OpenAI just killed us. One problem, they didn't actually solve the biggest problem of AI at work: Context Let me prove it to you... Here's a wild example of ChatGPT Work going head-to-head with PromptQL:

Holy f***king shit. Grok 4.5 is a monster. On our internal evals on maintaining a company wiki: Grok 4.5 is fable-level on reasoning and a little lower than opus 4.8 on rigor. @elonmusk and @cursor_ai have just been COOKING. Just added Grok-4.5 on PromptQL for everyone.

Fable 5 is a BEAST! We gave 4 frontier models the same real job: clean up a bloated page in the wiki our team + our AI use as a shared skills repo. Fable was the only one that worked by a huge gap, with a generalizable intuition. ~40 sections of accumulated notes, duplicated & stale context, broken links. THE SETUP Our product (PromptQL) gives a team a shared AI agent plus a human-readable wiki the agent maintains as a shared skills and shared knowledge repo. This particular page had bloated into ~40 sections, had the same guidance repeated in 4 places, had dated status notes that were stale for a few months, and section titles that were not named according to the principle the agent is usually supposed to follow while maintaining wiki pages. Think of it as a dark corner of our wiki that because of a combination of human & AI actions become a particular weakness for either agent or human to maintain. We asked each model to restructure it with identical instructions. WHY THIS IS A GOOD MODEL TEST Maintenance punishes eloquence and rewards diligence. The only useful output is the fully rewritten page. The task also had hidden constraints (the page format supports flat sections only, no nesting) that are typical of knowledge work. We did this work on a duplicate copy of the page intentionally, while the original page was also kept around. WHAT FABLE 5 DID (10/10 on our rubric) • Read all ~40 sections before judging. The only model that did. • Found the actual redundancy: one piece of guidance repeated 3x, another 4x. Collapsed each to a single home. • Rewrote 40 incident-named sections into 9 skill-shaped ones ("how to do X" instead of "the June 12 issue"). • Added a routing map at the top so the reader can jump straight to the right section • Deleted stale point-in-time facts and kept the durable ones. • Caught the broken link. Every other model missed it. • Noticed the page looked like a duplicate, flagged it — and then did the assigned job anyway. • Delivered the complete rewritten page in one response. No outline. No "here's what I would do." WHAT THE OTHER THREE DID • GLM 5.2 wrote an articulate review memo about the page instead of restructuring it, but only after reading ~15 of the 40 sections. It shipped the diagnosis as the deliverable with a headline recommendation to delete the page ignoring the fact that we'd told it about the deliberate duplicate of the page. However, it produced the single sharpest observation of the batch about keyword placement. • Opus 4.8 preserved every word verbatim from the original page and wrapped the mess in nested sub-headings that was invalid formatting. It also needed a second prompt before it would output the full page. • Sonnet 5 reasoned beautifully about what the page WAS but applied every correct principle at whole-page granularity to justify discarding the content, and never shipped a complete rewrite. THE INTUITION TO TAKE AWAY Consider a task that is: read everything → judge fact by fact (rigorously) → rewrite completely — long, unglamorous work where the model has to respect the constraints of the system it's operating in and spend real output tokens finishing the job. For such tasks, Fable 5 is currently in a class of its own. Add that to your model routing classifier and thank me later.