What is Step 5 Preview?

Step 5 Preview is StepFun's frontier model, announced September 20, 2026 with same-day API access. It is a 600B-parameter sparse mixture-of-experts model with about 27B parameters active per token, a 1M-token context window with 1M maximum output, and native text, image and video input. The API price is $1.00 per million input tokens and $2.70 per million output tokens, with a 95% discount on cached input. Artificial Analysis scored it 44 on its Intelligence Index, level with Kimi K3 Max, at a measured $0.71 per task. StepFun says full open weights follow on October 15, 2026.

Is Step 5 Preview cheaper than Kimi K3, DeepSeek V4 Pro and GLM-5.3?

It sits in the middle. Per million tokens as of September 20, 2026: DeepSeek V4 Pro is $0.66/$1.98 off-peak (double at peak), Step 5 Preview is $1.00/$2.70, GLM-5.3 is $1.40/$4.40 and Kimi K3 is $3/$15. On Artificial Analysis's measured cost to run its full evaluation, Step 5 Preview came in at $0.71 per task against about $2 for Kimi K3 Max, because it uses fewer output tokens per answer. DeepSeek V4 Pro remains the cheapest list price, but only outside its 01:00–04:00 and 06:00–10:00 UTC peak windows.

Which of these models has open weights?

As of September 20, 2026: Kimi K3 has open weights (released July 27, 2026). DeepSeek V4 Pro and V4 Flash are open weight. GLM-5.3's open weights were announced as staged and had not shipped at launch on August 14. Step 5 Preview is API-only today; StepFun has committed to releasing full weights on October 15, 2026, and its Hugging Face repository currently contains only a placeholder file. Treat any 'open' claim for Step 5 as a promise until the checkpoint lands.

Which model should I pick for long-context agent work?

For very long inputs, Step 5 Preview and DeepSeek V4 Pro both offer 1M-token context; Step 5 also allows up to 1M output tokens, which is unusual, and accepts up to 60 images or a video per request. Kimi K3 leads on raw intelligence among open-weight models but is the most expensive of the four at $15 per million output tokens. GLM-5.3 is the strongest on coding-agent benchmarks per its vendor and has a $18/month Coding Plan. If you need multimodal input at low cost, Step 5 Preview is the only one of the four with native video input.

Quick Answer

Step 5 Preview vs Kimi K3 vs DeepSeek V4 Pro vs GLM-5.3

Published: September 20, 2026

The four models side by side

	Step 5 Preview	Kimi K3	DeepSeek V4 Pro	GLM-5.3
Maker	StepFun (Shanghai)	Moonshot AI (Beijing)	DeepSeek (Hangzhou)	Z.ai / Zhipu (Beijing)
Released	September 20, 2026	July 27, 2026 (weights)	GA August 13, 2026	August 14, 2026
Architecture	600B sparse MoE, ~27B active	Open-weight MoE	MoE, 1M ctx	743B base, post-trained
Context	1M in / 1M out	Long context	1M in / 384K out	Long context, thinking mandatory
Input modalities	Text, image (60/request), video	Text	Text	Text
API price (per MTok)	$1.00 / $2.70	$3 / $15 (flat)	$0.66 / $1.98 off-peak; $1.32 / $3.96 peak	$1.40 / $4.40
Cached input	95% discount	—	$0.022 off-peak	$0.26
Open weights	Promised October 15, 2026	Yes	Yes	Staged, not shipped at launch
AA measured cost per task	$0.71	~$2 (K3 Max)	—	—

Verified September 20, 2026 against StepFun’s platform docs, Artificial Analysis and the vendors’ pricing pages. Two caveats before quoting: DeepSeek prices double during its peak windows, and Artificial Analysis re-versioned its Intelligence Index in September, so scores from August posts are not comparable with the new 44-point Step 5 number.

What Step 5 Preview actually is

StepFun announced Step 5 Preview on September 20, 2026 and opened its API the same day. The company describes it as “StepFun’s frontier model for production-scale Agent applications,” aimed at programming and professional knowledge work.

The headline specs from StepFun’s own documentation:

600B total parameters, sparse mixture-of-experts, roughly 27B active per token.
1M-token context window, with both maximum input and maximum output listed at 1M tokens.
Native text, image and video input. Up to 60 images per request (low or high detail), and MP4/QuickTime/Matroska video under 128 MB, ideally under five minutes.
Model ID step-5-preview; OpenAI-compatible chat API.

Pricing is $1.00 per million input tokens and $2.70 per million output tokens, and cached input is discounted 95%. Artificial Analysis calculated a blended $0.51 per million tokens at a 7:2:1 cache-hit/input/output ratio, which made StepFun the cheapest provider of its own model.

On Artificial Analysis’s Intelligence Index, Step 5 Preview scored 44, matching Kimi K3 Max, with an output speed of about 100 tokens per second and a measured $0.71 to run the full evaluation versus roughly $2 for Kimi K3 Max. StepFun’s own demos emphasise long-horizon autonomy: it reports the model optimising H100 GPU kernels unattended for 24 hours to reach 508 TFLOPS.

The open-weights caveat is important. StepFun says the full weights ship on October 15, 2026. As of September 20 the Hugging Face repository contains only a .gitattributes file, so today Step 5 is a hosted API, not a model you can run.

Kimi K3 — the intelligence leader, at a price

Moonshot’s Kimi K3 shipped open weights on July 27, 2026 and has led the open-weight tier on Artificial Analysis since. Its API price is a flat $3 input / $15 output per million tokens — no peak/off-peak games, but the most expensive output of the four, and more than Claude Sonnet 5 ($10). The “K3 Max” configuration is what Artificial Analysis benchmarked against Step 5 at 44 points; it consumed about 2.8x more money to finish the same evaluation.

Choose K3 when you want the highest-capability weights you can self-host and the token bill is not the constraint. See the Kimi K3 local hardware guide.

DeepSeek V4 Pro — the price floor with a schedule

DeepSeek V4 Pro reached GA on August 13, 2026 with 1M context and 384K maximum output. On August 16 DeepSeek moved to peak/off-peak pricing: $0.66 / $1.98 off-peak and exactly double, $1.32 / $3.96, during peak (01:00–04:00 and 06:00–10:00 UTC). Cache hits are $0.022 per million off-peak. For US and European working hours you mostly land off-peak; Asia-Pacific business hours mostly land in peak.

Choose V4 Pro when you can batch work into off-peak windows and want open weights with a 1M context. Do not quote a single flat V4 price — there isn’t one.

GLM-5.3 — the coding-agent specialist

Z.ai’s GLM-5.3 (August 14, 2026) is the same 743B base as GLM-5.2 with new post-training, priced at $1.40 / $4.40 per million tokens with $0.26 cached input. Thinking mode is mandatory. The $18/month Coding Plan and 50% off-peak points (outside 14:00–18:00 UTC+8 weekdays) make it the cheapest way to run a Chinese frontier model inside Claude Code-style agents. Open weights were announced as staged and had not shipped at launch.

Choose GLM-5.3 when the workload is coding agents and you want a subscription rather than per-token metering. See GLM-5.3 vs Opus 5 vs GPT-5.6 Sol for coding.

Head-to-head on the questions developers ask

“Which is cheapest?” List price: DeepSeek V4 Pro off-peak. Measured cost per finished task on Artificial Analysis’s harness: Step 5 Preview at $0.71 is the only one of the four with a published figure this month, and it beat Kimi K3 Max by ~2.8x. Token efficiency, not list price, decided that.

“Which handles the longest documents?” Step 5 Preview and DeepSeek V4 Pro at 1M input. Step 5 uniquely lists 1M output, which matters for agents that write whole repositories or long reports in one pass.

“Which can see images and video?” Only Step 5 Preview, natively. The others are text-only through their first-party APIs.

“Which can I self-host today?” Kimi K3 and DeepSeek V4 Pro. GLM-5.3 and Step 5 are promises for now.

“Which is smartest?” On Artificial Analysis’s current index, Step 5 Preview and Kimi K3 Max tie at 44. Independent numbers for GLM-5.3 and V4 Pro on the re-versioned index were not published as of September 20.

Bottom line

Step 5 Preview is the first Chinese model to combine 1M-context, native video input and sub-$3 output pricing in one API, and it matches Kimi K3 Max on Artificial Analysis’s index at roughly a third of the cost. Until October 15 it is API-only, so teams with a hard self-hosting requirement should stay on Kimi K3 or DeepSeek V4 Pro. For a broader view of the Chinese open-weight race, see DeepSeek vs Moonshot vs Z.ai vs MiniMax.