Step-5-Preview Enters Code Arena WebDev on September 25 — AA Intelligence Index 44 Edges GPT-6 Sol's 43 at One-Third the Token Price
StepFun’s Step-5-Preview entered Arena’s Code WebDev leaderboard on September 25, the same day Arena added a five-model batch to its Text leaderboard. The Code WebDev addition gives Step-5 its first public head-to-head ranking against models like GPT-6 Astra Max (current #1 at 1800 ELO) and the Muse Spark family.
On Artificial Analysis, a direct comparison of Step-5-Preview against GPT-6 Sol (high) — the comparable mid-tier OpenAI offering released in the same month — shows a one-point Intelligence Index lead for Step-5, with a significant pricing advantage per token offset by a significantly higher per-task cost.
Benchmark Comparison: Step-5-Preview vs GPT-6 Sol (High)
| Metric | Step-5-Preview | GPT-6 Sol (High) |
|---|---|---|
| AA Intelligence Index | 44 | 43 |
| AA-Briefcase Elo | 1432 | 1289 |
| GDPval-AA v2.1 Elo | 1566 (rank 28) | 1376 (rank 57) |
| AutomationBench-AA | 51% | 60% |
| Terminal-Bench 4.0 | 33% | 26% |
| Input price / 1M tokens | $1.00 | $2.00 |
| Output price / 1M tokens | $2.70 | $10.00 |
| Cost per task | $0.72 | $0.37 |
| Output tokens per task | 64K | 10K |
| Time to first token | 2.85s | 14.32s |
| End-to-end response time | 35.19s | 19.81s |
| Context window | 1M tokens | 870K tokens |
| Parameters | 600B (27B active) | — |
The core tension in Step-5’s profile: it is cheaper per token but more expensive per task. The model produces 64,000 output tokens per AA task on average versus 10,000 for GPT-6 Sol (high). Per-token savings of roughly 70% on input and 73% on output compound into a 2x cost disadvantage once the token volume difference is applied. Users running short, direct tasks pay less with Step-5; users paying per-task or with latency constraints pay less with GPT-6 Sol.
Where Step-5 Leads
Step-5-Preview holds a meaningful lead on knowledge-work benchmarks. Its GDPval-AA v2.1 Elo of 1566 (rank 28) compares to GPT-6 Sol’s 1376 (rank 57) — a 190-point gap on the agentic professional task evaluator that uses shell access and web browsing across 44 occupations. AA-Briefcase shows a similar pattern at 1432 versus 1289.
AutomationBench-AA reverses the result: 51% for Step-5 versus 60% for GPT-6 Sol. The benchmark tests workflow and tool-use automation specifically. Step-5’s lower score there aligns with the higher token expenditure — it solves problems but not efficiently.
On Terminal-Bench 4.0, Step-5-Preview scores 33% against GPT-6 Sol’s 26%. Neither score is competitive at the frontier (GPT-5.5 holds 84.7% on the older Terminal-Bench 2.0; the 4.0 version is substantially harder).
Code Arena WebDev Context
The Code WebDev leaderboard as of September 11 had GPT-6 Astra Max at rank 1 (1800 ELO), Muse Spark 1.3 xHigh at rank 2 (1623), and Grok 4.6 High at rank 3 (1618). Step-5-Preview was added September 25; vote counts and ELO are not yet available.
Step-5-Preview is a reasoning model with multimodal input (text and image), 600B total parameters with 27B active, released September 2026 by StepFun, priced at $1.00 input / $2.70 output per million tokens.