StepFun's 1M-token context enables complex long-form writing assistance

Continuity is the silent killer of long series — the eye color that flips in book three, the timeline that quietly snaps in book five, the reader review that catalogues every slip in devastating detail. A new model out of China just showed up with a context window big enough to hold your entire back catalogue in its head at once.

StepFun's Step 5 Preview is a sparse mixture-of-experts model — 600 billion parameters total, but only 27 billion active on any given task — built mainly for coding and "agentic" office work, not fiction. Nobody at StepFun is pitching this to novelists. But the number that matters travels well past its intended audience: a 1-million-token context window, roughly the equivalent of 1,500 pages in a single pass. A full-length novel runs 80,000 to 100,000 words, call it 110,000 to 135,000 tokens — so a window this size has room for seven or eight complete books, or one very long series bible, in one working conversation.

That matters more than raw intelligence scores for a specific kind of author problem. If you write trope-heavy series — romantasy, cozy mystery, anything with a returning cast — you've almost certainly hit the wall where your AI assistant forgets book one by the time you're asking about book four, forcing you to re-paste summaries every session just to get a straight answer on whether your hero's eyes were blue or green two books ago.

Step 5 itself isn't that tool yet. It's priced for developers ($1 per million input tokens, $2.70 output), aimed at software and agent work, with open weights due October 15. But context windows this size tend to trickle down fast, and once one lands inside the tools you already use — Sudowrite, Novelcrafter, or a future version of ChatGPT or Claude — the "wait, didn't we establish that already?" problem gets a lot smaller. Worth watching for the version of this that ships inside your actual workflow, not the one built for engineers.