Today I am announcing that Fable 5.1 now powers the Shofield AI platform. We have integrated it into our governed model layer for selected autonomous workflows where long context, sustained execution and model economics matter. Suitable work can now use Fable 5.1 while Shofield AI keeps the workflow, tools, permissions, evaluations, data path, cost controls and human authority inside one operating model.
Not every Shofield AI workload will run on Fable. Our model gateway routes work by capability, risk, latency, geography and cost. A demanding research or operational step may use Fable 5.1; routine extraction may belong on a faster, lower-cost or privately deployed model. The client outcome—not loyalty to one provider—decides the route.
Anthropic released Fable 5.1 beside Mythos 5.1, two access tracks for the same model. Fable is broadly available for safeguarded enterprise work. Mythos is restricted to vetted US organisations: Anthropic's developer documentation currently describes direct API access through Project Glasswing, while its Mythos page separately describes invite-only life-science access. One makes sustained autonomy more commercially viable; the other shows why consequential capability requires narrower authority and faster revocation.
What Fable 5.1 changes inside Shofield AI
Fable 5.1 can now support selected long-running work across Shofield AI: complex research, document and case operations, service resolution, sales preparation, software maintenance and other workflows that benefit from sustained reasoning over reusable context. It is not switched on indiscriminately. Each route still has to earn its place through workflow-specific evaluation, acceptable cost, an approved data posture and safe failure behaviour.
Where usage is billed by token, Fable 5.1 keeps Fable 5’s list price for input and output tokens, but Anthropic cut cache-read pricing from $1 to $0.25 per million tokens. Cache reads matter when an agent repeatedly works from material it has already processed. Think of an engineering agent returning to the same repository, a case worker checking the same policy corpus, or a revenue workflow working through a qualified account record over several turns.
Anthropic estimates about 25% lower total cost for typical workloads and up to 45% for highly agentic, context-heavy work. Those are vendor estimates, not a business case. A lower cache-read rate does not automatically mean a lower cost per accepted task. Output tokens, retries, tools, latency and human review still matter. The practical question is not ‘what is the token price?’ It is ‘what did a verified completed task cost us?’
Fable 5.1 posts the highest reported CursorBench 3.2 score
Cursor independently evaluates ambiguous, multi-file coding tasks drawn from real Cursor sessions. These are max-effort results measured in Cursor’s production agent harness.
- Claude Fable 5.1 Max73.4% · $9.64/task
- Claude Fable 5 Max70.5% · $17.32/task
- Claude Opus 5 Max70.0% · $8.23/task
- GPT-5.6 Sol Max67.2% · $5.69/task
Operator takeawayFable 5.1 has the highest reported max-effort score at 73.4% and costs materially less per measured task than Fable 5. The score gap over Fable 5 and Opus 5 is small, so the responsible conclusion is ‘evaluate it’—not ‘it will win every workflow.’
Two access tracks raise the operational bar
Fable 5.1 is built for long sessions across research, documents, computer use and large codebases. AWS describes work spanning applications and running for hours. Mythos is neither a generally available Shofield AI option nor a normal enterprise tier. Anthropic currently limits it to vetted US organisations and documents separate trusted-access paths for cyberdefense and life-science work.
Fable can identify software vulnerabilities, but its deployed safeguards prevent penetration testing, exploit generation and binary-based scanning. Mythos gives approved specialists more permissive safeguards for defensive research; it is not a normal enterprise entitlement. When any AI worker can call tools or touch records, it needs a named owner, least-privilege identity, approvals, monitoring, a kill switch and rollback.
For public-sector security or regulated research teams, restricted access should change procurement. Start with an approved defensive purpose, named specialists, scoped systems, evidence handling, legal authority, auditability and rapid revocation. The stronger the capability, the more specific the operating permission must be.
That is why Shofield starts with the workflow and its decision rights, not a provider logo. The model may supply reasoning. The operating system must supply the authority boundary. More capable autonomous work is a reason to strengthen that boundary, not relax it.
Same model capability. Different operating authority.
Anthropic says Fable 5.1 and Mythos 5.1 share the same underlying model capabilities. Access, safeguards and permitted operating context are what separate them.
| Operating dimension | Fable 5.1Broad enterprise access | Mythos 5.1Restricted trusted access |
|---|---|---|
| Core capability | Fable 5.1 capabilities and specifications. | The same underlying model and capabilities as Fable 5.1. |
| Access | Available to all Claude API customers and through listed partner platforms. | Restricted to vetted US organisations. Direct API documentation currently describes Project Glasswing-only access; Anthropic separately describes invite-only life-science access. |
| Safeguard posture | Production safeguards for broadly available enterprise work. | More permissive safeguards for vetted defensive cyber and life-science work. |
| Shofield AI status | Enabled only for selected workflows that pass routing, evaluation, data and cost gates. | Not enabled. Shofield AI does not claim access or offer Mythos to clients. |
| Operating controls | Named owner, least privilege, approvals, monitoring, pause and rollback. | The same baseline plus an approved specialist purpose, scoped systems, evidence handling, legal authority and rapid revocation. |
| Documented retention | 30 days; zero-data retention only if Anthropic expressly authorises it. | 30 days; zero-data retention only if Anthropic expressly authorises it. |
Governance takeawayMythos is not a higher subscription tier. It is a restricted authority model for approved specialist work. Capability may be shared; permission must not be assumed.
A model migration is a production event
There is a useful warning in the release notes. Neither Fable nor Mythos 5.1 supports forced tool use, and older models cannot read Fable thinking blocks. Fable also checks whether earlier turns, tools or system instructions have changed and can invalidate those blocks; Mythos does not run that particular check. A router that silently swaps models mid-conversation can lose relevant reasoning context. An integration built around a forced tool call can return an error instead of doing the job.
Agent reliability comes from the model and its runtime. Before migrating, test malformed tool outputs, denied permissions, unavailable services, fallbacks, stale retrieval, revoked credentials and mandatory human approvals. Measure quality, cost and safe failure—not whether the demo looked impressive.
The best result may be a mixed estate: Fable 5.1 for the difficult, long-running step; a faster or lower-cost model for routine classification; and a specific fallback when policy or availability requires it. That is precisely why a model gateway and versioned evaluations are operational capabilities, not architecture diagrams.
Privacy and sovereignty are deployment decisions, not release-note checkboxes
The most important caveat is often missed in the excitement about capability and price. Anthropic’s documentation says both Fable and Mythos 5.1 carry 30-day data retention. Fable is not available under zero-data retention unless Anthropic expressly authorises it; Enterprise Frontier Safeguards is an opt-in, phased path for eligible Fable customers to use customer-controlled storage and customer-side review. That is progress, not a blanket privacy guarantee today or an automatic Mythos entitlement.
For Europe, the UAE, Africa and regulated US sectors, ask exactly which account, provider path, region, retention setting, contract and controls apply. Data location does not prove sovereignty, and a general announcement does not prove your tenant’s data path. If the answer cannot be documented, the workflow is not ready for sensitive production.
The same named model can be reached through a direct API, cloud marketplace or enterprise path with different controls. Procurement needs an architecture record: account, provider, safeguards, storage boundary, version, retention commitment and exit conditions. Without it, saying ‘we use Fable 5.1’ says almost nothing about risk.
How to start using Fable 5.1 through Shofield AI
Start with one workflow where sustained reasoning could create a measurable operating gain: sales preparation, case analysis, service resolution, contract review, software maintenance or regulated document operations. Give it one accountable owner, define the decision that remains human and establish the baseline: accepted-task quality, completion time, human intervention, tool failures and full cost—not just tokens.
Shofield AI can then map the workflow, choose the model route and connect the intended tools under explicit permissions. We run Fable 5.1 against the same evaluation set, with the intended policies, retention posture and fallback routes, and deliberately break the path: block an integration, change a permission, trigger a refusal and switch models. A route moves forward only when it improves the outcome without weakening authority, privacy or recoverability.
How Fable 5.1 is governed inside Shofield AI
Fable powers selected demanding steps inside a controlled workflow. It does not receive every task, every tool or unlimited authority.
Identity · permissions · approvals · evaluations · audit · budgets · monitoring · kill switch · rollback
- 01Qualified workflow
One revenue or operational outcome with an owner, baseline and stop conditions.
→ - 02Policy & identity gate
Purpose, data boundary, permissions, region and human decision rights are fixed before routing.
→ - 03Model Gateway
Fable 5.1 handles suitable long-context work; approved alternatives handle routine, private, regional or fallback steps.
→ - 04Permissioned execution
Only approved tools, records and actions are available for the mission and timeframe.
→ - 05Evidence & human decision
Quality, full cost, intervention, safe failure and audit evidence determine approve, pause, revoke or rollback.
Architecture takeawayThe model supplies capability. Shofield AI supplies the route, evidence and authority boundary required to turn that capability into a controlled operating result.
You do not need to replace your full technology estate or commit every AI workload to one provider. Begin with a bounded 30-day implementation, measure the result and expand only after the evidence is strong. That makes adopting Fable 5.1 through Shofield AI a controlled operating decision rather than an open-ended transformation programme.
The relevant question is no longer whether a frontier model can work for hours. It is whether your organisation can prove what it was allowed to do during those hours.
Start with one governed workflow
Fable 5.1 is a real opportunity where AI work is constrained by long context, deep reasoning or repeated use of the same knowledge. Mythos 5.1 is the parallel governance lesson: when frontier capability reaches cyber and life-science boundaries, access and oversight must rise with it. Neither model removes the need to choose the right route, test the integration, govern data and retain human authority.
The practical question is no longer whether Fable 5.1 benchmarks well. It is which high-value workflow your organisation should put into controlled production first. Shofield AI is now ready to help you do that. Bring us one revenue or operational workflow; we will assess its fit, configure the model and control path, establish the success measures and show you what should—or should not—move into production.
Shofield AI independently uses Fable 5.1 within its multi-model platform. This announcement does not imply an Anthropic partnership or endorsement, and it does not imply that Shofield AI or its clients have access to Mythos 5.1.