On September 1, 2026, Anthropic quietly shipped the next chapter of the model line it launched with Claude Fable 5 back in June — and buried inside the release notes was a pricing change that matters more to anyone building on Claude than the headline benchmark numbers do. Claude Fable 5.1 and Claude Mythos 5.1 are, under the hood, the same model wearing two different safety collars. But the number every developer actually cares about — what it costs to run this thing at scale — just dropped by three-quarters on one of its biggest line items.
Anthropic released Claude Fable 5.1 (generally available) and Claude Mythos 5.1 (restricted access) on September 1, 2026. Base API pricing is unchanged at $10 per million input tokens and $50 per million output tokens, but cache-read pricing on Fable 5.1 falls 75%, from $1.00 to $0.25 per million tokens — cutting effective costs by roughly 25% on typical workloads and up to roughly 45% on heavily agentic ones. Anthropic also announced Enterprise Frontier Safeguards, a free security architecture letting enterprises store Claude usage-monitoring data in their own AWS S3, Azure Blob Storage, or Google Cloud Storage under their own encryption keys, rolling out in phases starting this fall.
What Actually Changed in the Pricing
If you priced out Claude Fable 5 back in June, the headline rate hasn't moved: $10 per million input tokens and $50 per million output tokens, same as before. What Anthropic changed is the price of reading from the prompt cache — the mechanism that lets a model reuse context (a long system prompt, a tool schema, a codebase snippet) across calls instead of reprocessing it every time. That rate falls from $1.00 to $0.25 per million tokens, a 75% cut. Cache-write pricing, what you pay the first time content goes into the cache, stays where it was: $12.50 per million tokens for a 5-minute cache and $20 per million tokens for a 1-hour cache.
That distinction — write once, read many times — is exactly why this cut matters more than it looks on paper. Anthropic's own estimate is that the cheaper cache reads lower effective costs by roughly 25% for a typical workload, and by as much as roughly 45% for workloads that lean heavily on agents: think a coding agent that keeps re-reading the same repository context across dozens of tool calls in a single session.
Why Cache Reads Are Where the Real Bill Hides
Most people comparing AI pricing look at the input/output rate on a model's pricing page and stop there. That's fine for a one-off chat message. It's a bad way to estimate cost for anything agentic — a coding assistant, a research agent, a customer-support bot with a long knowledge base stapled to every call — because those workloads spend far more tokens re-reading cached context than they do on the actual new input each turn. A model that looks pricier on input/output but cheap on cache reads can end up costing less in practice than one with a lower headline rate and no caching discount at all. Fable 5.1's cut is Anthropic explicitly optimizing for that second category of user, which happens to be the fastest-growing one.
Mythos 5.1: Same Model, Locked Behind Two Different Doors
Fable 5.1 is available to anyone with an API key. Claude Mythos 5.1 — described by Anthropic as the same underlying model with a different safeguard configuration — is not. Access is currently restricted to the Life Sciences Verification Program, limited to organizations in the US; Anthropic says the Cyber Verification Program — used today for defensive cybersecurity work with certain Opus- and Sonnet-class models — will extend to Mythos-class access "in the near future," but does not grant it yet. The logic is the same one Anthropic used when it first split Fable 5 and Mythos apart in June: a model capable enough to meaningfully help with offensive cyber techniques or dangerous biological research needs gates that a general-purpose coding assistant doesn't. Separately, Anthropic says Claude Code users overall can expect about 60% fewer safeguard interventions per session on Fable 5.1 than on Fable 5 — a broad improvement for all users, not something tied specifically to Mythos access.
Enterprise Frontier Safeguards: Privacy Without Giving Up Monitoring
The second announcement on September 1 wasn't a model at all. Enterprise Frontier Safeguards (EFS) is a free security architecture Anthropic is building with enterprise customers to solve a tension that's dogged every AI vendor: companies want zero data retention for privacy and compliance reasons, but they also want Anthropic to be able to detect misuse of its own models on their account. EFS tries to have both by letting the usage-monitoring data live in the customer's own storage — their AWS S3 bucket, Azure Blob Storage, or Google Cloud Storage — encrypted with keys the customer controls, not Anthropic. Automated misuse detection can still run against that data without Anthropic ever holding a durable copy of it.
EFS is rolling out in phases starting this fall, not all at once. In the meantime, Anthropic says eligible enterprise customers already get zero data retention on both Fable 5 and Fable 5.1 while the fuller architecture comes online. It's a notable move because it's free — a security architecture this specific is usually the kind of thing vendors reserve for the highest enterprise tiers, not something bundled in at no extra cost.
| Metric | Fable 5 (June 2026) | Fable 5.1 (Sept 2026) |
|---|---|---|
| Input tokens (per million) | $10 | $10 (unchanged) |
| Output tokens (per million) | $50 | $50 (unchanged) |
| Cache read (per million) | $1.00 | $0.25 (−75%) |
| Cache write, 5-min (per million) | $12.50 | $12.50 (unchanged) |
| Cache write, 1-hr (per million) | $20 | $20 (unchanged) |
| Terminal-Bench-Science 0.1 | 24.7% | 52.6% |
| Terminal-Bench 4.0 | 42.0% | 55.8% |
| AutomationBench | 17.1% | 31.4% |
What This Means If You're Comparing Models on Cost
For a developer or student stitching together an agent, a study bot, or a small SaaS tool, the practical takeaway isn't "Anthropic got cheaper" in the abstract — it's that the shape of your workload determines how much this actually saves you. A single-turn chatbot with no repeated context barely notices the cut, because it rarely hits the cache at all. An agentic coding tool that reloads the same repo context, tool definitions, and instructions dozens of times per session is exactly the shape of workload Anthropic priced this cut for, and that's where the 45%-range savings Anthropic cites would land. If you're already price-comparing Claude against GPT, Gemini, or open-weight models for a project, cache pricing — not just the sticker input/output rate — is now a line item worth actually checking on each platform's pricing page rather than assuming it's negligible.
If you're building anything agentic — a coding assistant, a multi-step research tool — check whether the platform you're using actually supports prompt caching and at what rate. The headline input/output price is only half the real cost story once a workload starts reusing context across calls.
- Sept 1, 2026: Anthropic launched Claude Fable 5.1 (public) and Claude Mythos 5.1 (restricted access) — the same base model with different safeguard levels.
- Base API pricing is unchanged: $10/million input tokens, $50/million output tokens.
- Cache-read pricing drops 75%, from $1.00 to $0.25 per million tokens; cache-write pricing is unchanged at $12.50 (5-min) and $20 (1-hr) per million tokens.
- Anthropic estimates roughly 25% lower effective cost for typical workloads and up to roughly 45% lower for heavily agentic ones.
- Mythos 5.1 access is currently limited to the (US-only) Life Sciences Verification Program; Cyber Verification Program access to Mythos-class models is planned but not yet live.
- Fable 5.1 shows large benchmark jumps over Fable 5, including Terminal-Bench 4.0 rising from 42.0% to 55.8%.
- Enterprise Frontier Safeguards, also announced Sept 1, is a free architecture letting enterprises hold Claude usage-monitoring data in their own encrypted cloud storage, rolling out in phases starting this fall.
01What's the actual difference between Claude Fable 5.1 and Mythos 5.1?
They're the same underlying model with different safeguard configurations. Fable 5.1 is generally available to any API user. Mythos 5.1 has additional capability unlocked but is currently restricted to the Life Sciences Verification Program (US-only); Anthropic says Cyber Verification Program access to Mythos-class models is planned but not yet available.
02Does Fable 5.1 cost more or less to run than Fable 5?
The base per-token rate is identical — $10/million input, $50/million output. The change is cache-read pricing, cut 75% from $1.00 to $0.25 per million tokens, which Anthropic says lowers effective cost by roughly 25-45% depending on how much of your workload relies on cached context.
03What is Enterprise Frontier Safeguards (EFS)?
A free security architecture Anthropic is rolling out starting this fall that lets enterprise customers store Claude usage-monitoring data in their own AWS S3, Azure Blob Storage, or Google Cloud Storage account, encrypted with their own keys, so misuse detection can run without Anthropic retaining a durable copy of the data itself.
04Can regular developers access Claude Mythos 5.1?
Not currently. It's currently gated behind the Life Sciences Verification Program (US-only); access via the Cyber Verification Program for cybersecurity-focused organizations is planned but not yet live for Mythos-class models. Fable 5.1 is the version available to everyone else.
05Why should a student or indie developer care about cache pricing specifically?
Because agentic workloads — coding assistants, research agents, anything that reloads the same context repeatedly across a session — spend far more tokens on cache reads than on fresh input. A platform's headline per-token price can hide most of the real cost if you don't also check its caching rate.
Comparing cache rates and benchmark deltas across a single provider's release notes is one thing — doing it across Claude, GPT, Gemini, and DeepSeek at once, before you've committed a subscription to any of them, is the harder problem most students and developers actually face. That's the gap LumiChats is built around: one login, every major model including Claude's latest lineup, and pay-per-day pricing under $1/day, so switching between them to see which one actually earns its keep on your workload doesn't cost a monthly commitment either way.
