Step 5 Preview: Pricing, Benchmarks and the October 15 Open Weights

Last updated: October 9, 2026

Step 5 Preview is StepFun's new flagship, a 600B-parameter sparse mixture-of-experts model with 27B active per token and a 1M-token context window, listed on OpenRouter at $1.00 input and $2.70 output per million tokens. It scores 44 on the Artificial Analysis Intelligence Index, at a task cost Artificial Analysis puts at $1.03. StepFun says the weights ship on October 15, so for the next six days this is an API-only model, and the useful question is whether to test it now or wait for the download.

Official Sources#

What Shipped#

StepFun calls Step 5 Preview its "new flagship model for agentic work", aimed at software engineering and professional knowledge work, with finance as the stated specialty. The architecture is a sparse MoE: 600B parameters in total, 27B active per token. It takes text and image input and returns text, with a 1M-token window and up to 64,000 completion tokens on OpenRouter.

Two dates matter. StepFun announced the model in September (Artificial Analysis lists a September release), and it reached OpenRouter on October 8, which is why it climbed Hacker News this week. The weights are a promise, not a fact: StepFun says "the model will be released with open weights on October 15" and has not named a license in the announcement. Artificial Analysis still labels it proprietary today. Plan around the date, not the license, until the model card is up.

Benchmarks#

Everything below comes from StepFun's own table unless marked otherwise. Treat the vendor-reported rows as a claim to reproduce, and note StepFun compares its "High" setting against rivals at "Max".

The pattern is consistent. On software-engineering benchmarks it sits with the open frontier and below the two closed flagships. The announcement itself concedes that on the most challenging long-running tasks "a meaningful gap to the frontier remains". The weak spot is the terminal: 33.3% on Terminal-Bench v4 trails GLM-5.3 by about eight points. If your agent lives in a shell, that is the row to retest on your own repo.

The independent number is the Artificial Analysis Intelligence Index: 44, rank 40 of 226 models tracked, against a median of 26 for comparable reasoning models. The same page flags the catch. Step 5 Preview generated 160M output tokens to run the index, against a median of 81M, which it calls "very verbose". Cheap per token and chatty per task is a combination that only shows up when you price a whole session.

Pricing and Context#

Prices checked on October 9, 2026 against the OpenRouter page. Artificial Analysis lists the same $1.00 and $2.70 and a 90% cache discount.

A worked session, with the assumptions stated so you can redo it: a 40-turn agent run where each turn sends 60K tokens of context (90% cache hits) and produces 2K output tokens. Per turn, 54K cached tokens cost $0.0027, 6K uncached cost $0.006 and 2K output costs $0.0054, about $0.014. Forty turns is roughly $0.56. The model's verbosity pushes the output line up, and output is the expensive line, so measure your own token mix before trusting any per-task figure.

OpenRouter's own telemetry from the first days is worth reading before you commit. On October 9 the page showed a cache hit rate of about 95%, a tool call error rate under 2% and a structured output error rate of about 17%. Those are early, moving numbers from one provider, but a structured-output error rate near 17% is the figure to test if your pipeline relies on JSON schemas.

Try It Today#

The OpenRouter page documents an OpenAI-compatible endpoint. This is the curl it publishes, with the model id swapped in:

export OPENROUTER_API_KEY=sk-or-v1-...
curl -N https://openrouter.ai/api/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $OPENROUTER_API_KEY" \
  -d '{
  "model": "stepfun/step-5-preview",
  "stream": true,
  "messages": [{"role": "user", "content": "Hello"}]
}'

Any coding agent that accepts an OpenRouter or OpenAI-compatible provider can point at the same id; see the OpenCode developer guide for wiring a custom provider. Run your own repo's failing-test task through it before reading anything into the tables above.

What People Are Actually Saying#

The Hacker News thread (130+ points) is split in a useful way.

• The bull case is speed and price. One commenter notes OpenRouter showing an average of 115 tokens per second and says they would "pay a little extra for faster inference", while flagging that 1.5s latency may blunt that for agentic work. Another calls Step 3.5 Flash and Step 3.7 Flash "easily my favourite local models" this year and hopes the new one is "a lot faster than GLM".
• The bear case is the index. A commenter links the Artificial Analysis page and says it "doesn't look competitive along any dimension". Replies push back on speed: Step 5 and DeepSeek run at much higher tokens per second than most of the open field, which a single score does not capture.
• The local case does not work yet. One commenter who ran earlier Step models on 128GB of shared memory, and wanted to compare this one with Qwen Flash Next, found it is 600B-A27B and wrote "no way to run that". That is the honest read: even quantised, this is a multi-GPU or large-memory job, unlike the 125B Qwen that runs on a gaming PC through Strata.
• The comparison table from the thread puts Step 5 at an index of 44 with 160M output tokens, next to Kimi K3 (max) at 44 with 160M, GLM-5.3 at 45 with 210M and DeepSeek 4.1 Flash at 39 with 250M. It is one commenter's reading of Artificial Analysis, so check the live page.

Which One to Pick#

Monday's plan: add stepfun/step-5-preview as one candidate in your model-routing bake-off, run a 10-task slice of your own repo, and log tokens per task rather than price per token. If the weights land on October 15 with a permissive license, rerun the slice against a self-hosted copy. See the best local coding LLMs roundup for the hardware side.

FAQ#

What is the API model id for Step 5 Preview?#

On OpenRouter the id is stepfun/step-5-preview, served by StepFun as the only provider at the time of writing. It supports tool calling and JSON-schema structured outputs, though early telemetry from that single provider showed a structured-output error rate of about 17%, so test your schemas.

How much does Step 5 Preview cost?#

$1.00 per million input tokens, $2.70 per million output tokens and $0.05 per million cached input tokens, with a 1M-token context window. Artificial Analysis measures $1.03 to run one Intelligence Index task and flags the model as verbose.

Is Step 5 Preview open source?#

Not yet. StepFun says open weights will be released on October 15, 2026, but its announcement names no license and Artificial Analysis currently lists it as proprietary. Check the Hugging Face model card on the day before building on it.

Sources#

Continue Reading#

• Kimi K3 Developer Guide: What the 2.8T Open Model Changes - the open model Step 5 Preview is benchmarked against
• GLM-5.3 Free and Cheap: Every Route and Price (October 2026) - the other open-weight rival, with route-by-route pricing
• DeepSeek 4.1 Flash as a Daily Coding Model: Cache, Peak Hours and Limits - the cheap option to price a session against
• The Best Local Coding LLMs in 2026: Run Enterprise-Grade AI Without the Cloud - what actually fits on your own hardware
• Run Qwen3.8-Flash-Next on a Gaming PC With Strata - the local counterpoint at 125B