Fable 5.5 Release Rumours, OpenAI Spy Arrest and a Big Claude Code Update

Key takeaways

• Fable 5.5: reported DNS and nine-region Azure footprint, plus brief-access reports. Possibly out this week, but unconfirmed
• OpenAI arrest: an engineer is charged over an alleged hidden audio device in an executive meeting room. OpenAI says there is no sign model weights, code or customer data were taken
• Claude Code: redesigned desktop app, nested sessions, /btw side questions, saved environments and idle compaction to save usage
• Anthropic: reported Q3 revenue of $16.8bn, profitable on an adjusted basis, with an IPO talked about for November
• Also: stealth models Cat Jam and Wild Missing No, Grok Voice 3.0, an Nvidia model tease and Hood Warfare 2’s reported $1.1M in sales

This week’s AI news is packed. The headline question is whether the Fable 5.5 release is finally close. There is also a spy story at OpenAI that sounds made up but isn’t, and a Claude Code update that changes how you run several coding agents at once.

Last week we covered Gemini 4 Argon and the first Fable 5.5 catalogue sightings in our Gemini 4 Argon hands-on round-up. This article picks up the new evidence and the rest of the week’s stories. Anything unconfirmed is labelled clearly.

Is the Fable 5.5 release happening this week?

Short answer: maybe, but nobody official has said so. What has changed is the kind of evidence. Earlier sightings came from model catalogues. The new reports point to infrastructure.

Two more details push the speculation along:

• Brief-access reports. Some users say they were briefly routed to Fable 5.5 before losing access. That could point to final testing.
• IPO timing. Anthropic is widely expected to go public, possibly in November. The creator suggests a big model launch just before an IPO would make strategic sense. He also says Anthropic could hold Fable 5.5 back to launch closer to the listing.

The video also shows a long “civilisation from day one to now” film that someone reportedly made by giving Fable 5.5 a single prompt. Every frame was said to be generated through code, with no stock footage. It is an impressive clip, but we can’t check which model actually made it. Treat it as a teaser, not a benchmark.

What happened in the OpenAI employee arrest?

This is the strangest story of the week. OpenAI says its security team found a hidden audio recording device in one of its executive conference rooms in San Francisco. That room is used for some of the company’s most confidential talks, including research, product plans, business strategy and government work.

Here is what has been reported. All of it comes from allegations in a criminal complaint and statements from OpenAI and the US Justice Department:

• The device was allegedly plugged into a TV’s USB port, set to encrypt its recordings and wrapped in toilet paper to hide it behind the screen.
• OpenAI found it during a sweep of the room on 30 September. The company says it used its own AI model to help work out where the device came from.
• Prosecutors say the suspect, a 35-year-old engineer, spent weeks asking an AI assistant about internal meeting schedules. Between 24 and 30 September, he allegedly asked a chatbot how to build, hide and remotely access recording devices.
• Security footage reportedly showed him near the room several times in September. A camera allegedly recorded him looking behind the TV on 5 October.
• He was arrested by the FBI on 9 October and fired by OpenAI. If convicted, he faces up to five years in federal prison and a fine of up to $250,000. A detention hearing is set for 19 October.

OpenAI says there is currently no sign that model weights, source code or customer data were stolen. The case is ongoing, the charges are allegations, and the suspect has not been convicted. You can read the official summary in the US Justice Department press release.

There is a bigger point here. As AI models become more valuable, the research behind them becomes a target. Physical security matters as much as cyber security.

What did OpenAI ship on days 8 and 9 of its 28-day pledge?

OpenAI has promised to release something new every day as part of its 28-day shipping pledge. Two of the smaller releases are useful for heavy users:

• Day 8: a usage-limit reset, which is welcome news if you had already hit your limits.
• Day 9: according to OpenAI’s Tibo, changes to how the ChatGPT and Codex desktop apps store their state on your computer. This should cut disk space use by roughly 60%.

It isn’t all good news for OpenAI. The Wall Street Journal reports that its computing chief is stepping down, and other employees are reportedly moving to Anthropic. People change jobs all the time, and Anthropic went through its own difficult patch a few months ago before recovering. Still, it is a trend worth watching.

What is new in the Claude Code update?

For anyone running several coding agents at once, this is the most practical story of the week. Anthropic has redesigned the Claude Code desktop app to make parallel work easier to manage:

One downside the creator spotted: the model selector has gone from the new interface, and he would like it back.

Idle compaction: the feature that saves usage

Idle compaction may save more money than any of the interface changes. If you leave a long conversation idle for close to an hour, Claude Code now compresses it before its prompt cache expires. When you come back, it only has to process a short summary instead of the whole conversation again. Anthropic’s changelog says this applies to large terminal conversations, and you can switch it off in settings if you would rather keep the full context.

Anthropic has also launched OpenTelemetry tracing for the Claude Platform in public beta. It works across the Messages API, Claude Managed Agents and the Claude SDK, so teams can debug agents and spot slow steps in the observability tools they already use.

Want to set this up properly? Our guide to Claude Code idle compaction and nested sessions covers the settings, and when to split work into child sessions versus keeping it in one.

How big is Anthropic getting before its IPO?

Two sets of figures are doing the rounds. Both are reports, not audited accounts:

• Revenue. The Financial Times reports that Anthropic told investors its third-quarter revenue reached $16.8bn. It also said it was profitable on an adjusted basis and cash-flow positive. If those figures hold up, that is a big milestone for an industry spending heavily on compute.
• Compute. Reported estimates put Google’s entire TPU capacity at about 2.2 gigawatts at the end of 2025. Anthropic is reportedly planning over 1 GW in 2026, another 5 GW in 2027 and possibly another 10 GW in 2028. That would be roughly 16 GW by 2028, more than seven times that 2025 estimate. Broadcom’s CEO reportedly confirmed that the 10 GW figure is for Anthropic. These are plans, not hardware already delivered.

What are the stealth models Cat Jam and Wild Missing No?

Two anonymous “stealth” models are free to try right now:

What else happened: Grok Voice 3.0, Nvidia and Hood Warfare 2

• Grok Voice 3.0. xAI (now SpaceXSI) says this is the world’s number one speech-to-speech model, with replies in under a second, 40% lower latency and lower prices than rival frontier voice models. It reportedly already runs Starlink customer support, where it is said to resolve about 70% of calls without a human agent.
• Nvidia tease. Nvidia posted that “so many models are launching this week and we have our own coming”, which suggests a new model soon. A new Nemotron is the obvious guess, but it is a guess.
• Hermes phone harness. A new open-source plugin lets an AI agent see and control an iPhone (through Mac mirroring) or an Android phone (through ADB). It can tap, swipe, type and open apps without needing an API for each one.
• AI on a PS5. Someone ran a 27-billion-parameter Qwen model on a PlayStation 5’s GPU at around 20 tokens a second. On the PS5’s CPU, it managed about one token a minute.
• Epic Games. Epic is rebuilding its game launcher with AI to make it much faster, and a preview is already out. It is also reportedly testing AI-assisted game development.
• Hood Warfare 2. This game was reportedly built largely with AI-written code and assets from existing asset stores. Estimates suggest about $1.1M in revenue, more than 200,000 owners and around 1,500 reviews, all at a $5 price. It shows that small teams can now ship things people actually pay for.

What should small businesses do with this news?

Our honest take: most of this week’s headlines are interesting, but only a few need action from a small business right now.

If you choose AI tools for a team, re-test on your own tasks rather than chasing every launch. Our guide on how to choose an AI tool for your business shows you how to run a fair trial.

Frequently asked questions

The bottom line

The Fable 5.5 release looks closer than ever, but it is still a rumour until Anthropic announces it. The update you can use today is Claude Code’s: nested sessions keep parallel work tidy, and idle compaction can cut wasted usage on long sessions. The OpenAI case is a reminder that physical security still matters. Get your test set ready, update your tools and wait for official launches before you change anything important.

Related posts

• Gemini 4 Argon Hands-On: Front-End Speed, Vision and AI Rumours
• Claude Code Mods: How to Build Your Own Automation Shortcuts
• How to Choose an AI Tool for Your Business: Benchmarks vs Real-World Results

Garry Knight

I'm Garry Knight, the person behind Prodify Digital. I write about email list building, email marketing, SEO, AI search and the tools that connect them. My aim is to make online marketing easier to understand, so creators and small business owners can make informed decisions about building an audience and keeping people engaged. Here you'll find straightforward guides and product reviews that explain what something does, where it fits and which limitations matter. The focus is on clear explanations and useful next steps—not hype, shortcuts or promises of easy earnings.

See Full Bio