Two weeks after the US government pulled Claude Fable 5 offline over national security concerns, the evidence that it is returning is piling up fast. A new Claude Code update added a string of text reading "You've used your Fable 5 usage for this week," suggesting the model is being prepared to return as a weekly allotment included in subscriptions rather than a separate purchase. Separately, Fable 5 reappeared in Amazon Bedrock's model catalog, listed as active with available status when data sharing is enabled.

Neither of these confirms Fable 5 is fully back online yet. No successful API calls have been confirmed and the status page has not cleared. But model strings do not get added to production code for fun, and reappearing in AWS Bedrock's documentation is not an accident. The pattern looks a lot like the quiet staging that happens in the days before a relaunch.

The community reaction has been predictably enthusiastic, with "we are so back" trending in developer circles within minutes of the news spreading. The more measured read is that something is clearly happening, the weekly usage framing suggests Anthropic worked out some kind of compromise around access controls, and an official announcement is likely close. Whether it comes with restrictions for non-US users remains the open question most people are watching.

Google launched computer use as a native capability inside Gemini 3.5 Flash today, and the framing is worth paying attention to. Until now, computer use was a separate standalone model you had to specifically call. It is now just something Gemini 3.5 Flash can do by default, the same way it can search the web or use Maps.

What computer use actually means in practice is that you can give Gemini a screen and a goal and it figures out what to click, type, and navigate to get there, across browsers, mobile devices, and desktop environments. One Google engineer used it to audit their own documentation by telling it to visit a URL, run the code examples, and report back any issues. It navigated through the pages, took screenshots, and returned a structured report without any further input.

Google has also built in safeguards specifically for this capability, including the ability to require user confirmation before sensitive actions and an automatic stop if the model detects it is being manipulated by something on the page it is visiting. Computer use opens up a lot of power and a lot of potential for things to go wrong, and Google is being unusually transparent about the risks and how they are trying to address them.

Google has officially pushed Gemini 3.5 Pro to July. The current version is not where DeepMind wants it to be, with testers flagging issues around knowledge cutoff, token efficiency, and performance on long complex tasks. The model's training data still appears to stop before major late 2025 events, which is a real gap for a flagship model launching in mid-2026.

The extra time is being used to gather feedback from testers on Antigravity and LMArena and refine accordingly. New pretrain checkpoints are already surfacing internally, which suggests active work rather than a model sitting in a queue. The delay also raises the possibility that Google bundles 3.5 Pro with other announcements in July, including what some are calling a Gemini super app. Sometimes a delay is just a delay. Sometimes it becomes a bigger launch.