AI News Feed
Market watch
Anhui AI Security Lab Releases Three Solutions as OpenAI Agent Breach Raises Redline ConcernsMarkTechPost Tests Seven Voice Cloning APIs on Similarity, Consent and PriceTsinghua, Infinigence and Zhengxing Innovation Open-Source RPent as GPT-6 Astra Enters Robot TestingApple Reportedly Testing Bezel-Less, Quad-Curved Displays for iPhone 20 ProSilent Hill: Townfall Moves the Series to 1990s Scotland and First Person, TechRadar Review SaysDeveloper Introduces Mini-AGI, a Continual-Learning Model Trained on 8GB VRAM
Anhui AI Security Lab Releases Three Solutions as OpenAI Agent Breach Raises Redline ConcernsMarkTechPost Tests Seven Voice Cloning APIs on Similarity, Consent and PriceTsinghua, Infinigence and Zhengxing Innovation Open-Source RPent as GPT-6 Astra Enters Robot TestingApple Reportedly Testing Bezel-Less, Quad-Curved Displays for iPhone 20 ProSilent Hill: Townfall Moves the Series to 1990s Scotland and First Person, TechRadar Review SaysDeveloper Introduces Mini-AGI, a Continual-Learning Model Trained on 8GB VRAM
Large Language Models

StepFun Releases Step 5 Preview: 600B-Parameter MoE Model With 1M-Token Context for Agentic Work

StepFun has launched Step 5 Preview, a sparse MoE model with about 600 billion total and 27 billion active parameters, a 1M-token context window, and API pricing of $1.00 per million input tokens and $2.70 per million output tokens. Open weights are scheduled for October 15, 2026.

AI News Feed · 2026-09-21 16:44

The launch argument is cost. StepFun states the model delivers comparable intelligence at a substantially lower task cost, a claim its launch materials frame as a Pareto frontier. Step 5 Preview is available now as a hosted API and on the StepFun platform; self-hosting depends on open weights, which StepFun says will be published on October 15, 2026. At 600 billion parameters, the weights would require roughly 1.2 TB of storage in BF16 before any KV cache, pointing to multi-GPU server hardware once they ship.

The documented specifications list the model ID as step-5-preview, a 1 million token context window, text, image and video input with text output, and selectable low, medium and high reasoning effort. Streaming, tool calling, JSON Mode, JSON Schema and prompt caching are supported. StepFun states that on research tasks the model coordinated 950 web fetches within a single agent action, and it documents a Claude Code integration through its Step Plan.

On architecture, Pandaily reported that StepFun did not widen the network but stacked 92 Transformer layers in a narrow-deep layout. The research team argues deeper stacks give longer paths for implicit multi-hop reasoning, which it says matters during long prefill when agents search, run code and read tool returns. Training leans on on-policy, long-horizon reinforcement learning, with bit-wise train and inference alignment across MoE routing. Other listed techniques include MTP-3 speculative decoding, FP8 MoE and KV-cache offload. StepFun reports more than 3x end-to-end speedup for long-horizon reinforcement learning.

Benchmark results reported by the company, which ran Step 5 Preview at High effort while rivals ran at Max, put the model at 66.4 on FrontierFinance against 69.7 for Claude Opus 5 and 55 for GPT-6 Astra, and at 83.3 on DRACO against 87.6 and 76.8 respectively, according to RuntimeWire. On coding, StepFun reports 67.7 on DeepSWE v1.1, 49.0 on StepCodeBench and 80.5 on ProgramBench, with GPT-6 Astra and Claude Opus 5 ahead on all three; StepCodeBench is StepFun's own benchmark. The company also ran two agent experiments of 24 hours each, tuning an H100 kernel to 508 TFLOPS against 493 for Claude Opus 5, and raising Qwen3-30B-A3B on AIME24 from 53.3 percent to 60 percent through automated post-training.

Artificial Analysis supplies the independent check, scoring Step 5 Preview at 44 on its Intelligence Index against a median of 24 for reasoning models in a similar price tier, and measuring output at 99.8 tokens per second on StepFun's API. The same evaluator puts median prices for comparable models at $1.88 per million input tokens and $10.00 per million output tokens. One caveat accompanies the price advantage: the model generated 160 million output tokens on the index run, against a median of 92 million, so verbose reasoning consumes part of the per-token saving.

Editor's Summary

StepFun's Step 5 Preview is a 600B-total, 27B-active MoE model with a 1M-token context window, text, image and video input, and API pricing of $1.00 per million input tokens and $2.70 per million output tokens. Company-reported benchmarks show it trailing Claude Opus 5 and GPT-6 Astra on finance and coding, while Artificial Analysis scores it 44 on its Intelligence Index against a tier median of 24. Open weights are scheduled for October 15, 2026, which will determine whether the model can be self-hosted.