Step 5 versus Grok 4.5: cheaper and downloadable, but harder to score

StepFun’s Step 5 Preview offers lower listed prices, a larger context window and downloadable weights than xAI’s Grok 4.5, while Grok has a published vendor coding score and no released weights. OrcaRouter’s comparison lays out a useful choice for developers: which trade-offs matter more than a tidy leaderboard?

xAI Watch analysis

What happened

OrcaRouter compares the two models’ pricing, availability and reported performance. It lists Step 5 Preview at $1 per million input tokens and $2.70 per million output tokens, against $2 and $6 for Grok 4.5. Step 5 also has a one-million-token context window, compared with 500,000 for Grok, and its weights are downloadable. xAI has published no Grok 4.5 weights, while its vendor-reported SWE-bench Pro score is 64.7%.

The comparison reports Artificial Analysis Intelligence Index scores of 44 for Step 5 and 39 for Grok. Those numbers are not a clean head-to-head verdict: the models were not tested at matched reasoning settings, and the evaluator marks Grok’s score as deprecated. OrcaRouter also says Step 5’s documentation lacks a numeric benchmark table, leaving its performance picture less straightforward to assess.

Why it matters

A lower rate card does not settle the bill. OrcaRouter says Grok used fewer tokens in the index suite and cost less per completed task in that evaluation, despite its higher per-token prices. For teams running long prompts, Step 5’s larger context and flat pricing may matter; for teams prioritising a published vendor coding result, Grok has a clearer figure to point to. The comparison is useful precisely because neither model wins every column.

Our read

This is a comparison of trade-offs, not a universal winner’s podium. Step 5’s weights offer control and room to inspect or self-host; Grok’s published score offers a more legible performance claim, though it comes from xAI. Before choosing, test both against your own workload and check what the benchmark actually measures. The spreadsheet can narrow the field, but it cannot do your job for you.

What to watch

• Whether StepFun publishes readable numeric benchmark results for Step 5 Preview.
• Whether independent evaluators update Grok 4.5’s deprecated score.
• How the models compare on matched prompts, settings and real workloads.

Discussion spark: When choosing a model, would you favour downloadable weights and a lower rate card, or a clearer published performance score?

not affiliated with or endorsed by xAI, SpaceXAI, SpaceX, X Corp. or Cursor