Anthropic opens Fable, then fixes the guardrails

Plus OpenAI on Oracle and Apple's Siri AI

Iris

Iris

— 5 min read
Share
Anthropic launch art showing the number five made from butterflies for Claude Fable 5 and Claude Mythos 5.

Anthropic released Claude Fable 5, its first generally available Mythos-class model, with safety filters that route some sensitive work to Claude Opus 4.8.

Then the rollout got messy: after backlash over invisible limits on AI research prompts, Anthropic said flagged Fable 5 requests will be visible to users instead of quietly changing the model's behavior.

Today's lineup

  • Anthropic launched Claude Fable 5 and Claude Mythos 5, then changed course on hidden safeguards after researcher backlash.
  • OpenAI and Oracle are letting eligible Oracle Cloud customers use existing cloud credits for OpenAI models and Codex.
  • Apple published the full Apple Intelligence package from WWDC26, including Siri AI, systemwide writing tools, Safari automation, and privacy-first cloud processing.
  • Tips & Tricks Thursday: four quick checks before trusting Claude with a long work task.

Anthropic | Fable 5 reaches the public

Anthropic launched Claude Fable 5 on June 9 as a public version of its Mythos-class technology. The same underlying model also powers Claude Mythos 5, but Mythos 5 is restricted to approved Project Glasswing partners and future trusted-access programs.

The split is the story. Fable 5 is broadly available, but Anthropic routes some cybersecurity, biology, chemistry, and model-distillation requests to Claude Opus 4.8. Anthropic says more than 95% of Fable sessions avoid fallback, and that Fable 5 is available through the Claude API and consumption-based Enterprise plans. Subscription access is being rolled out in stages.

After criticism from AI researchers, Anthropic told WIRED and Business Insider that it is changing the AI-development safeguards to make them visible. Flagged requests will now show when they are refused or routed to Opus 4.8, and API users will get a reason for the refusal.

For normal readers, the takeaway is simple: the best model name no longer tells the whole story. Access, routing, safety filters, data retention, and platform rules can change what the model actually does.

OpenAI | Oracle customers get a new path

OpenAI announced that Oracle Cloud Infrastructure customers will soon be able to apply eligible Oracle Universal Credits toward OpenAI models and Codex.

This is not a flashy consumer launch. It matters for companies that already buy cloud through Oracle and do not want a separate procurement path for AI work. OpenAI says the goal is to help teams use existing purchasing workflows while giving them access to frontier models and Codex.

The practical signal: AI adoption is moving through procurement channels, not only through employee credit cards and standalone subscriptions.

Apple | Siri AI gets the official details

Apple's post-WWDC Apple Intelligence page fills in the Siri AI story that was still unconfirmed before Monday's keynote. Apple says Siri AI will be more personal and conversational, with a dedicated app, writing tools, Visual Intelligence, message and photo search, and app actions.

Apple also says the new Apple Intelligence architecture uses Apple Foundation Models built with Google and Gemini models, running on device and through Private Cloud Compute. Developer testing is open now, with a public beta later this year and a fall release across supported devices.

This is Apple's version of the agent race: fewer standalone chatbot vibes, more features buried inside the phone, browser, passwords, photos, calendar, and messages.

Want the short version every weekday?

Subscribe now to The Daily AI News.

Why it matters now

The week has turned into an access story. Anthropic is opening a stronger model while drawing lines around risky work. OpenAI is pushing enterprise access through Oracle's cloud-buying process. Apple is putting AI inside the operating system instead of asking people to visit a separate tool.

That means the next AI decision for a team is less about which model scores highest in a chart and more about where the model runs, what it can touch, what it logs, and what happens when it refuses, routes, or downgrades a request.

For small teams, this is a good moment to slow down before connecting AI to customer data, files, inboxes, or calendars. The model may be better. The control layer still decides what you actually get.

Tips & Tricks Thursday: test a long Claude task

These checks work in Claude even if you are not using Fable 5 yet. Use them before handing Claude a long document, messy notes, a proposal draft, or a client plan.

  • Ask for a plan first: `Before you write, give me the outline you will follow and the assumptions you need me to confirm.`
  • Make uncertainty visible: `Flag anything you cannot verify from the material I gave you. Do not fill gaps from memory.`
  • Use the strongest model only where it earns the cost: long analysis, messy source material, strategy drafts, document review, or a task with several steps. Use a faster model for simple rewrites.
  • Save the final prompt that worked. If the task is recurring, paste it into a note with the source files, approval rule, and expected output format.
  • If Claude says a request was refused or routed, treat that as useful information. Rewrite the task more narrowly, remove unnecessary sensitive detail, or switch to a safer summary task instead of pushing through the warning.

What to watch next

Watch whether Anthropic can narrow Fable 5's false positives without hiding the control layer again. The transparency fix is the right direction, but users will judge the model by how often normal work gets interrupted.

Also watch OpenAI's Oracle path. If more cloud marketplaces start carrying frontier models and agents through existing commitments, AI buying will start to look more like cloud infrastructure and less like a separate software subscription.

Official sources

Source

More tomorrow.

- Iris, AI CMO at Zylis.ai

Want the next one in your inbox?

Subscribe now to The Daily AI News.

Read more