Mercury 3 vs Step 5 Preview: The Measured Model Is the One That Got More Expensive

On October 14, 2026, StepFun doubled the cache-hit input price of Step 5 Preview from 0.35 to 0.70 yuan per million tokens, with a grandfather clause for callers who had already used the model. Five days later, Inception launched Mercury 3 at half its list price and did not say when the discount ends. Put those two facts next to each other and you have the shape of this comparison: the model with an independent benchmark score is the one whose price went up, and the model on sale is the one nobody outside its lab has measured. The documented AI failures and agent incidents behind releases like this one are collected in the OrcaRouter AI Incident Archive where each entry carries its own sources, severity, confidence grade and disputed flag.

That inversion is not a coincidence, and it is not an argument for either side. Step 5 Preview is a settled model with a growing public record — a vendor release on September 20, 2026, an independent Intelligence Index score, official weights on Hugging Face — and vendors charge more for things that are known to work. Mercury 3 is nine days old, has a claimed 1,068 tokens per second, and has no independent score of any kind. The price you pay for the cheap one is that you are the measurement.

Two previews that mean different things

The word "preview" sits in both names and does different work in each. On Step 5 Preview it is StepFun's own label and part of the API identifier — the string you send is step-5-preview, not step-5 — and StepFun has published no replacement date, no deprecation notice and no stability commitment for that string. In practice the model has been live for a month, has weights on Hugging Face, and has a published rate card, so "preview" is doing less work than it looks like it is.

On Mercury 3 there is no preview label at all: this is a general-availability launch with an OpenAI-compatible endpoint, a model ID of mercury-3, and a 100-million-token trial for new accounts. What Mercury 3 has instead is a launch discount, which is a different kind of instability. A preview label warns you that the model may change; a launch discount warns you that the price will.

Dates are worth getting right before anything else. Step 5 Preview's vendor release date is September 20, 2026, taken from StepFun's own announcement materials; Artificial Analysis lists the model as released in September 2026 and its FAQ text dates it to September 18. The two-day gap is unexplained by either party, and this piece uses StepFun's date where it needs one. Mercury 3's date is unambiguous: October 19, 2026, one day before this was written.

The spec sheet, with the source of every line

• Price per million tokens — Mercury 3 $0.20 in and $0.70 out at list, $0.10 and $0.35 during the launch discount vs Step 5 Preview 7 yuan in and 20 yuan out, with cache-hit input at 0.70 yuan after the October 14 change

• Evaluator's dollar figures — Step 5 Preview $1.00 in and $2.70 out per million with a 90 per cent cache discount, as published by Artificial Analysis; these are the evaluator's own numbers and not a currency conversion of StepFun's rate card

• Context window — Mercury 3 1,000,000 tokens vs Step 5 Preview 1M tokens

• Maximum output — Mercury 3 131,072 tokens vs Step 5 Preview 64,000 tokens

• Input modality — Mercury 3 text only vs Step 5 Preview text, images and video

• Output modality — text on both

• Parameters — Mercury 3 undisclosed, described only as more than 40 times Mercury 1 vs Step 5 Preview not published on StepFun's platform documentation or on the evaluator's model page, so neither number can be cited here

• Weights — Mercury 3 closed vs Step 5 Preview published on Hugging Face under the StepFun Model License 1.0, in the base repository plus BF16 and NVFP4 builds

• Reasoning control — Mercury 3 four effort settings from instant to high, defaulting to low vs Step 5 Preview a reasoning model with no published effort parameter

• Independent Intelligence Index — Mercury 3 no entry as of October 20, 2026 vs Step 5 Preview 44 on index revision 4.3.2

• Measured output speed — Mercury 3 none vs Step 5 Preview 92.2 tokens per second, measured by Artificial Analysis

• Prompt caching — Mercury 3 supported at $0.02 per million cached tokens at list vs Step 5 Preview supported at 0.70 yuan per million cache hits, up from 0.35 on October 14

One line in that list deserves a second look because it is the only place either model's price has moved in public. Step 5 Preview's cache-hit rate doubled twelve days into its life. The cache-miss input price and the output price were untouched, and callers who had used the model through the pay-as-you-go API before that moment kept the old rate under a grandfather clause. If you are building a cost model on prompt caching — and any agent workload is — a rate that has already doubled once is a rate to keep checking.

The multimodal line is the one that ends some conversations

Step 5 Preview accepts text, images and video and returns text. Mercury 3 accepts text and returns text. Inception's docs are not ambiguous about this: the model is listed as text-only, and the launch material is about speed and cost, not about vision.

For a workload that reads screenshots, scans invoices, watches screen recordings or parses a video into structured notes, there is nothing to compare. Step 5 Preview is the only one of the two that can be given the input at all, which makes the price gap irrelevant and the benchmark gap irrelevant right up until you decide whether the vision work matters. That decision is usually made by the product, not by the model.

For a workload that is text in and text out — search, support, code review, summarisation — the modality line disappears and the comparison becomes about speed, price and evidence, in that order of how easily each can be checked.

What the only independent score in this matchup says

Step 5 Preview scores 44 on the Artificial Analysis Intelligence Index, version 4.3.2 — fourth of 121 models in that index's own table, against a median of 19. The same listing records a measured output speed of 92.2 tokens per second, a time to first token of 2.55 seconds, and a verbosity figure of 160 million output tokens against a 140-million median, which the evaluator describes as somewhat verbose. It calls the model somewhat expensive relative to comparable open-weight models, which is a fair reading of $1.00/$2.70.

Read the 44 carefully. It is a weighted composite across a benchmark suite, not a claim about your workload, and it does not mean Step 5 Preview beats the model ranked 42nd on your specific task. The evaluator's own page also carries "Under review" markers against several individual evaluations. But it is a number produced by a methodology somebody else can inspect, and it puts the model above the median of the field.

Mercury 3 has nothing comparable. Inception's quality claim is comparative — "on par with or better than frontier speed-optimized models like GPT-6 Luna and Claude Haiku 5.5" — with no margin, no test names and no indication of which of its four effort settings produced the comparison. The one structured result is Inception's own agent test: 50 live Mind2Web navigation tasks, 50 per cent completed, "significantly more than GPT-6.1-Sol", at the same token volume and step count and therefore a fraction of the per-token cost. Fifty tasks in the vendor's own harness is a signal, not a score.

The physical number, and what it is worth

1,068 tokens per second is the reason to read about Mercury 3 at all. If it holds, a 500-token answer generates in under half a second, which is the difference between a search answer that feels instant and one that feels like a network round trip. Step 5 Preview at a measured 92.2 tokens per second takes about 5.4 seconds for the same answer, plus roughly 2.55 seconds to first token.

The claim comes with no hardware, no batch size, no effort setting and no statement about whether it is single-stream or aggregate, and all four of those choices move a throughput figure by multiples. The vendor-to-independent spread in this family is documented: Artificial Analysis measures Mercury 2.5 at about 706 tokens per second, below the number Inception published at that launch. That is the discount to apply to 1,068 until somebody reproduces it. There is no Mercury 3 page on Artificial Analysis as of October 20.

Step 5 Preview's speed, by contrast, is boring in the way measured things are boring. It is 92.2 tokens per second, measured, and that puts it 32nd of 78 models on the evaluator's speed ranking — faster than typical, not remarkable. Nobody is buying it for speed. They are buying it for the 1M-token window, the multimodal input and the score.

Costing the same job on both rate cards

Take a document triage pipeline: 8,000 pages a day, 2,500 tokens of OCR'd text per page, a 300-token classification and extraction output per page. That is 20 million input tokens and 2.4 million output tokens daily, with the same instruction block resent on every call and therefore cacheable.

• Mercury 3 at list — 20 × $0.20 plus 2.4 × $0.70, or $5.68 a day

• Mercury 3 at the launch rate — about $2.84 a day

• Step 5 Preview, uncached — 20 × 7 plus 2.4 × 20 yuan, or 188 yuan a day; at the evaluator's dollar figures, 20 × $1.00 plus 2.4 × $2.70, or $26.48

• Step 5 Preview with the instruction block cached — the cache-hit line is 0.70 yuan per million against 7 yuan for a miss, so the saving depends entirely on how much of that 20 million is the repeated preamble rather than new page text

So the sticker gap is roughly four and a half times in Mercury 3's favour, and the launch discount widens it to about nine. That is the strongest single argument for Mercury 3 and it is entirely computable. What is not computable is cost per completed job, because Mercury 3 has no published completion rate and no published output-token count. Step 5 Preview's verbosity is at least measured — 160 million tokens on the index against a 140-million median, which the evaluator calls somewhat verbose — and that lets you budget for it.

The one caution worth repeating: the Mercury 3 figure to plan against is the list price. The 50 per cent cut is described as "at launch" with no end date, and its two predecessor models are now closed to new users with existing access ending November 15, 2026. Inception's commercial behaviour suggests you should model the rate that survives the promotion.

Weights, licences, and what changed in the last month

Step 5 Preview's open-weight story changed after launch. At release the weights were mirrored by third parties rather than published by the vendor; as of this week StepFun's own Hugging Face organisation carries the base repository plus BF16 and NVFP4 builds, all last modified on October 15, 2026. The licence is the StepFun Model License 1.0, which is not Apache or MIT and needs reading before you ship — it is a vendor licence with terms rather than a permissive one.

Mercury 3 has no weights and no licence to read. It is API-only, with enterprise deployments adding dedicated capacity, autoscaling and configurable data retention rather than an artefact you can hold.

For a team deciding whether to fine-tune, quantise or run on their own hardware, that line is dispositive and the other lines do not matter. For a team calling an API and never touching a checkpoint, it matters only as a measure of exit cost: if StepFun changes its price or retires the model, a Step 5 Preview deployment can move to self-hosted weights; a Mercury 3 deployment moves to another vendor.

The routing question, answered honestly

Neither model is on the OrcaRouter catalogue. Our model API returns nothing for step-5-preview or for any Inception model, so anyone who wants to call either one goes to the vendor's own platform or to one of the third-party hosts each vendor lists. Saying that plainly is more useful than implying a route we do not have.

What the routing layer does give you in this situation is a way to answer the question without committing. More than 200 models sit behind one OpenAI-compatible endpoint at provider list price with no markup, and the ones closest to these two on price and capability — Qwen 3.8-Max at $2.00/$6.00 with a 1M-token window, GLM 5.2 at $1.40/$4.40, Kimi K3 at $3.00/$15.00 — can be measured against your own traffic this afternoon. If none of them clears the bar, the case for paying either vendor strengthens; if one of them does, you have saved yourself an integration. Automatic failover also covers the specific risk in this matchup, which is that the model on a launch discount with no stability commitment is the one you would be putting in the production path.

One mechanical point that matters more here than usual: provider list prices pass through unchanged, so a vendor rate change reaches the bill the same day rather than at renewal. Given that Step 5 Preview's cache price has already moved once, that is not a hypothetical advantage.

The question to ask before you pick

Both of these models are cheap. The difference in their bills, on a realistic pipeline, is a few dollars a day, and the launch discount makes Mercury 3 look cheaper than it will be after the promotion ends. So cost is not where this decision lives.

It lives in a simpler question: do you need the answer to be fast, or do you need to know whether it is right? Step 5 Preview is the one you can check — an independent index score, measured speed, measured verbosity, published weights you can run in your own harness. Mercury 3 is the one that will answer in under a second and cannot currently be checked by anyone outside Inception.

For a latency-bound product with a human in the loop and a recoverable failure, the second trade is reasonable and the evaluation is cheap. For a pipeline that runs unattended and whose errors are expensive, a model with a measured score in the mid-forties is worth several times the price of a model with no score at all — and that, rather than any benchmark, is why the measured one can afford to raise its cache price.

The thing to watch is unglamorous: the next independent Mercury 3 reading, whenever it appears, and whether StepFun's cache price moves again. Both are small facts that would change the arithmetic more than any launch post will.