🚨Anthropic are finished
You now have to spend $100 a month to use Fable 5 starting July 20th, unless you want to use usage credits (basically pay as you go, keep in mind Fable is very expensive)😂
The pro plan is pointless now, I don't understand why anyone wouldn't switch to ChatGPT or Grok.
Hi Ayush!
This is unrelated to the post but maybe here is the fastest route to reach out to you.
I have one feature request that I think matters for where the market is going: granular API key permissions.
Right now a key is either full-access or read-only. That’s a good start, but it’s binary, and my real setup needs something in between. Concrete example from my stack this week:
•My AI agent monitors sales, verifies subscription configs, and reconciles attribution. Read-only is perfect for that.
•The same automation needed to register a webhook endpoint. That’s a write, but a harmless one. To do it, I had to hand over a full-access key that can also refund every customer and cancel every subscription. That’s the exposure I want to close.
•My app backend needs to create checkout sessions and cancel subscriptions when a member clicks cancel. It should be able to do exactly that, and never touch refunds or payouts.
What I’d love to see: per-resource, per-action permissions on API keys, Stripe-restricted-keys style. Something like:
•payments: read
•refunds: none (or read)
•subscriptions: read + write
•webhooks: read + write
•customers: read
•payouts: none
This matters double in the AI-agent era. More and more of your customers will be wiring agents into the API, and least-privilege keys are the only real safety model for that. Even your own MCP server’s docs point users to “API keys with limited permissions” as the proper security layer. Granular keys would complete that story.
Two small bonus items while I have you:
1.Per-key audit visibility (last used, what it called) would make key hygiene much easier.
2.Your MCP server’s code-execution tool doesn’t work on Windows. It talks to Deno over a Unix domain socket, and Deno’s unix transport isn’t supported on Windows, so every execute call dies. A TCP-on-localhost fallback would fix it.
2 days in, already at 66% weekly on Max x20.
Tried splitting work between Opus (execution) and Fable (brain) to save tokens. Didn’t matter. Usage still went through the roof.
At this rate the $200 Codex plan is starting to look reasonable. I need advice.
@ClaudeDevs@claudeai@AnthropicAI@thsottiaux@sama