September 2: Claude Fable 5.1, enterprise safeguards, and two new ChatGPT workflows

The Aug. 31–Sept. 1 window produced four updates worth a practical test. Claude Fable 5.1 changes the model and cost equation; Anthropic’s enterprise safeguards change where sensitive-use controls can live; OpenAI’s other two releases turn ChatGPT into a more bounded tool for websites and public clinical information.

What shipped: Anthropic released Claude Fable 5.1 on Sept. 1, 2026, and made the same underlying model available as Claude Mythos 5.1 with a higher safeguard level. Fable 5.1 is generally available. Mythos 5.1 is reserved for trusted-access programs, including approved Project Glasswing customers. 1

Anthropic says Fable 5.1 costs about 25% less than Fable 5 for typical token-billed workloads because cache-read pricing is lower. Anthropic estimates savings of up to about 45% on highly agentic work. The platform migration guide lists input at $10 per million tokens, output at $50, and cache reads at $0.25. 2

Anthropic reports higher scores for Fable 5.1 on several internal evaluations, including 55.8% on Terminal-Bench 4.0 and 73.4% on CursorBench 3.2.0. Those are Anthropic’s reported results, so treat them as launch evidence to test rather than a neutral leaderboard. 1

Who can use it: Fable 5.1 is available through the Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud, and Microsoft Foundry. The default context window is 1 million tokens, and a request can return up to 128,000 output tokens. Adaptive thinking is always on. 2

Why it matters: Developers and AI startups can retest long-context agents with lower cache-read costs. The migration has a sharp implementation edge: forced tool choice is unsupported, so code using tool_choice values any or tool returns a 400 error. Anthropic’s guide points developers toward auto, stronger instructions, and strict schemas where available. 2

First test: Run one cached, tool-using workflow against Fable 5 and Fable 5.1. Measure cache reads, total output, tool-call reliability, and the cost of the complete task. Check the retention setting before moving production traffic: the migration guide describes 30-day retention as the default and ZDR access as requiring Anthropic authorization, while Anthropic’s separate enterprise announcement gives eligible customers a ZDR path during the EFS rollout. 23

What shipped: Anthropic announced Enterprise Frontier Safeguards, a planned way to combine zero-data-retention-style privacy with automated safeguards for serious misuse. The announcement is dated Sept. 1, 2026. Anthropic says the rollout will begin in phases later in fall 2026. 3

Who can use it: Eligible customers receive ZDR on Fable 5 and Fable 5.1 while EFS is being rolled out. The planned surfaces include Claude Code, Claude Enterprise, Claude Platform, Amazon Bedrock, Claude Platform on AWS, Google Agent Platform, and Microsoft Foundry. 3

The design puts customer data in cloud infrastructure controlled by the customer. A customer can use an existing Amazon S3, Azure Blob Storage, or Google Cloud Storage account, bring its own encryption keys, set access policies, and keep audit logs. Automated monitoring reviews rolling activity windows for serious misuse, while alerts go to the customer’s own reviewers. 3

Why it matters: Regulated teams may gain a clearer route to frontier models while keeping storage, keys, and human review inside their own environment. EFS itself carries no Anthropic charge, although the customer pays its cloud provider for storage, reads, writes, and egress. The plan leaves model behavior, API pricing, and rate limits unchanged. 3

First test: Treat EFS as an architecture decision, not a feature available today. Ask security and platform teams which logs must remain customer-controlled, who reviews an alert, and which cloud account will hold the data. Keep the answers beside the rollout plan until Anthropic makes the relevant surface available.

What shipped: OpenAI says ChatGPT Work and Codex can use tools that supported websites provide directly in the desktop app’s built-in browser. ChatGPT can discover the tools for a task, and the address-bar arrow shows which tools a page provides. The release notes identify WebMCP as the underlying mechanism and date the update Aug. 31, 2026. 4

Who can use it: The feature depends on a supported account, model, and webpage. The current release is in the built-in desktop browser. Existing website-access and sensitive-action confirmations still apply. 4

Why it matters: Developers can expose task-specific website actions instead of asking an agent to navigate every screen like a person. Marketers and operators can test workflows such as drafting a report, preparing a campaign object, or collecting structured information, while the confirmation step remains part of the path for sensitive actions.

First test: Build one reversible action on a staging page. Check whether ChatGPT discovers the tool, describes the action correctly, asks for confirmation at the right moment, and leaves a usable result. Keep destructive or financial actions out of the first trial.

What shipped: Eligible ChatGPT for Clinicians users in the United States can install Healthcare Public Data in ChatGPT from the Plugin directory and connect the sources they want. The plugin brings together nine read-only apps covering biomedical research, clinical trials, medication information, Medicare data, and provider records. 45

The apps stay read-only and outside patient charts. OpenAI warns users to leave protected health information out of searches sent to public sources. 5

Why it matters: Clinicians can search several public healthcare source types from one ChatGPT workflow. Healthcare builders can study a bounded retrieval experience with clear limits: public information, read-only connections, United States eligibility, and a privacy rule that still depends on the user’s query.

First test: Use a de-identified, public-information question such as a trial lookup or medication-source comparison. Verify which app answered, open the underlying source, and keep patient information out of the search.

The useful pattern across these releases is practical access with explicit boundaries. Fable 5.1 changes the economics of long-running model work; EFS addresses where enterprise controls may live; WebMCP gives websites a way to expose actions; and Healthcare Public Data keeps a clinical workflow on public, read-only material. The next workflow test should match the gate that currently slows your team down.