HomeAI models
StepFun: Step 3.5 Flash
StepFun
•
Step 3.5 Flash
•
Released 
April 2026

StepFun
Step 3.5 Flash

Agent-tuned reasoning model with switchable low/high effort, cutting token use in high-frequency tool-call loops across a 256K window.

Modality:
Text
PDF
model ID
stepfun/step-3.5-flash
Start for free
arrow black icon
View API Docs
arrow white icon

Output Speed *

N/A
tok/s

Intelligence Index *

17
/ 100

Context Window *

262144
tokens

Input price

0.6
Anytoken

Output price

1.8
Anytoken
Step 3.5 Flash 2603: Reasoning You Can Dial Down for Agent Loops Step 3.5 Flash 2603 is StepFun's agent-optimized revision of Step 3.5 Flash. It is a text-in, text-out reasoning model with a 256K-token context window. StepFun tuned it for high-frequency agent workloads, keeping flagship reasoning and tool calling, improving token efficiency, and adding a low-reasoning mode. Its main value is fewer reasoning tokens per call across long tool-calling chains. It suits coding agents, orchestration loops and multi-step automation where most calls do not need deep thinking. Access Step 3.5 Flash 2603 through AnyAPI.ai

Performance

Token Efficiency, Not Peak Intelligence, Is the Point

Step 3.5 Flash 2603 is built to spend fewer reasoning tokens per call. StepFun reports that default high mode keeps roughly the same reasoning scores with 14% fewer tokens, while low mode cuts consumption by 56% with a slight score drop. These are vendor-reported figures. Artificial Analysis independently measures about 135 output tokens per second on StepFun's API. For agents making hundreds of calls, per-call savings compound. Route routine steps to low effort and reserve high for planning.

Benchmarks

Artificial Analysis: Index 17, With Stronger Long-Context Scores Than Step 3.5 Flash

Artificial Analysis rates Step 3.5 Flash 2603 at 17 on its Intelligence Index, behind Step 3.7 Flash at 19. Component scores reported through OpenRouter show IFBench 66.5%, AA-LCR 63.0% (original Step 3.5 Flash: 50.0%) and Terminal-Bench Hard 32.6%. Weaker areas are an AA-Omniscience non-hallucination rate of 8.7% and τ²-Bench Telecom of 87.4%, both below the original. Speed is 135 tokens/s with a 3.27s time to first token.

Output Speed

*
N/A
tok/s

Intelligence Index

*
17
/ 100

MMLU *

Broad world knowledge and problem-solving
0
%

GPQA *

PhD-level scientific reasoning across physics, biology, chemistry.
83
%

HLE *

Adherence to multi-step structured instructions.
25
%

LiveCodeBench *

Tool-calling reliability in long agentic loops.
0
%

Technical Specifications

What the model supports

Step 3.5 Flash 2603 accepts text and returns text. Image and video understanding was added in Step 3.7 Flash, not this model. The 256K-token context window is the headline limit. A per-request maximum output was not verifiable for this model, so test long generations. Reasoning effort accepts only low or high, with no medium tier. Artificial Analysis provider data lists function calling and JSON mode support.
Verified Specifications — 
Step 3.5 Flash
*
Input modalities
Text
PDF
output modalities
Text
Context window
262144
 tokens
Maximum output tokens
65536
Reasoning
Yes
Knowledge cutoff
April 2026
Pricing (standard)
0.6
 AnyTokens in
 / 
1.8
 AnyTokens out

Limitations & Trade-offs

Where Step 3.5 Flash falls short

1
Text-only input. Step 3.5 Flash 2603 accepts and returns text. StepFun's Step 3.7 Flash adds native image and video understanding. Screenshot-driven GUI agents, document-image parsing and video analysis therefore need a different model or a separate vision step. For pure text and code agents this is not a constraint. For multimodal agents, evaluate Step 3.7 Flash or another vision-capable model instead.
2
Modest measured intelligence. Artificial Analysis scores the model 17 on its Intelligence Index, below Step 3.7 Flash at 19. Humanity's Last Exam is 24.5% and CritPt is 2.3%, which shows limited headroom on frontier-level problems. StepFun also acknowledges that low effort slightly lowers reasoning scores. Hard research, advanced math or high-stakes analysis belongs on a higher-tier reasoning model.
3
Weak factual calibration and slow first answer. The AA-Omniscience non-hallucination rate is 8.7%, below the original Step 3.5 Flash at 14.3%, so closed-book factual answers need retrieval or verification. Time to first token is 3.27s, somewhat above the median for its price tier, and reasoning precedes the answer. That makes it a poor fit for latency-sensitive chat interfaces.
4
Superseded status and unclear output cap. Artificial Analysis marks 2603 as deprecated for benchmarking and suggests Step 3.7 Flash. A maximum output limit is not clearly documented for 2603: StepFun's examples use a max_tokens of 250000, while a third-party gateway config lists 65,536. New projects, or those needing long single responses, should check limits and evaluate Step 3.7 Flash first.

Best-Fit Workloads

Where this model earns its place

01

Coding agent loops

‍
StepFun specifically improved compatibility with coding tools and agent frameworks. Artificial Analysis reports 32.6% on Terminal-Bench Hard. That makes it a practical worker for repeated edit, run and inspect cycles in terminal or IDE agents. It is a fast, economical agent engine rather than a top-tier solver, so escalate difficult debugging to a stronger model.

02

Tiered-effort agent orchestration

‍
The low/high effort switch lets one model cover both routine and demanding steps. StepFun reports 56% lower token consumption in low mode and 14% lower in high mode. Use low for tool selection, parameter filling and simple routing, and high for planning or recovery steps. These savings come from the vendor's own testing, so confirm them on your traces. Only two effort levels exist.

03

Long-context codebase and document reasoning

‍
The 256K-token context window is paired with an AA-LCR score of 63.0%, up from 50.0% for the original Step 3.5 Flash, which suggests improved long-document reasoning. That suits repository-level questions, long log analysis and multi-document synthesis in text. Inputs must be text, and the weak non-hallucination rate means answers about facts not in context should be grounded or verified.

04

High-volume background automation

‍
Artificial Analysis lists it as the lowest blended-price model among StepFun's tracked models. Combined with tool calling, JSON mode and measured output speed of 135 tokens/s, it fits non-interactive pipelines such as ticket triage, data-extraction agents and scheduled workflows. Slow first-answer latency makes it a poor choice where users are waiting in real time.

Pricing in anytokens via AnyAPI
Input
0.6
₳
Output
1.8
₳
Cache write
—
₳
Cache read
—
₳
Get API access
arrow black icon
Read Docs

on this page

PerformanceBenchmarksTechnical SpecificationsLimitations & Trade-offsBest-Fit WorkloadsIntegrationFAQ

Integration

Access Step 3.5 Flash via AnyAPI.ai

Access Step 3.5 Flash through AnyAPI.ai using a unified API built for multi-model AI applications. Integrate Step 3.5 Flash without maintaining a separate provider-specific connection, and keep the flexibility to test, switch, or combine models as your application requirements evolve.

01

One API integration

Access Step 3.5 Flash and other AI models through the same API workflow instead of maintaining separate integrations for every provider.

02

Easy model switching

Test Step 3.5 Flash against alternative models or switch models as your performance, capability, or cost requirements change without rebuilding your application around another provider API.

03

Flexible for production

Use Step 3.5 Flash from experimentation through production while keeping your AI stack flexible as workloads, traffic, and model requirements evolve.

04

Multi-model applications

Use Step 3.5 Flash for the workloads where it performs best and combine it with other models for tasks that require different capabilities, performance, or efficiency.

Start free
arrow black icon
Read Docs

Frequently Asked Questions

Answers to common questions about integrating and using this AI model via AnyAPI.ai

Step 3.5 Flash 2603 has a 256K-token context window, according to StepFun's model documentation. It accepts text input and returns text output. A per-request maximum output limit is not clearly documented for this model, so treat context window and output length separately when budgeting long generations.

StepFun tuned 2603 from Step 3.5 Flash for high-frequency agent workloads, adding a low-reasoning mode and improving token efficiency, inference speed and coding-tool compatibility. StepFun reports 14% fewer tokens in high mode and 56% fewer in low mode. Artificial Analysis scores both models 17 on its Intelligence Index.

Yes. StepFun documents the reasoning_effort field for 2603 with low and high levels. The Anthropic-style Messages API likewise accepts output_config effort of low or high for this model. There is no medium level, and high is the default reasoning mode.

It is designed for coding agents, but the measured evidence is moderate. StepFun optimized it for coding tools and agent frameworks, and Artificial Analysis reports 32.6% on Terminal-Bench Hard with an Intelligence Index of 17. Treat it as a fast, economical agent worker rather than a top-tier coding model, and validate it on your own repositories.

It is not multimodal: it takes text input and returns text. Artificial Analysis lists 2603 as proprietary and the original Step 3.5 Flash as open. For image and video input, StepFun offers Step 3.7 Flash.

* Benchmark data source: Artificial Analysis artificialanalysis.ai