Vladimir Arustamov
stepfun opened step 5 preview to builders on oct 8: a listing on openrouter and a week of free access in partner tools. stepfun says its task cost is "substantially lower" than similarly capable models.
we checked that claim against artificial analysis, which ran one benchmark index on step 5 and on the two gpt-5.6 models either side of its score. the claim holds. the token counts show how the saving is made, and why tokens per task is the number to watch this week.
what the free week covers
step 5 preview is a 600b-parameter mixture-of-experts model with 27b active parameters per token, a 1m-token context window and image input, per stepfun. it was announced in september. on oct 8 stepfun posted that it is live on openrouter and that a week of free access is rolling out across opencode, cline, nous research and kilocode "and more". the post gives no end date.
openrouter lists $1.00 per m input tokens and $2.70 per m output tokens, with cache hits at $0.05 per m as a launch discount. stepfun says open weights follow on oct 15 and names no license.
what artificial analysis measured
artificial analysis (aa) runs each model through the same evaluations and reports a score, a cost per task and the cost of the whole run. we took the two gpt-5.6 models closest to step 5's score of 44: terra, two points lower, and sol, three points higher.
| model | score | price per m tokens, in / out | output tokens, full run | cost per task | cost of full run |
|---|---|---|---|---|---|
| step 5 preview | 44 | $1.00 / $2.70 | 156m | $1.03 | $1190 |
| gpt-5.6 terra (max) | 42 | $2.00 / $12.00 | 122m | $1.40 | $2501 |
| gpt-5.6 sol (max) | 47 | $4.00 / $20.00 | 90m | $1.99 | $3465 |
figures from aa's comparison pages, read on oct 8; the pages carry no date. (max) is the setting aa lists for the openai models.
step 5 costs 26% less per task than terra and 48% less than sol. the full run, which aa prices separately, comes out 52% and 66% lower. that supports stepfun's claim against both models. models that score lower can cost less: aa puts gpt-5.6 luna at 37 and $320 for the whole run.
where the saving comes from
price per token. step 5's output token costs $2.70 per m, against $12.00 for terra and $20.00 for sol: 4.4x and 7.4x less. its input token costs half of terra's and a quarter of sol's.
step 5 also writes more. across the whole run it produced 156m output tokens, 1.3x terra's and 1.7x sol's. the gap on the bill (1.9x per task against sol) is smaller than the gap on output price (7.4x), and the extra tokens are one reason.
what the free week should test
stepfun's own table has step 5 behind gpt-6 astra and claude opus 5 on deepswe v1.1, 67.7 against 74.1 and 74.0, with step 5 run at high effort and the others at max. the announcement says a gap to the frontier remains on the longest tasks. the lower cost comes with a model that trails the frontier on those tests.
price per token is set by the vendor. tokens per task depends on your prompts, so it is the number to count during the free week.
sources
- stepfun, step 5 preview announcement – specs, open weights date, stepfun's benchmark table
- stepfun on x, oct 8 – the free week and the openrouter listing
- openrouter on x, oct 8 – list prices and the launch cache discount
- artificial analysis: step 5 preview vs gpt-5.6 sol – scores, prices, cost per task, tokens
- artificial analysis: step 5 preview vs gpt-5.6 terra – same measures for terra
- artificial analysis: step 5 preview vs gpt-5.6 luna – the lower-scoring comparison
