StepFun Step 5 Preview Deep Dive: How a 600B Sparse MoE at $1 / $2.70 per MTok Is Rewriting the September API Price War

On September 20, 2026, Chinese AI company StepFun officially released Step 5 Preview API — a 600B-parameter sparse MoE flagship with only 27B parameters active per token, native vision input, a 1M-token context window, and an Artificial Analysis Intelligence Index of 44 (tied with Kimi K3 Max).

The pricing is what sharpened the entire industry’s attention:

• Input $1.00 / MTok
• Output $2.70 / MTok
• Cache hit $0.05 / MTok (95% discount)
• Task cost $0.71 (Artificial Analysis measured average)

Using Artificial Analysis’ “comparable capability tier” median as a reference ($1.88 input, $10.00 output), Step 5 Preview’s output price is 27% of the median — and its input price is only 53% of the median.

This is an engineering victory for extreme sparse-MoE cost engineering, and a market hit upgrading the price war to “1/7 of GPT-5.6 Sol.”

This article dissects it across five dimensions:

1. Architecture — why a 4.5% active ratio is an unusually sparse “engineering bet”;
2. Pricing — the math behind $1 / $2.70 / 95% discount, and how it stacks against the September price war;
3. Capability — AA Intelligence Index 44, Terminal-Bench 4.0 33.3%, and why it earns the “frontier-tier” label;
4. Open-weights suspense — the Hugging Face repo exists but only contains .gitattributes; BF16 lands on 10/15;
5. NixAPI integration strategy — what to do today, and what to change after 10/15.

---

1. Architecture: 600B Total / 27B Active — Why 4.5% Active Ratio Is “Unusually Sparse”

Step 5 Preview uses a sparse Mixture-of-Experts architecture. StepFun’s official numbers:

Item	Value
Total parameters	600B
Active per token	27B
Active ratio	4.5%
Context window	1M tokens
Modalities	Text + native vision input
Target workloads	Software engineering + professional knowledge work + finance

A 4.5% active ratio is unusually sparse. Current mainstream MoE active ratios usually sit between 8% and 15% (GPT-OSS-120B ~ 5B/117B = 4.3% is the outlier, Mixtral 8x7B ~ 13B/47B = 28%, DeepSeek V4 Pro published numbers around 8%).

A lower active ratio means less compute per token and theoretically lower unit-inference marginal cost — this is the underlying support that lets Step 5 Preview price output at $2.70.

But a lower active ratio also means stricter demands on the router: get the expert wrong, and the output collapses. Pushing active ratio down to 4.5% means StepFun’s router must make correct expert decisions with 95%+ probability — the most demanding engineering challenge outside the MoE experts themselves.

A number from Orcarouter is worth highlighting: output speed 99.8 tokens/s, TTFT 2.96 seconds. These two numbers show that Step 5 Preview’s inference stack has been engineered to handle long agent loops — an agent loop running 40 turns of code modifications won’t stall because of inference latency.

---

3. Pricing: The Math Behind $1 / $2.70 / 95% Discount

3.1 Price comparison table

Model	Input / MTok	Output / MTok	Cache input	Notes
Step 5 Preview (StepFun)	$1.00	$2.70	$0.05	95% discount
Step 3.7 Flash (StepFun)	$0.20	$1.15	$0.04	prior flash
Step 3.5 Flash (StepFun)	$0.10	$0.30	$0.02	older flash
GPT-6 Astra (OpenAI)	$10.00	$50.00	$1.00	flagship
GPT-5.6 Sol (OpenAI)	$4.00	$20.00	$0.40	last-gen flagship
Claude Fable 5.1 (Anthropic)	$10.00	$50.00	$0.25	flagship
Gemini 3.8 Flash (Google)	$0.75	$3.75	$0.075	intro pricing
DeepSeek V4 Pro 0813 (DeepSeek)	$0.435	$0.87	$0.0036	spot
Claude Sonnet 5 (Anthropic)	$2.00	$10.00	—	mid-tier

3.2 vs the “comparable capability tier” median

Artificial Analysis sets the comparable-tier reference median at: input $1.88, output $10.00.

Dimension	Median	Step 5 Preview	Ratio
Input	$1.88	$1.00	53%
Output	$10.00	$2.70	27%
Cache hit	n/a	$0.05	(implied ~95% off)

The 27% output price means Step 5 Preview is cheaper than every model in the tier except DeepSeek V4 Pro — and DeepSeek has structural peak/off-peak advantages on absolute low-price.

3.3 The engineering meaning of 95% cache discount

Step 5 Preview’s cache-hit price is $0.05 / MTok — a 95% discount off the cache-miss $1.00. What does this mean?

• Agent loops: same code base + same system prompt repeats over 40 turns; cache hit rate can hit 80%-90%;
• RAG re-writes: same knowledge base repeats across multiple re-writes;
• Repo-scale tasks: multi-turn review of one git repo — 60%+ cumulative hit rate is normal.

Stack these scenarios up and real long-context task cost can be 50%-70% below the headline — this is Step 5 Preview’s real killer move on agent loads.

3.4 Task cost $0.71 — the arithmetic

Artificial Analysis’ “per-task cost” of $0.71 is based on a standardized Intelligence Index evaluation run. The math:

task_cost = (input_price × input_tokens + output_price × output_tokens) / task_count

One full evaluation run cost approximately $923 (back-calculation: “Step 5 Preview used 160M output tokens, output accounts for ~$432 of the total”). At ~$0.71/task, this sits below the median for its tier.

Note: 160M output tokens is far above the tier median of 92M — Step 5 Preview is a verbose model. Verbose = more thinking = higher output bucket cost. If your workload is “short answer + many turns”, Step 5’s cost advantage is more pronounced; if it’s “single long output”, its relative advantage gets partly offset by the 70% extra output tokens.

---

4. Capability: AA Intelligence Index 44 + Terminal-Bench 4.0 33.3%

4.1 Artificial Analysis Intelligence Index v4.3.2

Rank	Model	Index
1-23	GPT-6 Astra / Claude Opus 5 high / Gemini 3.8 Pro / etc.	45-58
24-27	Step 5 Preview	44
24-27	Kimi K3 Max	44
25	GLM-5.3	45
28-30	Claude Opus 5 medium reasoning	44

Step 5 Preview ties Kimi K3 Max, ties Claude Opus 5 medium reasoning, trails GLM-5.3 by 1 point.

4.2 Terminal-Bench 4.0

Model	Terminal-Bench 4.0
Step 5 Preview	33.3%
DeepSeek V4.1 Flash	26.8%
Kimi K3	12.6%

TB 4.0 33.3% in the current frontier tier clearly exceeds DeepSeek V4.1 Flash and Kimi K3 — this is the hard number supporting StepFun’s positioning as an “agent workload model.”

4.3 What this means

• Not SOTA: Intelligence Index 44 is not the frontier top. But as “one tier below SOTA at 1/3 the price,” it redraws the Pareto frontier — this is what Orcarouter’s “Advancing the Pareto Frontier” headline really means;
• Agent-friendly: TB 4.0 33.3% + 1M long context + 95% cache discount, stacked together, is a structural advantage on agent workloads;
• Verbose is a feature: Agent tool-calling inherently needs more output tokens. Step 5’s verbosity fits agent scenarios naturally — but on “short answer” tasks it costs money.

---

5. Open-Weights Suspense: HF Repo Open, BF16 Still on the Way

StepFun’s 9/20 announcement says: “The model will be released with open weights on October 15.”

But as of the announcement day, the Hugging Face repo stepfun-ai/Step-5-Preview-BF16 already exists but contains exactly one .gitattributes file — no weights, no license, no model card, no config.

This means:

1. API and weights are two separate releases — API is callable from 9/20, weights arrive in 25 days;
2. BF16 is the public-committed precision — the repo name literally reads Step-5-Preview-BF16; StepFun has anchored that bfloat16 weights (not FP32, not quantized) will ship;
3. License TBD — no license file exists yet; the most likely candidates are MIT / Apache 2.0;
4. HF naming locked early — to prevent other vendors squatting on the “Step-5” name or causing naming confusion.

Why StepFun is doing this — open weights mean “any vendor can download BF16 and infer,” ceding pricing power to the ecosystem. But StepFun simultaneously collects “closed-source inference” premiums via API. This is an “open-source + closed-source dual-track” business model — similar to Meta’s Llama strategy (open primary + API premium), opposite to Mistral’s Mixtral (closed-API only).

Implications for global developers:

• 9/20 ~ 10/14: only StepFun’s official API (or authorized aggregators like NixAPI);
• After 10/15: download BF16 and self-host — making private deployment / private cloud inference possible, especially important for data-sensitive customers (finance, healthcare, legal);
• Business-model impact: StepFun’s API revenue must compete against “open weights + self-hosted” deployment cost. Step 5 Preview API has to keep providing cheaper + better-maintained value (inference optimization, version updates, compliance certification) or get replaced.

---

6. NixAPI Perspective: What to Do Today, What to Change After 10/15

6.1 Today (9/20 ~ 10/14): use the API

step-5-preview is already on aggregate API platforms like NixAPI (following the same pattern used for Gemini 3.8 Flash, DeepSeek V4 Pro):

# Call Step 5 Preview through NixAPI
curl https://nixapi.com/v1/chat/completions \
  -H "Authorization: Bearer ***" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "stepfun/step-5-preview",
    "messages": [{"role": "user", "content": "Refactor this Python module and run tests."}]
  }'

Zero business-code changes — upgrading from stepfun/step-3.7-flash to stepfun/step-5-preview is just a field change.

6.2 After 10/15: consider self-hosting

Deployment option	Use case	Cost
StepFun official API	Medium-volume, long-context agents	$1/$2.70 + 95% cache discount
BF16 self-host (H100/H200)	High-volume, data-sensitive	GPU capex + inference ops
BF16 self-host (DeepSeek-V3 style 8×H20)	Extreme cost control	Needs deep customization + MoE inference framework
NixAPI aggregation + smart routing	Flexible model switching	Pay-as-you-go

Single-card BF16 inference of 600B MoE isn’t feasible — even with only 27B active per token, BF16 inference needs at least 4×8 H100 cluster. The realistic minimum for self-hosting is 8×H100 / 4×H200 — at H100 spot rates around $3/hr, 4 cards 24/7 is roughly $8,640/month. This only breaks even at daily token consumption >= 100M.

6.3 Immediate business-side actions

1. Run a head-to-head experiment: use the same agent benchmark (SWE-Bench Verified or Terminal-Bench 4.0) to compare Step 5 Preview vs your current default model — measure the real cost of “output verbosity”;
2. Check your cache hit rate: in your NixAPI dashboard, look at cache hit rate over the past 30 days — if it’s > 50%, Step 5 Preview’s 95% discount hits your costs hard;
3. Design a downgrade path: make Step 5 Preview the “high-load default,” and DeepSeek V4 Pro / Gemini 3.8 Flash the “short-answer downgrade” — switch dynamically by prompt length / task type;
4. Watch the 10/15 release: subscribe to NixAPI Radar (nixapi.com/blog) to get BF16 + license details first.

---

7. Practical Advice for Indie Developers

7.1 If you build agents / long-context workloads

• Use Step 5 Preview directly — AA Index 44 + TB 4.0 33.3% + 1M ctx + 95% cache discount is one of the best “agent-friendly” combos in the current frontier tier;
• Budget by output bucket — Step 5 is verbose, 160M output per eval run says it “thinks a lot.” If your workflow is “read 100K prompt → output 10K answer,” output cost $2.70 × 10K = $27/M tasks is the dominant line;
• Highest value when cache hit rate > 70% — typical scenario for long prompts + multi-turn agents.

7.2 If you build short-answer / high-frequency API calls

• Step 5 isn’t optimal — DeepSeek V4 Pro $0.435/$0.87, Step 3.7 Flash $0.20/$1.15, Step 3.5 Flash $0.10/$0.30 are all cheaper “short answer” options;
• Step 5 fits “heavy reasoning + short input” — e.g., “read 50K legal text → answer yes/no + explain” type agents;
• Keep Step 3.7 Flash / DeepSeek V4 Pro as default; only switch to Step 5 when “must be smarter.”

7.3 If you’re cost-sensitive and want open-source / self-host

• Wait for 10/15 — BF16 weights + MIT/Apache license + 600B/27B MoE combo equals “you build your own inference infrastructure” beginning;
• Assess whether you can run it — BF16 600B needs at least 4×H200 / 8×H100 / 16×A100-80G cluster;
• First validate self-host with Step 3.5 Flash / 3.7 Flash — they’re already open-source + small (~50B / ~100B); Step 5 is the post-10/15 target.

7.4 Three things you can do this week regardless of stack

1. Add a model field to your product config layer — even if you use Step 3.7 Flash / DeepSeek V4 Pro today, future switching to stepfun/step-5-preview is a string change;
2. Track the 10/15 open-weights release — subscribe to NixAPI Radar (nixapi.com/blog) to get BF16 + license + deployment guide first;
3. Re-examine whether your “agent loop” depends on low IQ — many “agent stalls” are really base-model IQ issues. Step 5 Preview’s Index 44 may be the key to unlocking your current stall scenarios.

---

8. Conclusion

Step 5 Preview is one of the most important model releases of September 2026 — not because its capability is strongest (Index 44, not SOTA), but because it uses extreme 4.5% active-ratio sparse-MoE engineering to push frontier-tier marginal cost to 1/3 of the comparable-tier median.

• Architecture bias: 600B / 27B = 4.5% active, sparser than mainstream MoE;
• Price punch: output $2.70 is 27% of comparable-tier median, 1/7 of GPT-5.6 Sol;
• Agent-friendly: TB 4.0 33.3% + 1M ctx + 95% cache discount — currently one of the best agent combos;
• Open-source suspense: HF repo open but only .gitattributes; BF16 on 10/15 — the wait itself is a business signal.

For global developers, 9/20 ~ 10/14 is the API dividend window — use aggregators like NixAPI at $1/$2.70/95% discount to run business; after 10/15 evaluate self-hosting, fold “open BF16” into your cost structure.

Step 5’s story is just starting. We will keep updating the 10/15 weights release, license details, benchmark performance, and deployment strategy for global developers on NixAPI Radar.

---

Sources

• StepFun official announcement — stepfun.com/step-5-preview
• Artificial Analysis model page — artificialanalysis.ai/models/step-5
• Artificial Analysis provider comparison — artificialanalysis.ai/models/step-5/providers
• Orcarouter technical teardown — orcarouter.ai/blog/step-5-preview-open-weights
• mlllm.io coverage — mlllm.io/articles/3240-stepfun-releases-step-5-preview-reasoning-model
• Runtimewire coverage — runtimewire.com/article/stepfun-step-5-preview-600b-agent-model-pricing
• AI Weekly brief — aiweekly.co/alerts/stepfun-ships-step-5-preview-api
• Eyestech Systems Lab analysis — eyestech.in/stepfun-step-5-preview-pareto-frontier-moe-analysis
• NixAPI supported models & pricing · NixAPI API docs · NixAPI console