The biggest trap in this launch is a tense: Step 5 Preview is available, but its open weights are still a promise.
Step 5 Preview launched on September 20, 2026 as StepFun’s new flagship model for agentic work. StepFun describes a sparse mixture-of-experts system with 600 billion total parameters, 27 billion activated for each token, vision input and a one-million-token context window. The company says the model is available through its products and API. It separately says the weights will be published on October 15.
What ships now and what comes later
| Capability | Status on September 20 | What to verify |
|---|---|---|
| Product and API access | Available, according to StepFun | Account eligibility, model ID, rate limits and regions |
| Text and vision input | Documented | Supported image formats and image-token accounting |
| 1M-token context | Documented | Retrieval accuracy and cost at production lengths |
| Open weights | Promised for October 15 | License, files, precision and deployment requirements after release |
That distinction matters for developers planning self-hosting. An announced publication date does not reveal the eventual license, weight format, quantizations, tokenizer files or memory requirements. Until the artifacts are public, describe Step 5 Preview as an API model with a future open-weights commitment.
The 600B headline hides the active compute
A sparse model can hold many parameters while activating only a subset for one token. StepFun reports 600B total parameters and 27B active per token. That does not make the system equivalent to a dense 27B model: routing, expert diversity, memory movement and serving architecture still affect quality and latency. It does mean the total-parameter number alone is a poor predictor of API speed or self-hosting cost.
The API economics are unusually easy to test
Vercel’s AI Gateway lists Step 5 Preview at $1 per million input tokens, $2.70 per million output tokens and $0.05 per million cached input tokens. Prices can change, and a gateway price is not a universal price. Still, those units make a representative workload calculation possible.
- Measure one short coding task, one repository-sized task and one vision task.
- Record uncached and cached input separately.
- Count tool results returned to the model, not only the original prompt.
- Cap maximum output so a failed agent cannot generate indefinitely.
- Compare cost per accepted task with the GLM-5.3-Flash long-context profile.
A one-million-token window needs positional testing
Context capacity is not the same as reliable recall. Place the answer-bearing evidence near the beginning, middle and end of a long input. Add distracting files and conflicting instructions. Repeat the test with images. The result should show whether the model can retrieve, cite and act on evidence across the window rather than merely accept a large request.
Treat the benchmark table as a vendor claim
StepFun publishes results for coding, finance and agentic workflows, plus an expert review in which it says roughly 70% of evaluators judged the model capable of autonomously solving moderately high-complexity coding tasks. Those results are useful leads, not independent proof. The prompt set, harness, tool permissions, retry policy and judge model can change an agent benchmark materially.
A practical Step 5 Preview evaluation
- Pin the exact preview model ID and record the date.
- Run a fixed set of repository tasks with no hidden retries.
- Measure accepted patches, not completed generations.
- Test image reading with small text, diagrams and contradictory captions.
- Exercise tool failures, permission denials and timeouts.
- Compare the same workload with Atria Dawn Preview and another stable model.
- Re-run the deployment review only after weights, license and files actually appear.
What the October weight release must answer
The eventual files need a model card, license, tokenizer, configuration, checksums and clear precision information. Operators will also need expert-parallel serving guidance, tested inference frameworks and realistic memory requirements. A downloadable archive without those pieces would not establish practical openness. Check whether commercial use, modification and redistribution are permitted, whether vision components are included and whether the published checkpoint matches the hosted preview. Reproduce a small set of API results locally before assuming behavioral parity.
The practical verdict
Step 5 Preview is a substantial same-day model release because it combines sparse scale, vision, a one-million-token context and accessible API pricing. The most consequential part for independent deployment is still pending. Teams can evaluate the hosted model now, but any self-hosting conclusion should wait for the October artifacts and their license.
Primary sources
Checked September 20, 2026. Pricing, access and the promised weights can change; verify the current artifacts before deployment.
