StepFun's Step 5 Preview leaks at 600B — and the weights stay closed
StepFun's next flagship surfaced on a benchmark page before the company said a word about it, and it breaks the pattern the Shanghai lab built its reputation on. Tesla, meanwhile, is putting its Chinese in-car assistant behind a subscription.
StepFun's Step 5 Preview appeared on Artificial Analysis's model page dated September 18 — a 600-billion-parameter reasoning model with a 1M-token context window, text and image input, and a score of 44 on the Artificial Analysis Intelligence Index. The listing reports 27 billion parameters active per token, pricing of $1.00 per million input tokens and $2.70 per million output, and a cache-read price of $0.05 — roughly a quarter of the median output price for models in its tier. Reddit's LocalLLaMA spotted the page first, and a Chinese forum post claimed the launch venue was an AGI House "AGI Frontier" event in Shanghai; StepFun has not published a Step 5 page, and its own site still lists Step 3.7 Flash as the current flagship.
The number that matters is 44. That lands the model level with Kimi K3 and a single point under GLM-5.3, the highest-ranked open-weights model on the same leaderboard — except Step 5 Preview is listed as proprietary, with no weights available. StepFun's identity for two years has been open releases: Step 3.5 Flash and Step 3.7 Flash both shipped weights and a training recipe, and the lab's last launch was voice, not text — StepFun ships five voice models and takes two global No. 1 spots. A 600B flagship that closes is a different company posture, and it arrives while Chinese labs including Zhipu and Moonshot hold the open end of the frontier.
Two caveats before anyone reroutes production traffic. First, a preview on a third-party benchmark page is not a release: no announcement, no API documentation, no deprecation notice for what it replaces. Second, the model is expensive to run in a way the price sheet hides — Artificial Analysis reports it generated 160 million tokens to complete the index, against a median near 90 million, which means it thinks longer and bills for it. Cheap per token, costly per answer.
Tesla began pushing software version 2026.26.200.11 to Chinese cars, expanding the ByteDance Doubao voice assistant it launched in July to more vehicle models — and gating it behind the premium in-car entertainment subscription. The update adds selectable voice personas, folds trip tracking into the Energy app, makes range estimates account for battery aging, and adds a familiar-route navigation mode; it also ships parental controls, three traction-control modes, and a batch of security fixes. Tesla's own rollout notes put the assistant behind a paid tier, which is the part worth watching: the Doubao integration that arrived in July as a headline feature is now a subscription upsell, and Tesla is charging for the assistant rather than the car. We covered the original rollout in August — Tesla puts ByteDance's Doubao AI inside its cars.
What to watch: whether StepFun announces Step 5 at its Shanghai event with weights attached, and whether the closed version is a flagship-only exception or the start of a portfolio shift. Would you route production traffic to a 600B model that a third-party benchmark leaked before the vendor confirmed it exists? Tell us in the comments.
