update is aimed directly at that problem.
One genuinely new capability came with it. Fable 5.1 can now be used to discover software vulnerabilities, the defensive side of security work. It still will not write exploit code for those vulnerabilities.
That line matters, and it is enforced deliberately. If your work involves actual penetration testing or exploit development, that still routes to Opus models, not Fable.
If You Work in Life Sciences or Security Research
Anthropic opened a new Life Sciences Verification Program alongside this release, built with the US government, specifically to give vetted scientists access to Mythos 5.1’s expanded biology capabilities.
Open enrollment for scientists is coming soon rather than available immediately. If this applies to your work, that program, not a regular Claude subscription, is the path in.
Worth knowing the backstory here. The earlier Mythos model drew real scrutiny over its cybersecurity capabilities, serious enough that the US government ordered a temporary suspension before access was restored. Mythos 5.1’s restricted, program-only access is a direct response to exactly that concern, not an arbitrary limitation.
What to Actually Do This Week
If you already pay for Claude, you do not need to change plans or do anything to get Fable 5.1. It is live now on every platform. What you should do is look at how you currently structure long sessions.
If you have been closing out conversations and starting fresh for every small task, try keeping related work in one continuous thread instead, and watch what happens to your usage.
If you build with Claude the way we did when testing its artifact feature to build three working apps in an hour, this update is worth testing directly on a real project, since agentic, multi-step work is exactly where the cache-read savings show up most.
And if you are weighing Claude’s pricing against other AI subscriptions in general, this update is worth reading alongside Anthropic’s decision to lock in permanent pricing on Claude Sonnet 5. Two different models, same underlying pattern: Anthropic making its pricing more predictable at a moment when it has every reason to want developers staying put.
Join the discussion Tap to open the comment form +