Step 5 Preview vs GPT-5.6: One Flat Card Against a Ladder of Six Tiers

There is no single model called GPT-5.6 to put next to Step 5 Preview, and that is the first thing this comparison has to settle. the lab shipped the GPT-5.6 family on July 9, 2026 as three tiers — GPT-5.6 Sol at $4.00 and $20.00 per million tokens, GPT-5.6 Terra at $2.00 and $12.00, and GPT-5.6 Luna at $0.20 and $1.20 — each with a six-position reasoning dial that runs from none to max. Step 5 Preview, which StepFun released on September 20, 2026, is one model at $1.00 and $2.70 with no tiers and three reasoning settings. The two vendors have built opposite cost-control mechanisms: one vendor sells you a ladder you climb per request, and StepFun sells you one flat card. Which is cheaper depends entirely on how good your routing is, and that is not a number either vendor publishes. None of that is hypothetical: 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.

The headline comparison most pages run is Step 5 Preview against GPT-5.6 Sol, because Sol is the flagship and flagships get compared. That framing is misleading in both directions. Sol costs seven and a half times what Step 5 Preview charges per output token, and it scores three points higher on the independent index. But Sol is not the tier most Step 5 Preview buyers would actually switch from — GPT-5.6 Terra is, at $2.00 and $12.00, and Luna is what a high-volume pipeline would weigh instead. Pick the wrong tier to compare against and the conclusion flips without any of the underlying facts changing.

The three tiers against one card

• Input price — Step 5 Preview $1.00 per million tokens vs GPT-5.6 Sol $4.00, GPT-5.6 Terra $2.00, GPT-5.6 Luna $0.20

• Output price — Step 5 Preview $2.70 vs GPT-5.6 Sol $20.00, GPT-5.6 Terra $12.00, GPT-5.6 Luna $1.20

• Long-input repricing — Step 5 Preview one flat rate across the whole 1M-token window vs all three GPT-5.6 tiers stepping up at 272,000 input tokens: Sol to $8.00 and $30.00, Terra to $4.00 and $18.00, Luna to $0.40 and $1.80

• Cache — Step 5 Preview $0.10 per million on a cache hit, a 95% discount vs the vendor's published cache rates, which are applied per tier and which the vendor has revised with each of the family's two repricings

• Context window — Step 5 Preview 1M tokens vs GPT-5.6 tiers at roughly 1.05M tokens, effectively identical

• Maximum output — Step 5 Preview 64k tokens vs 128,000 tokens on all three GPT-5.6 tiers, twice StepFun's ceiling

• Modality — Step 5 Preview text, images and video in, text out vs GPT-5.6 tiers text, images and files in, text out; the trade is video in one direction and native file input in the other

• Reasoning control — Step 5 Preview three settings, low, medium and high vs six on GPT-5.6: none, low, medium, high, xhigh and max

• Independent score — Step 5 Preview 44 vs GPT-5.6 Sol 47 at max effort and 44 at xhigh; GPT-5.6 Luna 37 at maximum effort; GPT-5.6 Terra is not carried on the same index revision

• Cost per completed index task — Step 5 Preview about $0.71 vs GPT-5.6 Luna $0.18 and GPT-5.6 Sol at multiples of StepFun's figure, since Sol is priced at five times Step 5 Preview's output rate at the same measured configuration

• Weights — Step 5 Preview downloadable in FP8, BF16 and, since October 14, 2026, an NVFP4 build vs no weights for any GPT-5.6 tier

The row worth staring at is not the price. It is the effort dial. Step 5 Preview's 44 is one scored configuration at one effort setting, which makes it a floor and a ceiling at the same time. GPT-5.6 Sol measures 47 at max and 44 at xhigh — the same headline number as StepFun's model at a setting that costs materially less to run. That is not a coincidence worth building a narrative on, but it is a real feature of the comparison: the same vendor will sell you the same score Step 5 Preview publishes, at a lower effort level, for more money per token than StepFun charges at any setting.

The 272,000-token line is the structural difference

Both of these models are sold as long-context systems and one of them charges for it in a way the other does not. Step 5 Preview's rate is flat across the entire million-token window; a request at 900,000 tokens costs the same per token as a request at 900. Every GPT-5.6 tier changes rate once a single request passes 272,000 input tokens, and the change applies to the whole request rather than to the portion above the line: Sol goes from $4.00 and $20.00 to $8.00 and $30.00, Terra from $2.00 and $12.00 to $4.00 and $18.00, Luna from $0.20 and $1.20 to $0.40 and $1.80.

StepFun's model has a 64k output ceiling where every GPT-5.6 tier allows 128,000, so the two products are shaped differently at both ends. If your workload reads a lot at once — a repository, a document set, a long transcript — Step 5 Preview's flat rate is the single most valuable line on the card, and it is worth more than the four-dollar input price difference on any prompt that clears the line by a wide margin. If your workload writes a lot at once — a full-file generation, a long structured artefact — the 64k ceiling is the binding constraint and cheap tokens do not help you.

This is also where a routing layer earns its keep rather than being a slogan. A pipeline that sends a 400,000-token retrieval prompt to Sol pays the doubled rate on every call forever, and the fix is not a cheaper model — it is not sending that request to a tier that reprices. The same key that reaches Step 5 Preview is not available to us, but the shape of the fix is generic: keep the long-context path on a model with a flat window and keep the short, hard reasoning calls on the model with the effort dial, and measure both against cost per completed task rather than cost per million tokens.

Why the GPT-5.6 family is a strange thing to buy in October

The vendor's own pricing has moved twice underneath this family since launch, and it has moved in one direction. The two cheaper tiers were cut on July 30, 2026; the flagship followed on August 21, taking Sol from $5.00 and $30.00 to $4.00 and $20.00 — a promotional rate the vendor commits to only through November 21, 2026. Then on September 22, the lab released GPT-6 Sol at $2.00 and $10.00 and GPT-6 Luna at $0.10 and $0.50, which is exactly half the GPT-5.6 flagship on both meters and half of GPT-5.6 Luna on both meters.

The consequence is not that GPT-5.6 is bad. It is that GPT-5.6 is no longer the vintage of the vendor's own line that a new integration should default to, and every tier of it now has a cheaper sibling with a later release date and a higher number on it. That matters to this specific comparison because it changes what "Step 5 Preview vs GPT-5.6" is really asking. If the honest answer for a Sol workload is GPT-6 Sol at half the rate, then the matchup that decides anything is Step 5 Preview at $1.00 and $2.70 against GPT-6 Sol at $2.00 and $10.00 — a fourfold output spread rather than a seven-and-a-half-fold one — and the GPT-5.6 tiers are only in the conversation for teams with pinned integrations, published evaluation baselines, or an existing commitment to a specific model identifier.

There is a second reason the family question is live. Sol's promotional rate expires on November 21 and the vendor has not said what happens on November 22. Any budget built on $4.00 and $20.00 carries an unpriced event inside it, and the honest way to treat that is as a range rather than a number. StepFun has not repriced Step 5 Preview since release, which is a shorter commitment than it sounds given the model is four weeks old, but it is the commitment you have.

The tier that compares to Step 5 Preview is not the flagship

Set aside Sol and the matchup gets more interesting, because GPT-5.6 Terra at $2.00 and $12.00 is the tier a Step 5 Preview buyer is most likely to be replacing and GPT-5.6 Luna at $0.20 and $1.20 is the tier a high-volume pipeline should be measuring against. These are different arguments.

Against Terra, Step 5 Preview is half the price on input and a little over four times cheaper on output, and it wins context behaviour outright because Terra reprices above 272,000 tokens and StepFun's model does not. What Terra offers in exchange is twice the output ceiling, a larger context window by a margin that does not matter, native file input for pipelines that ingest PDFs as documents rather than as images, and a family of six effort levels that lets you steer cost per call rather than accept it. the vendor has not published a current independent index figure for Terra on the same revision as Step 5 Preview's 44, which means any claim that one is more capable than the other on Terra's side is currently unsupported — and this article is not going to invent one.

Against Luna the arithmetic reverses and Step 5 Preview loses on price by a wide margin: $0.20 and $1.20 against $1.00 and $2.70 is five times cheaper on input and a little over twice on output, and Luna's cost per completed index task of $0.18 against StepFun's $0.71 is a fourfold gap on the number that folds verbosity in. Luna scores 37 against Step 5 Preview's 44, and the evaluator notes that Luna's score is effort-dependent while StepFun's is a single configuration — a reminder that a model with a dial has many scores and quoting the max as "what it costs" is a category error. If your task is classification, extraction or routing, the tier to compare against StepFun's model is Luna, and the answer is that StepFun is not competitive on cost at all.

What actually decides it

Step 5 Preview is the right pick when three or more of these hold: the workload reads very long inputs and would cross the 272,000-token line regularly, video input is needed, a flat single-column rate card is worth more than a routing layer to your team, or downloadable weights matter because you intend to self-host or fine-tune. The weights point is not a footnote here — StepFun publishes FP8 and BF16 checkpoints under its own licence with an NVFP4 build added on October 14, 2026 configured for Blackwell at 394 GB with an unquantized KV cache, and the repositories gained a full vLLM and SGLang serving guide in mid-October with a published container image and dedicated reasoning and tool-call parsers. No GPT-5.6 tier can be downloaded, pinned or inspected, and no amount of index score closes that.

A GPT-5.6 tier is the right pick when you need a 128,000-token single response, when file input is in your pipeline, when an existing integration is pinned to a GPT-5.6 identifier and the migration cost exceeds the token savings, or when the reasoning dial is the thing you actually want — the ability to spend less on an easy request without changing models. That last one is a genuine design advantage and it is the reason to look at the family seriously rather than dismissing it as last quarter's vintage. What it is not is a reason to pick Sol over Terra or Luna for workload reasons. Route by difficulty and the family's own ladder does the cost control for you.

Both of the tiers this article compares are reachable through one endpoint on OrcaRouter: GPT-5.6 Sol, GPT-5.6 Terra and GPT-5.6 Luna sit in the catalogue at openai/gpt-5.6-sol, openai/gpt-5.6-terra and openai/gpt-5.6-luna, each at the vendor's own list price with 0% markup, which for this family has a specific consequence — the vendor has cut these rates twice since July, and a pass-through means a cut is live on your bill the day it lands rather than at the next contract cycle. Because the routing DSL composes several models into one call, the tier question above stops being an architecture decision: the same key that holds Sol at max for hard work can hold Luna for the easy majority, with automatic failover between them and no second integration to maintain. Step 5 Preview is not in that catalogue; call it through StepFun's own platform or self-host it from the published checkpoints.

The open question this comparison leaves is not which model is better. It is whether the vendor's own cheaper vintage makes the whole GPT-5.6 question moot for a new integration, and whether Sol's promotional rate survives November 21. Neither of those has an answer today. What is answerable today is the structural one, and it does not depend on either: one of these products charges you a flat rate for a million tokens and the other charges you a different rate once you pass 272,000. For a workload that lives in the long end of that window, that single line is worth more than the four-dollar difference on input, and it is the row to check against your own trace before you read any index score.