Step 5 Preview as a Cursor and Cline Backend: $1/$2.70 for a 1M-Token Context — and the Verbosity Tax Nobody Prices In

By AICoderScope Team · 13 min read
stepfunstep-5openroutercursorclinebyokpricingcost-analysislocal-llm

TL;DR: Step 5 Preview — StepFun’s 600B-total/27B-active MoE with a 1M-token context window — went live on OpenRouter October 8, 2026 at $1/$2.70 per million tokens, and hit #1 on OpenRouter’s trending list a day later. Its agentic coding scores are mid-pack (31.8% on Terminal-Bench 4.0 per Vals AI — below Claude Haiku 5.5’s 39.2% at a tenth of the input price), and it burned roughly 1.7x the median output tokens on Artificial Analysis’s index run, which inflates the real bill past the sticker. The genuinely interesting part arrives October 15: open weights.

Step 5 PreviewClaude Haiku 5.5Mistral Large 4 (preview)
Input / output per MTok$1.00 / $2.70$0.10 / $0.50 (≤100K prompt), $0.50 / $2.50 above$0.68 / $2.09 until ~Oct 20, then $1.36 / $4.18
Context / max output1M / 64K1M / 128K1M claimed (524K on the preview endpoint per Artificial Analysis)
Terminal-Bench 4.031.8% (Vals AI)39.2% (vendor)22.7% independent, 28.3% vendor
Open weightsPromised Oct 15, 2026NoPromised late Oct 2026
The catch~1.7x median output-token burn; single OpenRouter provider5x price cliff past 100K prompt tokensMid-pack scores; preview endpoint caps

Honest take: As a cloud BYOK backend, Step 5 Preview loses to Claude Haiku 5.5 on both price and measured agentic coding ability, so don’t rewire your Cline config for it today. The reason to care is dated October 15: if StepFun ships the weights as promised, this becomes one of the largest open-weight 1M-context models you can self-host — and that is worth a calendar reminder. Pay the $1/$2.70 now only if you specifically want to evaluate it before committing local hardware to it.

What is Step 5 Preview and what does it actually cost?

Step 5 Preview is StepFun’s flagship agentic model, announced September 20, 2026: a sparse mixture-of-experts design with roughly 600B total parameters and 27B active per token, a 1,000,000-token context window, 64K max output, and text-plus-image input. StepFun positions it for long-horizon agent work, with particular emphasis on software engineering and finance workflows.

Pricing, verified October 13, 2026 against the OpenRouter listing (stepfun/step-5-preview; OpenRouter also lists a dated snapshot, stepfun/step-5-preview-20261008):

Per million tokensStep 5 Preview
Input (cache miss)$1.00
Cached input read$0.05
Output (reasoning tokens billed as output)$2.70

StepFun’s own API charges the same $1.00/$0.05/$2.70 structure, so OpenRouter isn’t adding a visible markup. Two operational caveats before the math: OpenRouter currently routes Step 5 Preview through a single provider, so there’s no failover if that provider degrades, and the 1M context is flat-priced — unlike Claude Haiku 5.5, there is no threshold where the rate jumps.

That flat pricing sounds like the headline. It isn’t, because Haiku 5.5’s “expensive” tier is still cheaper: past 100K prompt tokens Haiku bills $0.50/$2.50 — half Step 5’s input rate and roughly the same output rate — and Haiku’s window is also 1M tokens. The 1M-context-without-a-cliff pitch that made sense for Mistral Large 4 a week ago doesn’t survive contact with Anthropic’s October 7 price sheet. Step 5 has to win on something else.

The verbosity tax: why the real bill runs ~1.7x the sticker

Here’s the number that should anchor any cost comparison: running its full Intelligence Index evaluation, Artificial Analysis measured Step 5 Preview emitting about 160M output tokens, against a median of roughly 85–94M across the models it tests. Reasoning tokens bill as output on this model. Same tasks, ~1.7x the billable output.

Apply that to the sticker. A workload we’ve costed before — 40 agent requests a day averaging 20K input and 1.5K output tokens, 22 workdays — looks like this on paper:

Monthly (17.6M in / 1.32M out)Step 5 PreviewHaiku 5.5 (short prompts)Mistral Large 4 (preview price)
Input$17.60$1.76$11.97
Output (sticker)$3.56$0.66$2.76
Output (at ~1.7x verbosity)~$6.06——
Realistic total~$23.70$2.42$14.73

A verbose reasoning model isn’t automatically a worse deal — if the extra thinking tokens bought a higher task completion rate, you’d pay the $23.70 happily. But the benchmark section below shows they don’t, at least not on agentic coding. When you compare per-token prices across models, multiply this one’s output column by its measured verbosity first; at 1.7x, Step 5’s effective output rate behaves like ~$4.60/M, not $2.70/M.

Throughput is the one place the measurements flatter it: Artificial Analysis clocks output speed around 83–100 tokens/second (most-cited figure 99.8 tok/s) against a ~65 tok/s median for its price band. Fast — it just spends a chunk of that speed on tokens you didn’t ask for.

How good is Step 5 Preview at coding, actually?

Mid-pack, on every number published so far — and every number needs a flag, because StepFun has not released a technical report and no independent SWE-bench Verified score existed as of October 13, 2026.

BenchmarkStep 5 PreviewSource status
Terminal-Bench 4.031.8%Vals AI, independent (33.3% per some aggregators)
DeepSWE v1.167.7%Vendor-reported
Artificial Analysis Intelligence Index44Independent

Calibration against backends we’ve already covered: Claude Haiku 5.5 posts 39.2% on Terminal-Bench 4.0 (vendor-run) at $0.10/M input, Sonnet 5.5 posts 70.6% at $2/M, and Mistral Large 4 managed 22.7% independent. On the DeepSWE v1.1 table, Step 5’s vendor-reported 67.7% would slot just below Kimi K3’s 68.0 and well under DeepSeek V4.1 Flash’s 74.2 — both open-weight rivals that already have weights you can download, which matters for the verdict below.

Read those two rows together and the shape is clear: Step 5 Preview is a legitimate frontier-adjacent agentic model, and it is not a reason to move off whatever you’re running. If a terminal-driving agent is the job, the cheap Anthropic model completes more of it for less money. If DeepSWE-style repo tasks are the job, two open-weight models already beat the vendor’s own number.

Where it might earn a slot anyway: tasks that genuinely consume the window. One million tokens is roughly a 3,000-file monorepo or two large codebases side by side. A single full-context turn costs about $0.80 in input on Step 5 — and $0.04 on a cached re-read, which is what turns “load the entire legacy codebase and interrogate it for an afternoon” from a stunt into a workflow. Haiku 5.5 can do the same trick at half the input price, but if you want a second opinion from a non-Anthropic model on a million tokens of context, the list of candidates is extremely short.

How do you set up Step 5 Preview in Cline?

Cline’s native OpenRouter provider handles it — no custom endpoint needed. Verified against Cline’s provider docs and OpenRouter’s integration guide, October 13, 2026:

  1. Create an API key at openrouter.ai/settings/keys and load credits — a saved key with zero credits fails on the first paid request, not when you save it.
  2. In Cline: gear icon → API Provider → OpenRouter, paste the key.
  3. In the model field, search for and select:
stepfun/step-5-preview

Tool calling works through this route — community testing since the October 8 listing reports Step 5 running agentic loops in Cline, Kilo Code, and OpenCode, and OpenRouter’s trending rank came mostly from exactly this traffic. There were also reports of a limited free-access window in Cline and Kilo Code around October 8–9; that came from router-affiliated blogs rather than Cline’s changelog, so check the model picker for a FREE tag rather than assuming it.

One Cline-specific note on the 1M window: context you accumulate is context you re-send, and at $1/M a packrat session gets expensive — a 40-turn session holding 300K tokens of files and tool results bills $0.30 per turn uncached. OpenRouter passes through prompt caching at $0.05/M reads, and Cline caches aggressively by default, so the realistic figure is closer to $0.015 per turn. If you’ve tuned context limits for local models before (the same discipline from our Ollama context-length fix), the habit transfers: give the agent what the task needs, not what the window allows.

How do you set up Step 5 Preview in Cursor?

Through OpenRouter’s override route — with real limitations you should read before deciding it’s worth it. Cursor has no native StepFun provider, and OpenRouter still labeled its Cursor integration Beta as of early October 2026.

The setup that OpenRouter documents:

  1. Cursor Settings → Models → toggle OpenAI API Key on, and paste your OpenRouter key into that field.
  2. Enable Override OpenAI Base URL and set it to the dedicated Cursor endpoint:
https://openrouter.ai/api/v1/cursor
  1. Add stepfun/step-5-preview as a custom model name and select it for chat/agent use.

The problem we’d flag from community reports, and the fix: using the generic https://openrouter.ai/api/v1 base URL breaks tool calls in Cursor’s agent mode — requests go through but the agent can’t edit files reliably. The /api/v1/cursor endpoint exists specifically to normalize Cursor’s request format; use it, not the generic one.

Know what the override does not do. Cursor’s own models — Composer and Tab completion — never route through a custom key, so autocomplete stays on Cursor’s backend and bills normally. Community threads also report Anthropic models throwing 422 errors while the OpenAI override is active, so expect to toggle it off when you switch back to Claude models, and BYOK in agent mode requires a paid Cursor plan. A Cursor forum moderator has stated OpenRouter isn’t an officially supported BYOK provider; it works today as documented by OpenRouter, but it’s a workaround, not a product commitment. If Step 5 becomes a daily driver for you, Cline is the lower-friction host.

Step 5 Preview vs Haiku 5.5 vs Mistral Large 4: which backend wins?

Haiku 5.5, for most readers of this site — same verdict as our Mistral Large 4 analysis last week, now with a second data point confirming the pattern: new 1M-context MoEs are launching into a price-performance zone Anthropic’s cheap model already occupies.

Your situationPickPrice per MTokWhere
Cost-efficient daily agent work, short-to-medium promptsClaude Haiku 5.5$0.10 / $0.50 (≤100K)our Haiku 5.5 breakdown
Sessions permanently above 100K tokens, EU hosting or image input requiredMistral Large 4$0.68 / $2.09 previewour Large 4 setup guide
Million-token context from a non-Anthropic lab; evaluating before self-hostingStep 5 Preview$1.00 / $2.70openrouter.ai
Open-weight agentic backend you can run or rent todayKimi K3 / DeepSeek V4.1 Flashsee breakdownour Kimi K3 analysis

The decision logic in one pass: Step 5 Preview costs 10x Haiku’s input rate and scores lower on the one independent agentic coding benchmark both have run. Flat 1M pricing was its differentiator for about 24 hours until you check Haiku’s long-prompt tier, which is still half the price. What neither Anthropic nor Mistral offers today is what StepFun has promised for Wednesday.

The October 15 open weights are the actual story

Two days after this article’s publication date, StepFun says the Step 5 weights go public. As of October 13: a Hugging Face placeholder repo exists with no weights, no model card, and no stated license. StepFun’s smaller Step 3.5 Flash line shipped under Apache 2.0, which makes a permissive license plausible — but that is an inference, not an announcement. Treat the date as a plan: Mistral Large 4’s weights are likewise still a promise (due October 31), and until files appear on Hugging Face, “open-weight” is marketing for both.

If the release lands as described, the local math gets interesting fast. A 27B-active MoE is the same activation class as the models people already serve on high-memory home-lab boxes — the weights are enormous (600B total parameters means hundreds of GB even quantized), so this is unified-memory-workstation or multi-GPU territory, not a single RTX 3090. But a self-hosted, 1M-context, DeepSWE-67-class agentic model with zero per-token cost and zero data egress is precisely the endgame the BYOK crowd has been assembling hardware for. For what current hardware can realistically serve, start with the best local coding LLM guide at runaihome.com; for the serving stack once weights land, Ollama’s current state is covered at aifoss.dev.

Our recommendation, concretely: spend $2–5 on OpenRouter credits this week to test Step 5 Preview against your own repo and your own prompts — that’s the entire point of a hosted preview — and make the real adoption decision when the weights, the license, and the first independent SWE-bench runs exist. Committing to it as a paid cloud backend today means paying a premium for a model the publisher is about to give away.

FAQ

What is the exact model ID for Step 5 Preview? stepfun/step-5-preview on OpenRouter (a dated snapshot, stepfun/step-5-preview-20261008, is also listed), and step-5-preview on StepFun’s own API. Verified October 13, 2026.

Does Step 5 Preview support tool calling for agent use? Yes — it runs agentic loops in Cline, Kilo Code, and OpenCode via OpenRouter, and StepFun markets it specifically as an agent model. In Cursor, tool calls work only through OpenRouter’s dedicated /api/v1/cursor endpoint, not the generic base URL.

Is the 1M-token context real, and what does a full-context call cost? The OpenRouter listing confirms a 1,000,000-token context with 64K max output. A single 800K-token prompt costs about $0.80 in input on a cache miss, or roughly $0.04 when read from cache at $0.05/M — caching is what makes repeated giant-context turns affordable.

Why is my Step 5 bill higher than the $2.70/M output price suggested? Reasoning tokens bill as output, and this model reasons at length: ~160M output tokens on Artificial Analysis’s index run versus a ~85–94M median. Budget output at roughly 1.7x whatever your task history says a typical model emits.

Should I use Step 5 Preview instead of Claude Haiku 5.5 in Cline? Not on current evidence. Haiku 5.5 is cheaper at every context size — including past its own 100K cliff — and scores higher on Terminal-Bench 4.0 (39.2% vendor vs 31.8% independent for Step 5). Pick Step 5 for non-Anthropic 1M-context work or to evaluate ahead of the October 15 open-weight release.

Sources

Last updated October 13, 2026. Preview pricing, routing, and the open-weight timeline change quickly; verify current state on the official pages before committing a team or hardware budget to this model.

Was this article helpful?

Know which coding tool is worth paying for

Hands-on comparisons of AI coding assistants and what each one costs to run — including the local-model path. Sent only when something changes. Unsubscribe anytime.