Step 5 Preview vs Claude Sonnet 5.5: Twelve Index Points, Ten Times the Task Bill

Step 5 Preview and Claude Sonnet 5.5 are the two models readers actually weigh against each other this month: a $1.00 / $2.70 per-million-token StepFun preview that opened on September 20, 2026, and the $2.00 / $10.00 Claude Sonnet 5.5 that shipped September 28. On the independent board they are twelve points apart — 44 against 56 on the Artificial Analysis Intelligence Index v4.3.2 — and twelve points is a real gap. It is also not the number that decides the bill. The number that decides the bill is cost per completed task, and there the same board reads $0.71 for Step 5 Preview against $7.60 for Claude Sonnet 5.5 at maximum effort. Ten and a half times the money, for twelve points. Almost every other row on the two cards is an argument about which half of that trade you can live with.

The two rate cards, row by row

Everything in this section is each vendor's own published pricing, read on October 9, 2026.

• Input — Step 5 Preview $1.00 per million tokens vs Claude Sonnet 5.5 $2.00 per million tokens

• Output — $2.70 vs $10.00 per million tokens; the row that dominates both bills

• Cached input — Step 5 Preview states a flat 95% input discount, which puts a cache hit at roughly $0.05 per million tokens vs Claude Sonnet 5.5's $0.20 cache read

• Cache write — Claude Sonnet 5.5 publishes $2.50 per million tokens for a five-minute cache write; StepFun has published no cache-write rate at all, so that half of the comparison is blank on one side

• Context window — 1M tokens on both cards

• Maximum output — Claude Sonnet 5.5 documents 128,000 tokens (and 300,000 on the Message Batches API behind a beta header); StepFun has published no output ceiling for Step 5 Preview

• Modality — both take text and images and return text only

• Reasoning control — Step 5 Preview exposes extended thinking with no published effort tiers; Claude Sonnet 5.5 exposes an effort ladder whose highest settings are exactly where its own cost per task blows up

• Weights — Step 5 Preview's BF16 checkpoint is promised for October 15, 2026 and has not landed; Claude Sonnet 5.5 is proprietary

Two of those rows do more work than the headline. The cache row is where the cheap card gets cheaper still, because a 95% discount on a dollar is a five-cent cache read and Anthropic's is twenty. And the output-ceiling row runs the other way: one vendor tells you where its edge is, the other advertises a million-token window without saying how much of it can come back out.

Where the twelve points actually come from

Both figures below are Artificial Analysis's, on Intelligence Index v4.3.2, and both are read from the model's own page rather than from a vendor table.

• Intelligence Index — Claude Sonnet 5.5 56, ranked #3 of 216 vs Step 5 Preview 44, ranked #45 of 200. Two different class sizes because the boards were rendered at different moments; the scores are the comparable part.

• Cost per index task — $7.60 vs $0.71. Anthropic's figure is for the configuration the board labels "Adaptive Reasoning, Max Effort, Default Fallback", which is the expensive end of that model's own ladder.

• Output tokens generated during the index run — 410M vs 160M, against a board median of 88M. Both models are verbose by the board's measure. One is more than four times the median, the other not quite twice it.

• Output speed — Step 5 Preview reads about 100 tokens per second against an average of 70. Artificial Analysis had not published an output-speed figure for Claude Sonnet 5.5 at the time of writing, so that row is one-sided rather than won.

Put the first two rows together and the shape of the decision appears. Twelve index points separate these models; a factor of ten and a half separates what they cost to run the same evaluation. The points are real — the ranking gap puts one in the top three models the board scores and the other in the top quarter — but a team paying $7.60 per task for a hard agentic workload is paying for deliberation, and the question is whether their tasks need it.

Vendor tables will not settle this, and here is the proof

Both labs publish a Terminal-Bench 4.0 number. Anthropic's Sonnet 5.5 announcement claims 70.6% for it, against 10.3% for the model it replaces. StepFun's own published Terminal-Bench 4.0 figure for Step 5 Preview is 33.3%. Both are vendor-run, both are unreproduced by any third party, and they are 37 points apart on a benchmark with the same name.

That gap is not evidence that one model is twice as good. It is evidence that two labs running the same-named harness on different configurations, agent scaffolds and effort settings produce numbers that cannot be subtracted from one another. This is the recurring problem with launch-day comparisons, and it is the reason every row in this piece that comes from a lab is labelled as such. Where a third party has measured both models on one harness, use that. Where it has not, treat the vendor figure as a claim about an intention.

Anthropic's own migration note from the Sonnet 5.5 launch is worth carrying into this comparison, because it applies to any preview endpoint you are evaluating. The vendor's guidance was that effort levels were recalibrated between Sonnet 5 and 5.5 and that teams should re-run their effort sweep rather than carry a setting over. The equivalent warning applies here in the other direction: Step 5 Preview is a preview, its serving configuration is still moving, and a setting tuned against this week's endpoint is not a setting you own.

Both of them talk too much, by different amounts

Verbosity is the row that turns a rate card into an invoice. Step 5 Preview generated 160 million output tokens to complete the index suite, where the board's median model generates around 92 million — roughly 1.7 times the median. Claude Sonnet 5.5 generated 410 million, or about 4.7 times the median.

At $2.70 per million output tokens, Step 5 Preview's verbosity is a large share of a small number. At $10.00, Claude Sonnet 5.5's is most of a very large one. Neither model is concise, and neither vendor's marketing describes it that way. The practical consequence is that a token-per-task measurement on your own traffic will tell you more about the real cost difference than the rate cards do, because the rate cards assume both models write the same number of words for the same answer. They do not.

The cache row, worked out on a real loop

Take an agent loop with a 40,000-token stable prefix — a system prompt plus repository context — resent on every turn for thirty turns, with the prefix written once and hit twenty-nine times. That is 1.16 million cached input tokens per session.

• Claude Sonnet 5.5 — one write at $2.50 per million tokens, roughly $0.10, then 1.16M read tokens at $0.20 per million, roughly $0.23. Total about $0.33 per session, before any fresh input or output.

• Step 5 Preview — reads at the stated 95% discount, so about $0.05 per million and roughly $0.06 per session. The write half is uncosted because StepFun has not published a cache-write rate.

On the rows we can actually price, the cached traffic costs about five times more on the Anthropic card, and the caveat is that one of the two write rates does not exist yet. This is the workload both models are sold for, which is why the row matters more than the headline discount percentages make it look.

Where each one can actually be called

Claude Sonnet 5.5 is in our catalogue as anthropic/claude-sonnet-5.5, at Claude Sonnet 5.5's own $2.00 and $10.00 with a $0.20 cache read, a 1,000,000-token context window and 128K max output — list price passed through at 0% markup, so a change on Anthropic's card is a change on ours the same day. Step 5 Preview is not in that catalogue, and we are not going to imply otherwise: it answers on StepFun's own API and, until the promised checkpoint lands, nowhere else at all.

That asymmetry has a practical use. If both models were behind one key with automatic failover, moving a workload from one to the other would be a model-string change and nothing else — which is the honest reason to keep the door generic. As it stands, the comparison is between a model you can call through your existing route and one you have to wire up separately, and the second one is a preview whose serving fleet is still being built. Putting unproven capacity behind a router with failover to something already qualified is the cheap insurance; a primary path pointed directly at a preview endpoint is not.

What would change this verdict

Three things, all datable. The first is October 15, 2026, when StepFun's BF16 checkpoint is due. If the licence permits fine-tuning and the weights are real, the twelve-point gap stops being a fixed property of the model and becomes something a team can attack on its own data — at which point the ten-times-cheaper card becomes a much more interesting starting point. If the checkpoint slips or the licence is restrictive, the gap stays exactly where the board put it.

The second is whether Anthropic's own effort recalibration lesson gets applied here. Sonnet 5.5's per-task advantage over Sonnet 5 exists at low and medium effort, not at maximum, and the reverse is true at max. A Step 5 Preview buyer who never sweeps effort is comparing a model at its default against a model at its default, which on the Anthropic side is not where the cheap, high-value work happens.

The third is a proper independent run of Step 5 Preview at a configuration comparable to the one Anthropic's number is measured at. Until that exists, the 44 is a single scored configuration with no ladder beside it, and the 56 has a whole effort curve under it. The comparison is honest as written; it is not yet symmetrical.

For a team deciding today rather than on the 15th: if the workload is long-context, cache-heavy and frequently repeated, the arithmetic on this page favours the cheap card by an order of magnitude and the capability gap is the price of admission. If the workload is hard, short, and graded by a human who will notice the difference between 44 and 56, the expensive card is the one that finishes the job. Neither answer survives contact with a workload that is both, which is what a router is for.

list price passed through at 0% markup , so a change on Anthropic's card is a change on ours the same day.

automatic failover on a router is what lets a preview endpoint be evaluated behind a qualified fallback instead of as the primary path.