What Is Step 5 Preview? StepFun’s 600B Agent Model Explained

Step 5 Preview explained: explore StepFun’s 600B MoE model, pricing, coding and finance benchmarks, 1M context, API access, and release status.

What Is Step 5 Preview? StepFun’s 600B Agent Model Explained

Updated September 21, 2026

Step 5 Preview is StepFun’s new flagship reasoning model for coding, long-running agents, professional research, and finance. The headline numbers are unusual: a sparse 600-billion-parameter Mixture-of-Experts architecture, 27 billion active parameters per token, a 1-million-token context window, and native text, image, and video input.

Those specifications make the Chinese AI model worth watching, but do not settle whether it fits a production workload. Independent tests give it a small intelligence lead over GLM 5.3 Flash and DeepSeek V4.1 Flash; both rivals are cheaper, and DeepSeek is faster.

One availability note matters before going further: Step 5 Preview is available through StepFun’s own API, but it is not yet available on GPTProto as of September 21, 2026. GPTProto plans to add it later. Developers who need a model through GPTProto today can evaluate GLM 5.3 Flash, DeepSeek Flash, or browse the text model catalog.

목차

What Is Step 5 Preview?

Step 5 Preview is a multimodal reasoning model developed by Shanghai-based StepFun. It is intended to plan, call tools, inspect results, revise its approach, and continue until it produces a deliverable.

Here are the main facts from the official model documentation and StepFun’s launch page:

Specification Step 5 Preview
Developer StepFun
Model ID step-5-preview
Architecture Sparse Mixture of Experts
Parameters 600B total, 27B active per token
Context window 1M tokens
Maximum output 64K tokens
Input Text, images, and video
Output Text
Reasoning control low, medium, or high
API status Available from StepFun
Open-weight status Planned for October 15, 2026
GPT Proto status Not available yet; integration planned

“Preview” is important. The API can be used now, but behavior, limits, pricing, and even naming may change before a stable release. StepFun says the model’s weights will be released on October 15, 2026. Until that happens, it should be treated as a hosted proprietary model—not as an open-weight model that can already be self-hosted.

Step 5 Preview Features and Architecture

600B parameters without activating all 600B

Step 5 Preview uses a sparse Mixture-of-Experts, or MoE, design. It contains 600 billion parameters in total, but only 27 billion are active for each token. A routing system selects the experts relevant to the current input.

“600B” describes total capacity; it does not mean every token incurs the cost of a dense 600B model. The active count is more relevant to inference efficiency, but hardware, batching, reasoning length, and provider capacity still determine real performance.

A 1M-token context and 64K-token output

The 1-million-token context window targets large repositories, reports, conversations, and document collections. Maximum output is 64,000 tokens. Large context still requires disciplined retrieval: more material can add irrelevant evidence, conflicting instructions, and cost.

Prompt caching makes repeated work over the same codebase, policy library, or research corpus much cheaper than repeatedly sending uncached input.

Text, image, and video input

The model accepts text, up to 60 images per request, and video. Images can be JPEG, PNG, WebP, or static GIF. Video formats include MP4, QuickTime, and Matroska; StepFun recommends clips under five minutes.

You may see third-party databases list only text and image input. That does not necessarily contradict StepFun. For example, Artificial Analysis records the modalities covered by its listing and tests, while StepFun documents video support in its own API. For implementation limits, use the first-party documentation; for comparative performance, use independent testing.

Tool calling is not built-in access to the outside world

Step 5 Preview supports streaming, function calling, JSON mode, JSON Schema, and adjustable reasoning effort. However, the model does not independently browse the web, execute code, or access a developer’s files. StepFun explicitly states that search, code execution, document handling, and other tools must be supplied by the integrating application.

The model can decide when and how to use a tool, but the surrounding agent system controls its tools, permissions, and result validation.

Step 5 Preview for Coding and Agentic Work

In a coding agent, the model receives a goal, reads files, applies changes, runs tests, observes failures, and iterates. Its long context can preserve repository structure and earlier decisions across that loop.

StepFun reports the following results for the model at high reasoning effort:

Benchmark Step 5 Preview score What it tests
DeepSWE v1.1 67.7% Repository-level software engineering
StepCodeBench 49.0% Multi-language engineering tasks
Terminal-Bench v4 33.3% Coding and terminal tool use
Agents’ Last Exam (ALE-CLI) 29.5% Computer-use agent tasks
GPQA Diamond 93.5% Graduate-level reasoning and knowledge
HLE 46.5% Broad, difficult expert questions

These are vendor-reported results. StepCodeBench is internal; some rivals use “Max” effort while Step 5 uses “High”; and some HLE-with-tools comparisons use different subsets. The scores show capability, not guaranteed superiority on a particular repository.

StepFun acknowledges a meaningful gap on the hardest long-running coding tasks. Test bug repair, feature changes, test generation, and rollback on your own code, then measure accepted tasks per dollar and minute.

Early community evidence is thin. In one Linux DO launch thread, a user liked the visible intermediate reasoning; Reddit discussion focused more on leaked or mirrored weights than production reliability. These launch-day anecdotes suggest questions to test, not a ranking.

Why Step 5 Preview Focuses on Finance

Step 5 Preview in finance is not simply about answering accounting questions. StepFun targets analyst workflows that gather filings and market data, reconcile definitions, build valuations, run calculations, and produce an auditable report. Sources, assumptions, formulas, and sensitivities must remain traceable.

StepFun reports 66.4% on the external FrontierFinance benchmark. It also reports 74.5% on FinStepBench LiveSearch, 60.6% on CorporateValuation, and 55.8% on Finance Deep Research. The last three are StepFun’s internal benchmarks, so they should carry less weight than independent tests until third parties reproduce the results.

A strong score does not make unsupervised financial decisions safe. Keep source links, calculation traces, approval gates, and human review—especially for investment, credit, tax, or compliance decisions.

How Good Is Step 5 Preview in Independent Tests?

Artificial Analysis gives Step 5 Preview an Intelligence Index score of 44. It measured output at 99.8 tokens per second, a 2.96-second time to first token, and $0.72 per Intelligence Index task on StepFun’s API.

The important warning is verbosity: 160 million output tokens across the evaluation versus a 92-million class median. A reasonable token price can still produce a high task cost, so production systems need output caps, appropriate reasoning effort, and stop conditions.

Independent testing also lists text and image input, not video. That reflects the scope of the evaluator’s model record; StepFun’s official API documentation remains the source for video capability. Keeping those two evidence types separate avoids turning a difference in test coverage into a false contradiction.

Step 5 Preview Pricing Explained

StepFun’s official pricing page lists three rates per 1 million tokens:

Token type Price
Uncached input $1.00
Cached input $0.05
Output $2.70

For Step 5 Preview, cache-miss input pricing includes writing new content to the cache. Output billing includes both reasoning tokens and the final answer, so hidden reasoning is not free.

For example, 100,000 uncached input tokens plus 10,000 output tokens cost about $0.127. With the input fully cached, the same request costs about $0.032. Tool loops, retries, and reasoning length can change the total.

This is StepFun’s price, not a GPT Proto price. GPT Proto has not added the model yet, so there is no legitimate GPT Proto Step 5 Preview rate to quote.

Step 5 Preview vs GLM 5.3 Flash vs DeepSeek Flash

This table uses Artificial Analysis data current on September 21, 2026. DeepSeek Flash means DeepSeek V4.1 Flash at maximum reasoning effort.

Model Intelligence Index Output speed Input / output per 1M tokens Cost per task Context Weights
Step 5 Preview 44 99.8 tok/s $1.00 / $2.70 $0.72 1M Planned Oct. 15
GLM 5.3 Flash 42 95.0 tok/s $0.15 / $0.50 $0.25 1M Open, MIT
DeepSeek V4.1 Flash 39 242.3 tok/s $0.30 / $1.20 $0.27 1M Open, MIT

GLM 5.3 Flash vs Step 5 Preview

Step 5 Preview leads by two Index points at a similar speed. GLM is far cheaper and its MIT-licensed weights are available now. Step 5 is more interesting for finance-specific workflows or official video input; GLM is the stronger starting point for cost and self-hosting.

Step 5 Preview vs DeepSeek Flash

Step 5 leads by five Index points, but DeepSeek is more than twice as fast, costs much less per task, and is already open-weight. Choose Step 5 only if its reasoning produces enough additional accepted work to justify slower, costlier runs.

GPT Proto currently offers both GLM 5.3 Flash and DeepSeek Flash. For a deeper comparison between those available options, see GLM 5.3 Flash vs DeepSeek V4 Flash.

How Can You Access Step 5 Preview?

Developers can call StepFun’s POST /v1/chat/completions endpoint with model ID step-5-preview. Its request shape follows the OpenAI chat-completions style. This example targets StepFun directly; it is not a GPT Proto endpoint.

curl https://api.stepfun.ai/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $STEP_API_KEY" \
  -d '{
    "model": "step-5-preview",
    "messages": [
      {
        "role": "user",
        "content": "Review this repository migration plan and list the three highest-risk assumptions."
      }
    ],
    "reasoning_effort": "medium",
    "max_tokens": 1200
  }'

The official API reference should be checked before production use for current parameters, errors, rate limits, and multimodal request formats.

If you prefer one balance and a unified interface across many models, GPT Proto plans to add Step 5 Preview later. Until the model appears in the live catalog, do not assume that a similar name, a third-party mirror, or an unofficial route is the GPT Proto integration.

Is Step 5 Preview Worth Trying?

Step 5 Preview deserves a controlled evaluation for long documents, coding agents, multimodal research, or traceable finance workflows. Its independent intelligence score is strong and its 1M context is useful.

It is not the default choice. GLM is cheaper and open-weight; DeepSeek is much faster; and Step 5 remains verbose, in preview, and unavailable on GPT Proto.

The best decision metric is completed, reviewed work—not parameter count or a single benchmark. Run the same representative tasks through each candidate, record correctness, intervention rate, latency, and total token cost, then choose the model that produces the lowest cost per accepted result.

Frequently Asked Questions

What is Step 5 Preview?

Step 5 Preview is StepFun’s 600B-total, 27B-active sparse MoE reasoning model for agentic coding, professional work, finance, and multimodal analysis. It supports a 1M-token context and text, image, and video input.

When was Step 5 Preview released?

Step 5 Preview entered API availability in September 2026. Artificial Analysis records September 18, while StepFun’s public launch announcement appeared around September 20. StepFun says open weights are planned for October 15, 2026.

How much does Step 5 Preview cost?

StepFun charges $1.00 per 1M uncached input tokens, $0.05 per 1M cached input tokens, and $2.70 per 1M output tokens. Reasoning tokens are included in output billing.

Is Step 5 Preview open source?

Not yet. StepFun says it plans to release open weights on October 15, 2026. The license and final release details should be checked when the weights actually appear.

Does Step 5 Preview support video?

Yes. StepFun’s official documentation lists text, image, and video input, with text output. Some independent model databases list only the modalities they tested, so their entries may show text and image only.

Is Step 5 Preview good for coding?

Vendor benchmarks suggest strong repository-level coding and agent behavior, but results vary across tests, and StepFun acknowledges a remaining gap on the hardest long-running tasks. Evaluate it on your own repositories before making a production decision.

Is Step 5 Preview better than GLM 5.3 Flash?

It scores slightly higher in the current Artificial Analysis Intelligence Index, 44 versus 42. GLM 5.3 Flash is substantially cheaper and already open-weight, while measured output speeds are similar.
What Is Jev? TypeSafe AI’s System One Model Explained

What Is Jev? TypeSafe AI’s System One Model Explained

TL;DR Jev is TypeSafe AI’s first System One model : a model designed to turn text or text-shaped application state into predefined, typed decisions with probabilities. Unlike a standard large language model, Jev is not meant to write an answer one token at a time. It evaluates bounded questions and returns structured results in parallel. That makes it interesting for classification, routing, scoring, validation, and agent action selection—but not for open-ended writing, coding, or multi-step reasoning. The practical way to think about Jev is not “a faster chatbot.” It is a probabilistic decision component that software can call when ordinary rules are too brittle but free-form generation is unnecessary. One Key for Your Team Availability note: Jev is not currently available through GPTProto. GPTProto may add support in the future. For now, you can explore the models that are already available in the GPTProto model catalog .

Michael Johnson | 2026-09-20

5 Best APIs for Tech Startups in 2026: A Lean MVP Stack

5 Best APIs for Tech Startups in 2026: A Lean MVP Stack

A startup rarely loses its first month because it chose the “wrong” database brand. It loses the month at the seams: mismatched permissions, payment events that fail to update subscriptions, leaked AI keys, or missing transactional emails. This is therefore a practical API stack for a subscription-based web product—especially an AI SaaS MVP—not a directory of unrelated tools. My recommended default is GPTProto for AI inference, Supabase for data and backend services, Stripe for payments, and Resend for transactional email . Clerk is the fifth option, but it is an upgrade rather than a requirement because Supabase already includes authentication. The stack can begin with no fixed monthly platform fees on the non-AI services, although model calls, successful payments, and excess usage still create variable costs. One Key for Your Team Pricing and plan limits in this guide were checked on September 18, 2026. Verify the linked product pages before committing a production budget.

Schuyler Stacy | 2026-09-18

What Is GPT-6 Sol? Release Status, Pricing, Tests, and What We Know

What Is GPT-6 Sol? Release Status, Pricing, Tests, and What We Know

GPT-6 Sol is currently a reported OpenAI model label, not a publicly documented model . Screenshots and community reports suggest that a Sol-tier GPT-6 model may be in testing, but OpenAI has not published a model page, API ID, price, specifications, benchmarks, or release date for it. A name appearing in an interface is evidence worth investigating, but it is not a supported production endpoint. The public OpenAI model catalog lists GPT-6 Astra and the GPT-5.6 family. It does not list GPT-6 Sol. Last checked: September 17, 2026. OpenAI had not published a GPT-6 Sol model reference, API price, benchmark, or release post. We will update this guide if that changes. Explore GPT API Key Question Verified answer Has OpenAI announced GPT-6 Sol? No official announcement found Is gpt-6-sol a documented public model ID? No Is there evidence that the label exists? Yes, in community screenshots and reports Is there a confirmed release date? No Has OpenAI published GPT-6 Sol pricing? No Are its context window and knowledge cutoff known? No Are there verified GPT-6 Sol benchmarks? No Can you call it through GPTProto? No confirmed GPTProto model page or supported model string exists yet

Michael Johnson | 2026-09-17

6 Best Affordable LLM APIs for AI Agents in 2026

6 Best Affordable LLM APIs for AI Agents in 2026

An affordable LLM API for an AI agent is not necessarily the model with the lowest input-token price. An agent may choose a tool, construct arguments, read the result, revise its plan, and call another tool before it produces a useful answer. A cheap model that makes invalid calls or needs several retries can therefore cost more than a slightly more expensive model that finishes the task once. This guide compares six agent-ready models available through GPTProto. The ranking considers API price, tool use, independent performance evidence, speed, context limits, and the practical risk of paying for unnecessary agent loops. It is a public-benchmark and pricing comparison—not a claim that we ran a private head-to-head test. One Key for Your Team Quick answer: GLM-5.3 Flash is the strongest default for most cost-sensitive agents. DeepSeek Flash is the faster open-weight alternative, while GPT-5.6 Luna is promising for lightweight, high-volume work once its live route price is confirmed. MiniMax M3 fits long document sessions, Gemini 3.8 Flash leads on multimodal speed, and Grok 4.6 is better treated as an escalation model for harder tasks.

Michael Johnson | 2026-09-15