AI News Today — July 19, 2026:
The 5 Biggest Stories You Need to Know
The Sunday the AI coding world had circled on its calendar. Anthropic’s Fable 5 free-access window closed tonight with no fourth extension and no Opus 5 in sight, a security researcher caught SpaceXAI’s Grok Build quietly shipping user credentials off-device, and OpenAI made its clearest coding-agent power move yet with an acquisition aimed squarely at Claude Code. Layered on top: Oracle’s workforce is absorbing the cost of the AI infrastructure boom, and Microsoft just handed Anthropic’s flagship security model a very well-funded rival.
Fable 5’s Free Ride Is Over — Usage Credits Start Monday at $10/$50 per Million Tokens
Anthropic let Claude Fable 5’s included-plan access lapse at 11:59 PM PT tonight after three separate extensions since early July, and this time there’s no fourth reprieve or Opus 5 announcement to soften the landing. From July 20, Fable 5 usage runs on prepaid credits priced at $10 per million input tokens and $50 per million output — double what Opus 4.8 costs and the priciest entry on Anthropic’s price list, though Batch API use halves it and cached input tokens drop to $1 per million. Anthropic says the credit-only period isn’t meant to be permanent and it will restore subscription access once compute capacity allows, but hasn’t given a date. For most everyday coding and writing work, Claude Sonnet 5 remains fully included in paid plans at its $2/$10 introductory per-million rate through the end of August.
Grok Build Was Quietly Uploading Users’ SSH Keys and Password Vaults
AI safety researcher Cereblab found that SpaceXAI’s coding agent Grok Build had been sending entire local repositories — full Git history, SSH keys, and even password-manager databases in some cases — to a Google Cloud Storage bucket, and doing it even after users had explicitly told the agent not to touch those files. Elon Musk responded on X promising a data purge, but SpaceXAI hasn’t said how long the behavior was live, whether the uploaded data was ever accessed, or which builds are affected. Anyone who has run Grok Build should treat their SSH keys and any locally stored credentials as compromised and rotate them now, since a promised purge isn’t the same thing as a confirmed fix.
OpenAI Buys Ona (Formerly Gitpod) to Give Codex Agents That Never Sleep
OpenAI acquired German cloud-development startup Ona — the company formerly known as Gitpod — specifically to let Codex run as a persistent, cloud-hosted agent instead of a process tied to an open terminal window. Close your laptop mid-task today and a local coding agent stops; Ona’s tech is meant to keep long-running jobs alive in the cloud until they finish or hit a wall that needs a human. The timing tracks Codex’s growth: OpenAI says the tool now clears 5 million weekly users less than three months after its public debut, and this is the third major coding-tool consolidation move of July, following SpaceX’s reported $60B purchase of Cursor and Anthropic’s backing of the Ode startup. Terms of the Ona deal weren’t disclosed.
Oracle Cuts Up to 30,000 Jobs — While Posting Record Cloud Revenue
Oracle began laying off an estimated 20,000 to 30,000 employees, roughly 18% of its workforce, concentrated in sales, engineering, and security — internal tracking reportedly showed 10,000 Slack accounts vanish almost overnight. The company says the savings are being redirected into AI data-center buildout, chiefly its share of the $500B Stargate infrastructure partnership with OpenAI and SoftBank. What makes this one notable is that Oracle isn’t cutting jobs to survive a downturn; its cloud division just posted record revenue. It’s the clearest example yet of a pattern showing up across the sector this month, where AI infrastructure spending is increasingly funded by headcount reductions at profitable companies rather than by losses.
Microsoft Is Building a Cut-Price Rival to Anthropic’s Mythos Security Model
Microsoft is preparing to launch “Project Perception,” an enterprise vulnerability-scanning tool that blends Claude, GPT-5.6, and Microsoft’s own models, positioned as an accessible alternative to Anthropic’s Mythos 5 — a model most companies currently can’t get access to at all. Mythos 5 is limited to roughly 40 vetted partners under Anthropic’s Project Glasswing program, a restriction that tightened further after June’s export-control disruption. Project Perception sidesteps that bottleneck entirely: it ships inside Microsoft’s existing Defender and Sentinel security products, needs no separate procurement process, and can deploy in regions like Europe where Mythos access remains effectively closed. It’s a win for Anthropic’s distribution, since Claude is one of the models doing the work, but also a real threat to Mythos’s pricing power in the broader enterprise security market.
That’s today’s wrap. The Fable 5 credit switch and the Grok Build leak are the two every developer should act on this week — one changes your bill, the other means rotating your keys. Follow NeuralPaws daily for the next AI news drop.