On September 1, 2026, Anthropic cut the price of cache reads on Claude Fable 5.1 by 75% — to $0.25 per million tokens — which Anthropic estimates lowers total cost by about 25% for typical workloads and up to roughly 45% for highly agentic ones. The coverage is all benchmarks, but for a bank, RIA, or carrier the number that matters is the cache-read line: compliance review is the most cache-heavy workload in the building, and this is the price change that flips it from pilot economics to production economics. The same announcement also makes model access a governance question: Claude Mythos 5.1 is the same model with relaxed safeguards for verified professionals, so who gets which tier — approved by whom, logged where — is now an entitlements decision, not an IT ticket.
That is the line item that decides whether your compliance review runs on AI or on associates.
On September 1, Anthropic announced Claude Fable 5.1 and Claude Mythos 5.1, and the industry coverage did what it always does: benchmarks, effort levels, leaderboards. Useful, if you are choosing a model. Irrelevant, if you already chose one and could not make the unit economics work.
Buried in the announcement is the number that changes that: cache reads — the tokens a model spends re-reading context it has already processed — now cost 75% less, at $0.25 per million tokens, wherever usage is billed by token. Anthropic estimates that cuts total cost around 25% for typical workloads and up to roughly 45% for complex, highly agentic work, while input and output pricing stays flat at $10 and $50 per million. Every pricing figure in this post is Anthropic's published pricing, from that announcement — not our analysis.
If you priced out an AI-assisted compliance review in 2025 and shelved it, this is the line item you shelved it over.
Think about what a compliance review actually reads. The same Form ADV. The same policy wrapper. The same credit memo template, the same procedures manual, the same approved-language library — re-read against every new document, every new account, every new piece of marketing, every day.
In token terms, the overwhelming majority of that workload is not new input. It is stable context, processed once and re-read thousands of times — precisely what a cache read is. A sales assistant re-reads a little context; a coding agent re-reads a repository; a compliance copilot re-reads the regulatory core of your entire firm, on every task. Anthropic's own usage data shows cache reads dominating the cost of context-heavy, tool-heavy work — the category compliance review falls squarely into.
Benchmarks tell you whether the model is good enough. The cache-read price tells you whether you can afford to run it at the volume compliance actually requires — every document, not a sample.
Most 2025 pilots failed the budget test for a simple reason: they were priced as if every token were new. Here is the before-and-after, using Anthropic's published rates:
| Line item | Before (Fable 5 rates) | After (Fable 5.1 rates) |
|---|---|---|
| Cache reads (re-reading your ADV, policies, templates) | $1.00 per million tokens | $0.25 per million tokens — a 75% cut |
| Input tokens (the new document under review) | $10 per million | $10 per million — unchanged |
| Output tokens (the review, findings, citations) | $50 per million | $50 per million — unchanged |
| Anthropic's estimated total, typical workload | Indexed at 100 | ~75 (about 25% less) |
| Anthropic's estimated total, highly agentic workload | Indexed at 100 | ~55 (up to about 45% less) |
Illustrative math on Anthropic's published rates: in a review task where most tokens are cache reads — re-reading the policy core — the cache line is the dominant cost of the program once you multiply full-context review by every document, every day, across a year. Cutting that line 75% does not shave the budget; it restructures it, which is why Anthropic's indexed estimate for cache-dominated work drops by up to roughly 45%.
That is the flip from pilot economics to production economics. A pilot tolerates a bad unit price because the volume is trivial; production cannot. The firms that ran the 2025 math honestly and walked away were not wrong then — but the input to that math changed on September 1.
Two things, both easy to miss under the pricing news.
Enterprise Frontier Safeguards (EFS). Anthropic's new safeguards architecture is designed to give enterprise customers privacy equivalent to a zero data retention policy — by storing data in cloud infrastructure controlled entirely by the customer, not Anthropic — while keeping safeguards state-of-the-art against adversarial use. It rolls out in phases beginning later this fall; until it does, eligible customers can use Fable 5.1 with zero data retention. The caveat, which we covered when the program was announced: zero data retention is not zero obligation — your recordkeeping, supervision, and contractual diligence duties travel with the data regardless of whose infrastructure holds it.
Claude Mythos 5.1. Mythos is the same underlying model as Fable 5.1 with different levels of safeguards, available only through Anthropic's trusted access programs — currently to a set of US organizations vetted for cyberdefense and life-sciences work, with broader access being coordinated with the US government. Anthropic says the capability gap between the two should narrow as safeguards improve, and its new safeguards already block 60% fewer false positives in cybersecurity contexts.
Here is the second-order point almost nobody is writing about. When a vendor ships the same model at two safeguard levels, gated by a verification program, model access stops being a procurement line and becomes an entitlements decision.
Someone inside your firm will ask for the Mythos tier — a security engineer who wants fewer false positives, a researcher who read about the biology access program. And someone has to answer: who inside the firm gets which tier, approved by whom, logged where?
That is not an IT ticket. It is a governance design decision with the same shape as entitlements you already manage:
| Governance question | What it looks like for tiered model access |
|---|---|
| Who qualifies? | Which roles map to Anthropic's verification criteria — and who inside the firm attests to that mapping? |
| Who approves? | Does tier access route through the same committee that approves other sensitive tooling, or is it an ad-hoc manager sign-off? |
| Where is it logged? | Is tier assignment recorded in your access-management system of record, reviewable in your next exam or audit? |
| How is it revoked? | Role changes and departures have to propagate to model tier, not just to systems access. |
| What usage is supervised? | Relaxed-safeguard output used in regulated work still falls under your supervision and recordkeeping obligations. |
Firms that treat this as a design decision will have a clean answer the day someone requests access. Firms that treat it as a ticket will discover their improvised answer in an exam.
Budgets for 2027 are being drafted this month. Five moves, all completable inside that window:
Vantage Point works with regulated firms on exactly this intersection — AI strategy, compliance architecture, and the governance design that makes AI programs approvable rather than alarming. Through our AI-driven personalization and analytics and compliance and security solutions services, senior consultants help you re-run the unit economics, instrument your cache ratio, design the entitlements policy, and scope a production pilot your CCO can sign. As an official Claude partner, we build on the platform whose economics this post is about. Senior consultants only — no junior handoffs; the experts you meet are the experts who deliver.
The price that shelved your pilot changed on September 1, and 2027 budgets lock this quarter. Vantage Point's senior consultants can re-price your compliance workload on the new cache economics, design your tiered-access policy, and scope a production pilot — on your budget calendar. Contact Vantage Point to schedule an AI economics review, or explore our AI services to see how we support regulated firms end to end.
Anthropic cut the price of cache reads on Claude Fable 5.1 by 75%, to $0.25 per million tokens, wherever usage is billed by token (such as the API). Input and output pricing is unchanged at $10 and $50 per million tokens. Anthropic estimates the change reduces total cost by about 25% for typical workloads and up to roughly 45% for highly agentic, cache-heavy work.
A cache read is when the model re-reads context it has already processed and stored, instead of paying full input price for it again. Compliance review re-reads the same stable core — your ADV, policy wrappers, templates, approved language — against every new document, so most of its tokens are cache reads. That is why a cache-read price cut reprices compliance work more than almost any other workload.
Anthropic frames the reduction as applying wherever usage is billed by token, such as the Claude API, and its "typical workload" savings estimate covers Fable usage across Claude Enterprise, Claude Code, and the API. Fable 5.1 is available on all platforms, including AWS, Google Cloud, and Microsoft Azure. Confirm how your specific subscription or contract passes the change through with your Anthropic account team.
They are the same underlying model with different levels of safeguards. Fable 5.1 is generally available; Mythos 5.1 is available only through Anthropic's trusted access programs, with safeguards designed for vetted cybersecurity and life-sciences work. Mythos is currently limited to a set of US organizations, with Anthropic coordinating with the US government on broader access.
Because two safeguard tiers turn access into an entitlements question: who qualifies, who approves, where the assignment is logged, how it is revoked on role change, and how output is supervised. Those are the same questions your firm already governs for other sensitive entitlements, and examiners will expect the same discipline — an improvised answer is a finding waiting to happen.
No. EFS is designed to provide privacy equivalent to zero data retention by storing data on infrastructure you control, but your recordkeeping, supervision, and diligence obligations travel with the data regardless of where it sits. We covered this in detail in our analysis of why zero data retention isn't zero obligation.
Re-run the math first. Re-price the workload at the new cache rates using Anthropic's 25–45% estimates, measure your actual cache ratio on one representative workflow, and design your tiered-access policy. If the economics flip — as they will for most cache-heavy compliance workflows — you will walk into 2027 budget season with a production-grade proposal instead of a rehashed pilot.
Vantage Point is a boutique CRM consulting firm helping businesses transform with Salesforce, HubSpot, and AI — 150+ clients, 400+ engagements, and a 4.71/5 average engagement rating. Learn more at vantagepoint.io.