Anthropic shipped Fable 5.1 today with a lot of coverage going to benchmarks. To me, the jumps look real. Terminal-Bench 4.0 increased from 42.0% with Fable 5 to 55.8% on Fable 5.1, and the "agentic science" benchmark more than doubled, going from 24.7% to 52.6%.
If you already have Claude in prod, hopefully you will now be exited about your cache bill (because it should be lower). Cache reads now cost $0.25 per million tokens, which is a 75% cut. Anthropic estimates that will be roughly 25% lower cost on typical deployments and closer to 45% on heavily agentic ones.
Your savings depend entirely on one ratio
Input is still $10 per million tokens and Output is still $50 per million tokens. Neither have changed with this update. The only price that has changed is the cache read itself, which is what you pay every time the model re-reads context it has already processed.
Prices in this post are Anthropic's published list prices for Claude Fable 5.1 as of September 3, 2026. Check Anthropic's pricing page for current rates. Merito does not set or discount Anthropic's API prices.
That makes the arithmetic is simple. Your savings work out to about 75% of whatever share cache reads represent of your current bill.
If cache consumption is a third of your spend, then you save around 25%. If they are 60% of your spend, you save ~45%. If you are barely caching at all, this release saves you close to nothing, but congrats, because you now know you can cache and it will save you a few bucks moving forward.
Most teams I talk to have no idea which bucket they are in. That is the actual problem I get excited about solving.
An example
Take an agent running code against a large repo. The system prompt, the repo context/map, and all instructions come to roughly 150,000 tokens of cached prefix, and a working session hits the model 40 times. That is 6 million cache read tokens in a single session.
At the old rate of $1.00 per million, those reads cost $6.00. At $0.25 per million, they cost $1.50.
Now scale it across your team. Fifty engineers, four sessions a day, twenty working days. That is 4,000 sessions a month. Cache reads now drop from about $24,000 to about $6,000.
Whether that $18,000 shows up as a 25% cut or a 45% cut to your invoice depends on what other line items are on the invoice. Long output-heavy generation will naturally move you toward the low end. Long-running agents that reread a big context over and over, sit at the high end.
Of course these numbers are illustrative and your prefix size and turn count will differ. The equation stays the same so pull your actual token mix, split it into cached reads, uncached input, and output, and you have your answer.
Where the savings leak back out
We have worked customers through enough Claude bills to know the money rarely stays saved. Usually, there are three patterns account for most of it.
Cache misses nobody is watching. A prompt prefix that changes on every call is something that never caches. Small things break it, like a timestamp injected at the top of a system prompt, or per-user data placed ahead of the stable content instead of after it. Teams often run this for months without noticing because the bill looks like usage growth rather than waste.
Everything running on the top model. Fable 5.1 earns its price on the longest and most autonomous work. Plenty of the traffic underneath it is routine, and routing that to a smaller model in the lineup costs a fraction. Cheaper cache reads make it tempting to stop thinking about model selection, which is a great opportunity for cost reduction.
No owner. Consumption governance is an island between platform engineering and finance, and in most organizations that means it sits nowhere. Usage climbs, somebody escalates, and the response is a spending cap rather than an AI FinOps Optimization.
What we would do in the next thirty days
Start by gathering metrics and measuring rather than migrating. Pull thirty days of token usage and break it into cached reads, uncached input, and output, per workload. You'll want the cache read share by workload, because a single blended average across the company doesn't tell the entire story. I'd even argue it hides the workloads that would benefit most.
Then find the workloads with a large stable prefix and a high turn count. Those are your agentic candidates and they have the strongest business case. Restructure prompts so the stable content sits at the front and the variable content sits behind it, then confirm the cache is actually being hit rather than assuming it.
After that, set a model tiering policy and document it. Decide which workloads justify frontier pricing and which ones run on something smaller, and revisit it quarterly as the lineup shifts.
Finally, run the numbers again on the pilot projects didn't fund. Several agentic use cases that failed a business case six months ago might make sense at these new prices.
One more thing worth watching
Anthropic also announced Enterprise Frontier Safeguards, rolling out in phases starting this fall. It pairs the privacy of zero data retention with misuse detection, keeps data in cloud infrastructure the customer controls, and requires no Anthropic human review of your data.
For regulated buyers who have been stuck between a security team that wants zero data retention and a business that wants frontier capability, that is the constraint loosening. It deserves its own post, and it will get one.
How Merito helps
Cost work on Claude is the part customers ask us about after their pilot succeeds and the invoices start rolling in. AI FinOps Optimization covers usage visibility, model selection, caching strategy, and the cost governance that keeps all of it from drifting back.
If you are earlier in the journey, AI Readiness and Governance handles the data, identity, and model governance controls your security team will ask about, and AI Rollout and Activation gets Claude deployed against your real release and identity controls with adoption that holds. For teams building on the API rather than buying seats, Agentic Build takes a prototype to production.
We wrote previously about why Claude adoption is already happening whether or not you control it. The pricing change makes that more urgent, since cheaper agents mean more of them, faster, with or without a policy behind them.
If you want the token mix analysis run against your own usage, book a scope conversation and we will get you off to the races.
