Step 5 Preview vs Claude Fable 5: The Flagship That Was Superseded and Never Repriced

The most useful thing to know about Claude Fable 5 in October 2026 is that the vendor still sells it at the same price as the model that replaced it. Step 5 Preview, the 1M-context sparse mixture-of-experts flagship StepFun released on September 20, 2026, costs $1.00 per million input tokens and $2.70 per million output. Claude Fable 5, released on June 9, 2026 as the lab's first Fable-class model, costs $10.00 and $50.00 — an eighteen-fold spread on the output meter — and it is charged that rate alongside Claude Fable 5.1, which sits on the same card at the same headline price with a point-and-a-bit more index score, a 75% cheaper cache read, and three more months of fixes. So the interesting comparison is not really Step 5 Preview against Claude Fable 5. It is what each vendor's price list tells you about which of its own models it wants you to call. Benchmarks age and prices move; the failure record does not: the documented AI incidents and agent failures behind models in this class are collected in the OrcaRouter AI Incident Archive where each entry carries its own sources, severity, confidence grade and disputed flag.

The answer is in the cache line. On Claude Fable 5 a cache read costs $1.00 per million tokens. On Claude Fable 5.1 it costs $0.25. Nothing else on the Claude Fable rate card moved between the two releases — $10 in, $50 out, unchanged — which means the only cost lever the vendor chose to pull was the one that matters to the workload a Fable-class model is sold for: a long agent loop re-reading its own context. Keeping Claude Fable 5 on the same $10 / $50 with a four-times-pricier cache read is not an accident of a forgotten price page; it is the mechanism by which the older model is left on the shelf while remaining technically available. This article prices both of them properly, and states plainly where Step 5 Preview is the wrong answer.

Both rate cards, interpreted rather than listed

• Input — Step 5 Preview $1.00 per million tokens vs Claude Fable 5 $10.00; ten times, and StepFun's number is a flat rate across the whole 1M-token window while the vendor's is also flat at this size

• Output — $2.70 vs $50.00; eighteen and a half times, and the meter that decides an agent's bill

• Cache read — Step 5 Preview states a 95% input discount, putting a re-read token at $0.10 vs Claude Fable 5 at $1.00, which is 10% of its own input price rather than 2.5% like its successor

• Cache write — Step 5 Preview does not publish a separate write rate vs Claude Fable 5 at $12.50 per million for a five-minute window and $20.00 for an hour, charged on top of the input rate

• Context and output — 1M tokens of context for both; 64k maximum output for Step 5 Preview against 128k for Claude Fable 5, the one limit where the vendor's older model is the larger of the two

• Input and output modality — Step 5 Preview takes text, images and video and returns text vs Claude Fable 5, which takes text and images and returns text

• Weights — Step 5 Preview ships downloadable checkpoints under StepFun's own licence, in FP8 and BF16 since release day and in an NVFP4 build published October 14, 2026 vs Claude Fable 5, which publishes no weights in any format

• Independent score — Step 5 Preview 44 on Artificial Analysis's Intelligence Index vs Claude Fable 5 at 50 on the same revision, with Claude Fable 5.1 at 53.35 and Claude Opus 5 at 51

• Cost per completed index task — Step 5 Preview roughly $0.71 vs Claude Fable 5 $8.75, and $7.63 for Claude Fable 5.1

Read the last two rows together, because they are where a reader is most likely to be misled. Claude Fable 5 scores higher than Step 5 Preview and costs more than twelve times as much to finish the same evaluation. Claude Fable 5 scores lower than its own successor and costs more to finish the same evaluation, which is the row the vendor's price page declines to print.

The two vendors are selling different things, and only one of them is tokens

StepFun's pitch for Step 5 Preview is a compute-efficiency pitch. At 27 billion active parameters out of 600 billion total, the model spends a fraction of the arithmetic per token that a dense model of its class would, and the independent measurements show it: Artificial Analysis records the model generating between roughly 88 and 100 output tokens per second across captures, and the host serves a million-token window at a flat rate rather than stepping the price up once a request crosses a threshold. The 5% cache-read price exists for the same reason. StepFun is optimising the number a developer sees on a per-token line, and it has a 600B flagship that behaves like a small model at inference time.

The vendor's pitch for a Fable-class model is not a token pitch and never was. The value proposition is a capability ceiling on long-horizon agentic work, and the vendor's own documentation for the family recommends reaching for it on demanding reasoning and multi-step workflows rather than as a default. That is a legitimate thing to sell and it prices accordingly, which is why the Fable tier has stayed at $10 and $50 through a release cycle while the cheaper Opus and Sonnet tiers have moved. The consequence is that a tenfold input-price gap between Claude Fable 5 and Step 5 Preview is not evidence that one of them is mispriced. It is two vendors optimising different terms.

What that means for a reader is that the choice cannot be made on the index. Claude Fable 5 measures 50 and Step 5 Preview measures 44, and a six-point gap on a composite of ten evaluations is not a statement about your task. The question is whether the twelve-times-higher cost of a completed task buys you anything your workload can detect. For a high-volume extraction, classification or summarisation pipeline the answer is trivially no. For a long-horizon agent doing research across a large repository, where a wrong turn costs an hour of wall-clock and a human review, it can be yes, and no amount of cheap tokens closes that.

The weights line is where Step 5 Preview stops being comparable

There is a category difference between these two models that no rate card captures, and it is the one that decides regulated deployments. Step 5 Preview is downloadable. The FP8 repository appeared at 03:15 UTC on the morning of release with its shards committed 36 minutes later, the BF16 repository followed, and on October 14, 2026 StepFun published an NVFP4 build at 394.18 GB quantised for NVIDIA's Blackwell hardware with an unquantized KV cache — a smaller machine than either of the earlier two checkpoints requires. StepFun's own model card carries a deployment table for the FP8 and BF16 paths, and the repositories gained a full serving guide across the middle of October: a published container image, worked tensor, expert and combined-parallelism commands, and dedicated reasoning and tool-call parsers so the model's output format survives an OpenAI-shaped API.

Claude Fable 5 has none of that and never will. It is an API-only product: no checkpoint, no quantisation choice, no self-hosting path, no way to pin a version you control or to inspect what is running. That is not a flaw in the model; it is a deliberate product boundary, and for most teams it is a feature rather than a limitation. But if a data-residency obligation, a procurement rule or an audit requirement sits anywhere in your decision, the two models are not in the same category and the comparison ends there. The clause "downloadable weights" is not a discount you can price against $10 per million tokens; it is a different answer to a different question.

It is worth being precise about the licence as well, because "open weights" is doing loose work in a lot of coverage. StepFun Model License 1.0 is a perpetual, worldwide, royalty-free grant to use, copy, distribute, modify and create derivative works from the weights, with a matching patent grant, and it is not an OSI-approved open-source licence. Two conditions attach: a commercial product or service that passes 100 million monthly active users or $20 million in revenue in a month has to display "Powered by StepFun AI" prominently and in its user agreement, and a hardware manufacturer shipping the model's capabilities in devices, or a model-as-a-service business past $20 million over any twelve months, needs a separate commercial licence. Internal use that does not expose the model to a third party is squarely inside the grant. That is a more permissive arrangement than no weights at all and a stricter one than MIT, and it is the document to read before you plan a product around the download.

Where Step 5 Preview is the wrong call

Three cases, stated without hedging. First, if your work is the kind where the top model's failure rate is the binding cost — long-horizon agentic coding, deep research, anything where you are paying a human to review the output — then a six-point index gap and a model with more effort settings is worth far more than the price difference, and Claude Fable 5 is the defensible choice despite being the vendor's superseded flagship. Second, if you want the Fable tier specifically, call Claude Fable 5.1 rather than Claude Fable 5: same headline price, better score, a cache read a quarter the price, and a longer maintenance runway. There is no workload for which the no-suffix model is the right pick on the merits, and the only reason it keeps traffic is that it is the identifier older integrations already hold. Third, if you need a 128k output ceiling — a single long generation rather than a long conversation — Claude Fable 5 is the only one of these two that offers it; Step 5 Preview caps a response at 64k.

Calling either one

Claude Fable 5 sits on OrcaRouter as anthropic/claude-fable-5, at the vendor's own $10.00 and $50.00 with the $1.00 cache read passed through unchanged, in the same catalogue as Claude Fable 5.1 at its $0.25 cache rate. That adjacency is the point: the version question above becomes a string change and a re-run of your own evals rather than a second contract, and because the pass-through is the provider's list price with no markup, the vendor's next rate move is live on our side the day it lands rather than at the next renewal. If you are testing Step 5 Preview against it, automatic failover is what keeps the experiment from becoming a dependency — route a slice at the unfamiliar endpoint behind a fallback to the model you already trust, measure cost per completed task rather than cost per million tokens, and raise the slice only if it holds.

Step 5 Preview is not in OrcaRouter's live catalogue, and the honest instruction is to call it through StepFun's own platform or to self-host it from the published checkpoints. That is not much of a limitation for this particular model, because the weights are the reason to pick it and a model you have downloaded is not a model you route.

The short version, if you are choosing today. Take Step 5 Preview for price, context, video input, throughput and downloadable weights, and accept six index points and a 64k output ceiling you will not notice unless you generate very long single responses. Take Claude Fable 5 only if you specifically want a 128k output ceiling or a vendor you already have a contract with — and if the answer is the Fable tier, take Claude Fable 5.1 instead of the model this page is nominally about. Neither of those is a compromise; they are different products with different jobs.

The last thing worth saying is about the direction of travel rather than either model. StepFun has spent October adding distribution and serving paths to a model released in September — a fourth checkpoint on the 14th, a GPU vendor's endorsement of it on the 15th, third-party API hosting on the 16th. the vendor has spent the same period leaving a superseded flagship in the catalogue at an unchanged price while its replacement gets the cheaper cache. Both are coherent strategies. Only one of them is likely to leave a developer choosing a model that its own maker no longer leads with, and it is the one that also happens to cost eighteen times more per output token.