JULY 19, 2026 — THREE STORIES THAT CHANGE HOW YOU USE AI CODING TOOLS TOMORROW
The Sunday the community had been watching delivered a pricing change, a security incident, and an acquisition — all in the same 24 hours.
- Fable 5 Credits Begin Monday — No Opus 5, No Fourth Extension — Free access ends tonight at 11:59:59 PM PT. From July 20: $10/$50/M. Batch API: $5/$25. Cache: $1/M input. Anthropic: will restore to subscriptions when compute allows, no date. Route general work to Sonnet 5 ($2/$10/M intro) from tomorrow. Full story + action plan →
- Grok Build Was Uploading Your Repos and SSH Keys Without Permission — Cereblab research: Grok Build uploaded entire repos, Git histories, SSH keys, and password manager databases to a GCS bucket, even when told not to open files. Musk vowed data purge. No incident report, no user notifications, no confirmed patch. Rotate SSH keys and passwords if you used Grok Build. Full story + immediate actions →
- OpenAI Acquires Ona (formerly Gitpod) — Codex Gets Persistent Cloud Agents — Codex at 5M weekly users. Ona acquisition enables Codex tasks to run continuously in cloud environments after you close your laptop. Terms undisclosed. Third major coding tool consolidation in July after SpaceX-Cursor and Anthropic-Ode. Full story →
Story 1 — Fable 5 Credits Begin: What Changes Tomorrow Morning
The extension cycle is over. This is the third time Anthropic has pushed back the original July 7 cutoff date — first to July 12, then to July 19. After July 19, Fable 5 moves from included plan access to metered usage credits. Starting July 20, all Fable 5 usage runs on prepaid usage credits at $10 per million input tokens and $50 per million output tokens — twice Opus 4.8. Anthropic has said it aims to restore the model to subscriptions once capacity allows. No date has been given.
The practical Monday morning change: default to Claude Sonnet 5 ($2/$10/M intro through August 31) for all standard workloads. Use Fable 5 credits only for tasks that specifically require its 80.4% SWE-bench Pro accuracy or 1M context with extended output. Before spending credits: set up prompt caching (cache hits drop to $1/M input) and route non-realtime work through Batch API ($5/$25/M). A heavy 50K-input/15K-output run is about $1.25, so roughly 30 runs a day amounts to $37.50/day or over $1,100 per month. The Honeycomb/Opus 5 signal has not materialised yet — it remains possible before July 31. Full credit cost breakdown and action plan →
Story 2 — Grok Build Data Leak: If You Ran It, Act Now
AI safety researcher Cereblab found that Grok Build, xAI's command-line interface, was uploading entire user repositories including full Git histories to a Google Cloud Storage bucket, even when instructed not to open files. One user had their entire home directory, including SSH keys and password manager databases, uploaded. Musk vowed a data purge. SpaceXAI has not published an incident report, has not confirmed whether the data was accessed by anyone, has not confirmed how many users were affected, and has not confirmed whether the behaviour has been patched in current versions. A vow to purge data is not a security remediation — it does not answer the questions that matter for risk assessment.
If you ran Grok Build on any machine with SSH keys, password manager databases, or code repositories containing sensitive data: rotate your SSH keys now, change passwords for any accounts in password manager databases on affected machines, and audit your Git history for hardcoded credentials. Do not resume Grok Build use on sensitive codebases until SpaceXAI publishes a full incident report with patch confirmation. Full incident details and immediate action steps →
Story 3 — OpenAI + Ona: Codex Gets Persistent and the Coding Agent War Escalates
OpenAI acquired German startup Ona, formerly known as Gitpod, to enhance its Codex coding tool by enabling persistent cloud-based agent operation. The acquisition aims to support Codex's growing user base, which now exceeds five million weekly users, amid fierce competition in the AI coding market. Persistent cloud agents solve the most practical limitation of current coding agents: tasks stop when your session ends. With Ona's cloud environment infrastructure, Codex can run a large refactor, test suite generation, or documentation pass overnight without requiring a local machine to stay active.
This is the third major consolidation in AI coding tools in July 2026: SpaceX acquired Cursor for $60 billion, Anthropic launched Ode (forward-deployed AI engineers in enterprise), and now OpenAI acquired Ona to enable persistent Codex agents. The competitive implication: Claude Code already supports persistent sessions; the Ona acquisition means OpenAI is investing to close that gap rather than accepting it as a structural disadvantage. Full analysis of what the Ona acquisition means →
The Week Ahead — Five Things to Watch
DeepSeek API migration — 5 days left (July 24 deadline). deepseek-chat and deepseek-reasoner stop working July 24 at 15:59 UTC. Migrate now: deepseek-chat to deepseek-v4-pro; deepseek-reasoner to deepseek-v4-flash (call v4-pro for heavy reasoning). This is a hard deadline — no grace period confirmed.
Opus 5 / Honeycomb — still possible before July 31. The Honeycomb EAP in Cursor on July 8 has not been acknowledged or denied. Credits beginning does not rule out an Opus 5 announcement this week. Watch anthropic.com/news and @claudeai on X. An Opus 5 with better token efficiency than Fable 5 would directly change the credit cost calculus.
Kimi K3 weights — July 27. Moonshot AI confirmed open weights by July 27. When weights ship, self-hosting K3 becomes possible — eliminating the China National Intelligence Law data residency concern for regulated industry users. The self-hosted K3 at $0 inference cost (your own GPU) vs $3/$15/M hosted changes the economics entirely.
Grok Build security response — watch for SpaceXAI incident report. If SpaceXAI does not publish a formal incident report within 48-72 hours, that absence is itself a signal. Regulated industry users (finance, healthcare, defence) may have mandatory data breach notification obligations that trigger regardless of SpaceXAI's response.
Claude Sonnet 5 intro pricing — ends August 31. The most underrated deadline of the month. Sonnet 5 at $2/$10/M is the best-value model available right now. After August 31 it steps up to $3/$15/M — matching GPT-5.6 Terra. Set up your Sonnet 5 production workflows now while it is the cheapest option.