The Intake
The Intake — Wednesday, September 2, 2026
On the substrate
Fable 5.1 ships with 75% cheaper cache reads and doubled science benchmark scores
Anthropic VentureBeat MacRumors
If you run Claude Fable 5 in agentic workloads where prompt caching carries most of the load, today's release changes the cost line. Fable 5.1 shipped September 1, 2026. Cache-read pricing drops from $1.00 to $0.25 per million tokens — a 75% reduction. Anthropic estimates 25% savings for typical workloads. The estimate rises to 45% for highly agentic tasks. Input and output pricing are unchanged at $10 and $50 per million tokens.
The benchmark numbers also moved. On Terminal-Bench-Science 0.1, Fable 5.1 scores 52.6%. Fable 5 scored 24.7% on the same test. On Terminal-Bench 4.0, Fable 5.1 scores 55.8%. Fable 5 scored 42.0%.
Mythos 5.1, which shipped alongside Fable 5.1, carries strengthened cybersecurity and life sciences safeguards. Anthropic says it reduces false interventions by approximately 60%. The reduction is measured against Fable 5. Access is restricted to US organizations that have cleared Anthropic's Cyber Verification Program or Life Sciences Verification Program. It is the first model under Anthropic's Enterprise Frontier Safeguards framework. That framework adds customer-controlled data residency. If your workloads are cache-heavy and you're running Fable 5 today, the cache-read cost change is the number to update first.
OpenAI launches WebMCP — a proposed standard for AI agents to call structured tools directly on websites
OpenAI community announcement Superpower Daily VKTR
If you build agents that interact with websites by navigating HTML or simulating clicks, OpenAI's August 25 announcement names the alternative now being standardized. WebMCP lets websites expose named, typed JavaScript functions that AI agents can discover and invoke directly. The agent runs through the user's existing authenticated browser session. Higher-stakes actions — purchases, for instance — require explicit user confirmation before execution.
Adoption is already underway. Millions of Shopify storefronts are WebMCP-enabled, OpenAI says. Expedia, Instacart, and Target are among early participants, per OpenAI. Chrome 146 includes a developer trial. OpenAI is proposing the standard through the W3C Web Machine Learning Community Group. Cloudflare, Vercel, Netlify, and Render are named as challenge partners.
The standard is currently scoped to callable tools only — it does not include MCP's resources, prompts, or sampling features. In ChatGPT, it is available to Work and Codex users; Enterprise and Edu workspaces are excluded. If you're building agents that need to call web-based tools through a structured interface rather than simulating UI interactions, this is the spec to evaluate now.
---
For operators
Fable 5.1 cache-read cut and Mythos 5.1 verified access: what changes for your Claude cost model
Anthropic VentureBeat MacRumors
If you run Claude in production with prompt caching, the cache-read line in your cost model changed yesterday. Fable 5.1 cache reads are now $0.25 per million tokens, down from $1.00. Batch processing runs at $5 per million input and $25 per million output tokens. Base input and output rates are unchanged from Fable 5.
The Mythos 5.1 access question is separate. If your deployment is in cybersecurity or life sciences and you're not currently enrolled in Anthropic's Cyber Verification Program or Life Sciences Verification Program, access depends on clearing that gate first.
WebMCP challenge closes tomorrow — the implementation decision for web-facing agent deployments
OpenAI community announcement Superpower Daily VKTR
If you run web-facing applications and are weighing whether to add WebMCP support, the challenge window closes tomorrow (September 3, 1 p.m. PT). The implementation question is longer-lived. Adding WebMCP means exposing named callable functions alongside your existing application logic and permissions — agents call those directly rather than navigating your UI.
Progress Software says its Telerik and Kendo UI toolkits can register components as WebMCP tools automatically, without custom code. If your users are on ChatGPT Enterprise or Edu, they won't encounter WebMCP agent interactions yet; those workspace types are currently excluded from the rollout.
---