Fable 5 is back - with strings attached: Usage credits, ID checks and a shared industry rulebook - all in one week.

🔭 The permission layer just became permanent infrastructure

The pattern. Emergency shutdown → controlled restoration → vetted access. Two competing labs ran that exact sequence independently, in the same seven days — and neither lost customers or credibility for it.

What that means for you. A model you rely on can now be narrowed to a list you’re not on, instead of simply switched off. The question worth asking has quietly changed from “is this model good enough” to “am I on the list.”

The honest caveat. Two events isn’t proof of a permanent shift — it’s evidence of one. If both labs quietly return to open access once this dispute cools, this reads as an overreaction. Worth holding both possibilities.

The bet. By the end of Q3 2026, at least one more frontier lab adopts a tiered, vetted-access model voluntarily — the fastest route back to market after a security scare, not the slowest. The full case — the permission economy →

You are my AI vendor risk auditor. Here is a

tool or model I depend on: what it does, how

critical it is, and how often I use it.

Score it on the same four axes labs now use

to grade AI jailbreak severity:

1. Capability concentration — if this vendor

changed its terms tomorrow, what could I

no longer do that I can’t easily replace?

2. Breadth of dependency — how many of my

workflows actually touch this one vendor?

3. Ease of substitution — how many honest

hours to swap to a real alternative?

4. Discoverability of change — would I

notice a quiet policy change before it

hit me, or only after?

Score each 1–5. Anything scoring 4–5 on two

or more axes is this month’s audit target.

🛠️ TOOL OF THE WEEK

LiteLLM

Open-source proxy that speaks every major model’s API — self-hosted, so no third party sits between you and your fallback.

★★★★ / 5

Last week we recommended OpenRouter as insurance against single-model dependency. Fair question this week: what happens if the insurer becomes a vendor you depend on too? LiteLLM is the self-hosted answer — you run the routing layer, not a third party. It exposes one OpenAI-compatible endpoint in front of 100-plus providers, including Claude, GPT, Gemini and any open-weight model you host yourself, with built-in fallback, retry and spend tracking.

Not technical? The principle still holds: know which of your tools you actually control, and which you’re simply renting access to.

Use if: you’re comfortable running a lightweight proxy and want zero dependency on someone else’s routing layer. Skip if: OpenRouter’s managed convenience is worth paying for — that’s a legitimate trade, not a mistake. Describe it to a colleague: “It’s the same adapter — except we own the socket.” Best use case: put an open-weight model behind it as the fallback that never needs an export licence, whoever’s government writes the next letter. LiteLLM →

💡 TIP OF THE WEEK

The grace period is not the policy

Here’s what got buried under the good news: Fable 5 didn’t come back “restored” — it came back with a countdown. Free inclusion for Pro, Max and Team plans runs out on 7 July; after that, usage credits. Enterprise needed credits switched on manually, or the model simply doesn’t work for your team. Again.

Pin the actual expiry date, not the headline date. Every “we’re bringing it back” post has two dates: today’s, and the one the free window closes. Calendar the second one.

Check whether “on” means “on for you.” Enterprise access needed a setting switched on — not a default. A restored model nobody enabled is still an outage.

Why it works: the sentence that costs you money is never the headline — it’s clause three, footnote one. Reading past the good news is the whole skill. Where it doesn’t apply: genuine no-strings restorations exist. The tell is whether a specific date and an action item both show up in the fine print. If neither does, relax.

Forward iPrompt → For the colleague who assumed “it’s back” means “nothing changed.”