30 April 2026: OpenAI Explains the Goblin Glitch, Anthropic Eyes $900B (AM)

The AI industry is explaining itself this morning in two very different ways: OpenAI has published a candid post-mortem on the personality quirks plaguing GPT-5, while Anthropic is reportedly circling a fundraise that would value it just shy of a trillion dollars. Closer to home, Microsoft says 20 million people now pay for Copilot, and Google Photos is teasing a feature that turns your camera roll into a digital wardrobe.

OpenAI has published an unusually frank explanation of why GPT-5 sometimes behaves like a goblin. In a post titled “Where the goblins came from,” the company traces the timeline and root cause behind the personality-driven quirks that have spread through its newest model since launch, including the odd lapses into mischievous, off-brand replies that users have been screenshotting and sharing for weeks. The piece walks through how the behaviour emerged during training and what fixes are now rolling out.

The notable thing is not the bug itself but the disclosure. AI labs rarely publish this kind of forensic detail, and the post sets a higher bar for what users and regulators can expect when models misbehave. For UK businesses building on top of GPT-5, it is a useful reminder that “stable” foundation models can shift in personality without the version number changing, which has direct consequences for tone-sensitive use cases like customer support and content moderation.

Anthropic is reportedly fielding offers that would value the Claude maker at up to 900 billion dollars. TechCrunch reports that the company has received multiple pre-emptive bids in the 850 to 900 billion dollar range, with talks underway for a new 50 billion dollar round. If it closes near the top of that range, Anthropic would be worth roughly 700 billion pounds, within striking distance of OpenAI’s most recent secondary marks.

The number is striking even by 2026 standards. It implies investors are pricing in not just continued enterprise adoption of Claude but also a meaningful share of the agent and coding tool market, where Anthropic has been gaining ground. The risk for buyers is concentration: at this valuation, Anthropic needs both Amazon Web Services and Google Cloud to keep underwriting compute, and any wobble in either relationship would be visible quickly.

Google Photos is rolling out an AI feature that turns your photo library into a virtual wardrobe. The tool, inspired by the rotating closet in the 1995 film Clueless, scans the clothes in your existing pictures and builds a browsable digital copy of your real wardrobe. Google says users will be able to mix and match items, generate outfit suggestions, and visualise combinations they have not actually worn together.

For UK households this is the most consumer-facing AI feature of the morning, and it lands in a market where Google Photos is already the default cloud library for many Android users. The practical question is whether the feature triggers any new data handling under the UK GDPR, given that wardrobe images and outfit metadata can be quite revealing about lifestyle and income. Google has not yet published a UK-specific rollout date, but anyone uneasy with the analysis can opt out at the album level.

Microsoft says paid Copilot users have passed 20 million and that engagement is growing. Speaking on the company’s earnings call, executives pushed back on the lingering perception that Copilot is widely sold but rarely used, claiming both the user base and the time spent inside the tool are climbing month on month. The 20 million figure covers paid seats across consumer Copilot Pro and the Microsoft 365 Copilot add-on for businesses.

For small UK firms still deciding whether the 24.70 pound per seat per month price tag is worth it, the engagement claim matters more than the headline number. Independent surveys earlier this year suggested that around a third of paid Copilot seats were being used less than once a week, and Microsoft’s case is now that those numbers have shifted. If that holds up in audit, Copilot starts to look less like a hopeful upsell and more like default productivity software.

Parallel Web Systems has hit a 2 billion dollar valuation just five months after its last fundraise. The AI agent infrastructure startup, founded by former Twitter chief executive Parag Agrawal, has raised another 100 million dollars led by Sequoia, doubling its valuation in less than half a year. Parallel builds tools that let AI agents browse, reason about and act on the open web, which has become one of the most contested layers of the AI stack.

For developers and product teams in the UK, Parallel is worth tracking because it competes directly with Browserbase and OpenAI’s own browsing tools to become the default agent runtime. The fundraise also confirms that investors are still willing to write large cheques for AI agent tooling even as broader software multiples have softened.

Here is everything else worth knowing from this morning’s AI news.

• OpenAI scales Stargate compute : OpenAI is adding new data centre capacity for its Stargate buildout, signalling more compute headroom for upcoming model upgrades.
• Google Cloud crosses 20 billion dollars in quarterly revenue : AI demand drove Google Cloud past the 20 billion dollar mark for the first time, though the company says capacity constraints held it back.
• Amazon AWS surges as capital spending climbs : AWS revenue is accelerating on AI demand and Amazon plans to keep spending heavily on infrastructure to keep up.
• Satya Nadella on the new OpenAI deal : Microsoft now gets to resell OpenAI’s models to its cloud customers without paying, and the chief executive says it intends to “exploit it”.
• Google adds 25 million paid subscriptions in Q1 : YouTube and Google One drove 25 million net new paid subs, taking the total to 350 million.
• Elon Musk back on the stand against OpenAI : Musk testified for a second day in his ongoing legal effort to unwind OpenAI’s structure, with his own past tweets surfacing as evidence.
• OpenAI publishes a cybersecurity action plan : A five-part plan focused on democratising AI-powered cyber defence and protecting critical systems.
• Ramp’s spreadsheet AI flagged for data exfiltration risk : Security researchers at PromptArmor say Ramp’s Sheets AI can be tricked into leaking financial data through prompt injection.

Featured image: Photo by Jonathan Kemper on Unsplash

This is a daily news update for informational purposes only. AI products and policies change rapidly. Verify details directly with providers before making decisions. Nothing here is financial or legal advice.

AI Daily is Cristoniq’s daily guide to developments in artificial intelligence, published every morning.