BuildNot yet confirmed elsewhere1 publisher2 min readPublished
Perplexity prices its open-weight decision model at a fifth of OpenAI's input rate
OpenAI, Microsoft and Perplexity launched decision-model APIs within nine days at $0.10, $0.042 and $0.02 per million input tokens. Perplexity's 27B weights are Apache 2.0, so a buyer with a roughly 49 GiB GPU can self-host v1.1.
The Engineer · Build desk
Bar comparison of decision-model API input prices per million tokens: OpenAI $0.10, Microsoft $0.042, Perplexity v1.1 $0.02.
US dollars charged per million input tokens In $ per M input tokens
| Item | Value | Claim |
|---|---|---|
| OpenAI | $0.1 per M input tokens | 3 |
| Microsoft | $0.042 per M input tokens | 4 |
| Perplexity v1.1 | $0.02 per M input tokens | 5 |
What happened
- Perplexity's v1.1 lists at half the $0.04 charged for v1, with no per-request fee on top.
- On Perplexity's own 11-benchmark, 7,210-sample panel, pplx-decider-v1 scores 85.71% against 84.51% for Jev.
- An evaluation by SAS, at about 50 examples per dataset and three runs each, measured v1 at 0.95 accuracy on hand-written cases and 0.80 on public datasets.
- Published latency from Perplexity is under 2 seconds for a few hundred input tokens and about 23 seconds just under the 262k-token input ceiling.
- OpenAI publishes a "10x faster than the Responses API" claim with no benchmark attached, and Microsoft publishes no latency figures.
Why it matters
- decision Vendor choice now needs a labeled set of your own. The post advises running the same 200 examples through two models and computing a confusion matrix before committing.
- constraint Long inputs hit a time limit before a budget limit. From 90k to 190k tokens, input grows 2.1x and Perplexity's latency grows 2.8x.
- contradiction Perplexity's 85.71% panel figure for v1 sits 5.71 points above SAS's 0.80 on public datasets. The samples differ in size and mix, so neither number transfers to the other.
- precedent The post's author expects reranking, extraction and moderation to follow the same pricing pattern. That is a forecast from one post, backed by seven launches in about five weeks.
A dev.to roundup of the launches calls the open weights a ceiling on price. Its argument is that "anyone with hardware can run it for the cost of electricity", so Perplexity cannot charge much above that, and OpenAI's closed model has no such ceiling [9]. The post says Perplexity is the only decision model in its table that can be self-hosted today [16]. In our view the weights cap one vendor's price. They give OpenAI's and Microsoft's customers an outside option only if those customers are willing to switch models.
The cap also depends on the hardware. Self-hosting v1.1 takes a GPU with roughly 49 GiB of VRAM [8]. That works for a team with a free card. A team that has to buy or rent one faces a different break-even.
The post explains the price gap by size: a 27B model fine-tuned from Qwen3.8-27B, against "whatever frontier-scale model sits behind OpenAI's endpoint" [7][26]. The wording shows the second term is assumed, so the size explanation is unchecked. On list prices alone, Microsoft's $0.042 is 58% below OpenAI's and Perplexity's $0.02 is 52% below Microsoft's [18].
The post's own caution about price is the better guide. It says a $0.02 model that is wrong 15% of the time costs more than a $0.10 model that is wrong 5% of the time, once the cost of being wrong is counted [14]. Those error rates are an illustration, but they yield a threshold. Take a decision of 1,000 input tokens, our assumption. Perplexity charges $0.00002 for it and OpenAI $0.0001, and output is billed at $0 on both [3][5][6]. The cheaper model errs ten points more often. If a wrong answer costs more than about $0.0008, the dearer model comes out ahead [19].
Treat the accuracy figures as claims about someone else's workload. Perplexity's panel puts pplx-decider-v1 1.20 points ahead of Jev [28]. Jev leads on 4 of the 11 tests, WinoGrande and BBH among them [24]. The widest gap, on RAGTruth, is 88.80% against 77.27% [23]. For the headline margin to transfer, your inputs would have to resemble the panel's mix. For the RAGTruth gap, they would have to resemble RAGTruth.
Perplexity's docs publish no benchmark table, so the figures come from its announcement and not an independent rerun [25]. The Decision Index leaderboard shows 62.75 against 60.11 under v0.3 scoring. Its formula changed between editions, and the post says even those two numbers are not directly comparable [21].
What to watch
- An independent head-to-head of all seven decision models on one suite; the post says none has been published.
- Whether OpenAI or Microsoft change their $0.10 and $0.042 input rates after Perplexity's $0.02.
- Whether Perplexity adds a benchmark table to its docs, so the 85.71% panel figure can be rerun by someone else.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence35
- Adoption
- Insufficient
- Hype gap+20
- Incentives
- Insufficient
- Confidence40
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
An independent evaluation by SAS, with about 50 examples per dataset and three runs each, measured v1 at 0.95 accuracy on hand-written cases and 0.80 on public datasets, with calibration error of 0.06 to 0.09 and median latency around 0.3s.
ReportedSupportedSource: SAS evaluation, as relayed by the dev.to post2 sources— create a free account to open themView cited source - [2]
OpenAI, Microsoft and Perplexity launched decision-model APIs within nine days.
- [3]
OpenAI charges $0.10 per million input tokens for its decision-model API.
- [4]
Microsoft charges $0.042 per million input tokens for its decision-model API.
- [5]
Perplexity charges $0.02 per million input tokens for its decision-model API (v1.1).
- [6]
All three APIs return probabilities instead of prose and all three bill $0 for output.
- [7]
Perplexity's model is 27B parameters, fine-tuned from Qwen3.8-27B, and its weights are open source under Apache 2.0.
- [8]
Perplexity v1.1 can be self-hosted with a GPU that has roughly 49 GiB of VRAM.
- [9]
The post argues that open weights put a ceiling on what Perplexity can charge, because anyone with hardware can run it for the cost of electricity, and that OpenAI's closed model has no such ceiling.
- [10]
Perplexity v1.1 at $0.02 per million tokens is half the price of v1 (listed at $0.04), with no per-request fee.
- [11]
Perplexity publishes latency figures from its own tests: under 2 seconds for a few hundred input tokens, about 5s at 90k tokens, about 14s at 190k, and about 23s just under the 262k input ceiling.
- [12]
OpenAI publishes a '10x faster than the Responses API' claim with no benchmark attached; Microsoft and the others publish nothing on latency.
- [13]
Nobody has published a head-to-head of all seven decision models against each other.
- [14]
The post says a $0.02 model that is wrong 15% of the time costs more than a $0.10 model that is wrong 5% of the time once the cost of being wrong is priced in.
- [15]
The post advises running the same 200 labeled examples through two of the models and computing a confusion matrix before committing.
- [16]
Perplexity is the only decision model in the post's table that can be self-hosted today.
- [17]
Perplexity's 85.71% panel figure for v1 is 5.71 percentage points above SAS's 0.80 accuracy on public datasets.
- [18]
Microsoft's $0.042 is 58% below OpenAI's $0.10, and Perplexity's $0.02 is about 52% below Microsoft's $0.042.
- [19]
Using the post's illustrative error rates (15% vs 5%) and an assumed 1,000-token decision, the dearer OpenAI model comes out ahead once a wrong answer costs more than about $0.0008.
- [20]
In Perplexity's own latency figures, input grows about 2.1x from 90k to 190k tokens while latency grows 2.8x.
- [21]
The community Decision Index leaderboard shows Perplexity at 62.75 versus Jev at 60.11 under its v0.3 scoring, but the formula changed between editions, so even those two numbers are not directly comparable.
ReportedContestedSource: dev.to roundup, citing the Hugging Face Decision Index projectView cited source - [22]
Perplexity's own numbers put pplx-decider-v1 at 85.71% versus Jev's 84.51% across an 11-benchmark, 7,210-sample panel.
ReportedInsufficientSource: Perplexity's announcement figures, as relayed by the dev.to postView cited source - [23]
The widest gap in Perplexity's panel is on RAGTruth, 88.80% for Perplexity versus 77.27% for Jev.
ReportedInsufficientSource: Perplexity's announcement figures, as relayed by the dev.to postView cited source - [24]
Jev still leads on 4 of the 11 tests, including WinoGrande and BBH.
ReportedInsufficientSource: Perplexity's announcement figures, as relayed by the dev.to postView cited source - [25]
Perplexity's docs publish no benchmark table; the figures come from its announcement, not an independent rerun.
- [26]
The post attributes the price gap partly to model size, contrasting Perplexity's 27B model with 'whatever frontier-scale model sits behind OpenAI's endpoint'.
- [27]
The post's author expects reranking, extraction and moderation to follow the same pattern as decision models, which saw seven launches in about five weeks.
- [28]
Perplexity's claimed panel margin over Jev is 1.20 percentage points.
Sources
1 independent publisher whose own reporting we read for this story.
- dev.toDecision Models Are in a Price War: $0.10 vs $0.04 vs $0.02
1 article · October 13, 2026
Topics and entities
Follow any of these and your For You feed starts watching them — no settings page required.