Step 5 Preview Outperforms DeepSeek V4 Pro in Coding Tasks at Fraction of Opus 5 Cost

Introduction

The author evaluates StepFun's newly released Step 5 Preview model through three practical coding tasks, comparing it side-by-side with DeepSeek V4 Pro. Each task uses identical prompts but different toolchains: Step 5 Preview via WorkBuddy with browser-based verification, DeepSeek V4 Pro via local DeepSeek Harness with syntax checks only. A third-party model (GLM) provides blind evaluations.

Step 5 Preview Model Overview

Step 5 Preview is positioned as "Advancing the Pareto Frontier" — a sparse Mixture-of-Experts (MoE) model with 600B total parameters, 27B activated per token, 1M token context window, and vision input support. Official benchmarks show Step 5 Preview (High) scoring 67.7 on DeepSWE v1.1, 80.5 on ProgramBench, and 49.0 on StepCodeBench (avg@4), surpassing Kimi K3 (43.9) and GLM-5.3 (40.2), though trailing Opus 5 (63.9). Cost per task is reported at 35% of GLM-5.3/Kimi K3 and 12.5% of Claude Opus 5.

WorkBuddy Integration Setup

To call Step 5 Preview, the author creates an API key on the StepFun open platform ( https://platform.stepfun.com/), then configures a custom model in WorkBuddy with:

API endpoint: https://api.stepfun.com/v1/chat/completions (Step Plan uses https://api.stepfun.com/step_plan)

API key from the platform

Model name: step-5-preview Advanced: enable tool calling; use provider defaults for input/output length

Case 1: 2D Animation — Sun Wukong Pilots Plane, Pelican Passenger

Step 5 Preview Output

Completed in 30 minutes 3 seconds . Uses Canvas 2D rendering. The animation depicts Sun Wukong with golden headband, monkey face, and tiger-striped clothing in the pilot seat; the pelican in the rear seat has a long beak, throat pouch, flight cap, goggles, and red scarf. Control stick, wings, tail, background mountains, and birds are all drawn. A single timeline drives all actions. Verification via headless Chrome and pixel checks caught two issues: premature phase-label change and reversed landing-gear animation, both fixed.

DeepSeek V4 Pro Output

Completed in 9 minutes 49 seconds via local Harness. Also uses Canvas and a unified animation clock. The screenshot shows runway and wheels still visible; Sun Wukong identified by headband, monkey face, and golden staff; pelican retains beak and throat pouch; two dark cockpit areas distinguish front/rear seats. However, the harness only performed node --check syntax validation — no actual browser preview or real-device acceptance test.

Comparison

Step 5 Preview shows richer character accessories and environmental details. DeepSeek uses simpler shapes but satisfies role/seat requirements. The two screenshots capture different flight phases (cruise vs. takeoff), so direct visual comparison is limited. On the "trigger turbulence" button during pause, Step 5 Preview logs the turbulence window while DeepSeek ignores the click. Dynamic behavior and button states need identical interactive testing; static images cannot confirm full compliance. Time difference (30m vs 9m) reflects toolchain and verification scope differences, not raw model speed alone.

Case 2: Token Bucket Rate Limiting Interactive Demo

Task: Build a Vite + TypeScript Chinese interactive demo. Default capacity 5, refill 2 tokens/second, time advanced only by button clicks. Required sequence: request 3, request 3, advance 1s, request 4, advance 5s, request 5. Expected final state: 0 tokens, virtual time 6s, 3 allowed, 1 rejected. Rejection must not deduct tokens; refill must not exceed capacity; reset must clear logs and stats.

Step 5 Preview Output

Time: 29 minutes 5 seconds . UI: left side shows bucket level with liquid and dots; right side groups config and action buttons; bottom has independent scrolling log area. A six-step guide highlights the next pending action. Screenshot shows after two requests: tokens 5→2, second request rejected (tokens stay 2), time 0s, allowed 1, rejected 1, progress "In progress 3/6" with hint "Next: advance 1s".

Delivery report: 37 test cases passed , type-check and build successful, Chromium verification of six-step demo, invalid input, reset, and 1440×900 layout. Fixed a long-log layout overflow by constraining log scrolling to its own region.

DeepSeek V4 Pro Output

Time: 11 minutes 6 seconds , 45 tool calls . Three-column layout: bucket state, config/actions, operation log. Color-coded log entries for allow, reject, time advance. Screenshot shows after step 3 (advance 1s): tokens 4, virtual time 1s, log consistent, stats allow 1 reject 1. Progress shows "Completed 3/6 steps" (different meaning from Step 5 Preview's "In progress 3/6").

Delivery report: 20 automated tests passed (10 algorithm, 10 controller), type-check and build pass, 36 browser checks via local Chrome. Notes: layout checks via DOM size and scroll assertions, no pixel visual diff, no mobile adaptation; desktop-only per requirements.

Blind Evaluation (Third-Party Model)

Functional verdict: tie. Both pass identical six-step demo, reset, invalid input checks, and re-run each other's tests. Differences lie in pedagogical expression and engineering details. Step 5 Preview explicitly indicates how many tokens are dropped when refill hits capacity, and provides a six-step checklist. Reviewer notes Step 5 Preview restricts manual actions during demo, covers every intermediate state and log-cap tests — more complete. DeepSeek's strength: concise implementation meeting explicit requirements with less code.

Case 3: CSV Order Analysis Workbench from Scratch

Scope: import order CSV, detect errors, filter/analyze, export current results. Both start in empty directories with Vite + TypeScript, same 18 valid orders and error sample, no backend. Tricky data: customer names contain commas, double quotes, newlines; payment/refund/cancel orders counted separately; amounts in integer cents; error file must not overwrite imported data. Pagination 10 rows/page, but export includes all filtered records and must be re-importable.

Step 5 Preview Output

UI: file ops at top, then filters, four metric cards, two charts, order table. Chinese column headers, date range inclusive, export shows record count toast. Time: 1 hour 26 minutes . 56 automated tests passed , type-check and build pass, system Chrome verification of import, pagination, combined filtering, error recovery, chart linkage, export of 12 records and re-import.

DeepSeek V4 Pro Output

Similar layout: filter area, metric cards, charts, order table. Daily net income bars show values directly, positive/negative vs zero axis easy to read. Table headers remain English ( date, customer, amount, etc.) despite Chinese UI requirement. Screenshot shows no date filter applied: all 18 records, payment ¥1,438, refund ¥458.50, net ¥979.50 — matches ground truth. Six daily net values: 199, 348, -141, 218, 178, 177.50; chart labels correspond.

Time: 13 minutes 25 seconds , 72 tool calls . 49 unit tests + 34 browser end-to-end checks passed , type-check and build pass. Coverage includes error import preserving old data, 10/8 pagination, cross-page export of 12 records, re-import consistency. Notes: model cannot view screenshots, so no pixel visual diff; browser tests rely on DOM assertions and visibility; local Chrome required, documented in README.

Blind Evaluation (GLM)

Anonymized: A=DeepSeek, B=Step 5 Preview. GLM first verified CSV byte-for-byte identity, then ran 32 core logic assertions and 80 browser operations on both, plus each project's own tests/builds. Core logic: 32/32 both. Browser checks: 79/80 both. Key behaviors (payment/refund stats, cross-page export, error import retention) all pass.

Issues found: DeepSeek resets filters on new import but leaves stale keyword "Zhang San" in search box; Step 5 Preview only shows this briefly when import triggered while input focused, clears on blur. Step 5 Preview advantages: duplicate ID error cites conflicting row; empty-state chart shows "no data"; bar chart keyboard operable; 1440×900 fits without page scroll. DeepSeek requires page scroll but buttons visible, table independently scrollable — still meets layout spec. DeepSeek delivers reproducible end-to-end scripts (34/34 re-run pass); Step 5 Preview has more unit tests but no equally reproducible browser scripts. Test count ≠ verification reproducibility.

Blind evaluation conclusion: Step 5 Preview slightly better, gap very small . Both complete main data flows; differences center on state consistency, error localization, empty-state messaging, and whether browser verification can be reproduced with the project.

Conclusion

Across three cases, Step 5 Preview delivers overall better results: Case 1 richer character/environment details; Case 2 six-step guide embedded in page; Case 3 clearer Chinese headers and filter/stat/export explanations. Trade-off: longer generation times — DeepSeek 9m49s, 11m6s, 13m25s vs Step 5 Preview 30m3s, 29m5s, 1h26m. Cost advantage remains: per-task cost 35% of GLM-5.3/Kimi K3, 12.5% of Claude Opus 5 — high value for money, worth trying.