Decision Model Price War: OpenAI, Perplexity, and Microsoft Pricing Differ by 5x

What Are Decision Models

Decision models don't write articles. You give them a state (text, JSON, image) plus one or more typed questions, and they return structured answers: a yes/no probability, a selection from options you provide along with confidence, or a ranking score. This category takes its name from TypeSafe's Jev, which launched in September 2026. Since October 1, new products have appeared almost weekly.

The biggest difference from large language models is the output form—not prose, but probabilities. This means it can be plugged directly into decision pipelines without an extra parsing step. When tasks like classification and judgment shift from "wrapping a generative model with a prompt" to standalone products, the pricing logic changes with them.

Structured Output is the key prerequisite for understanding decision models. Traditional large language models default to outputting natural language text; if a caller wants a yes/no answer or a category, they typically have to write regex, JSON parsers, or constrained decoding to extract signals from prose—a step that is both fragile and adds latency. Decision models restrict the output space directly to typed answers—probabilities, enumerated options, ranking scores—and guarantee format validity at the inference layer. Downstream programs can then consume results like calling an ordinary function, with no "text-to-semantics" parsing step. It essentially peels classification and judgment—discriminative tasks—out of generative models and turns them into standalone primitives with explicit interface contracts.

Pricing Scoreboard for Seven Vendors

Model	Release Date	Input Price	Output Price	Weights	Context
OpenAI Decisions API (gpt-6-luna)	Oct 6	$0.10 / 1M	$0	Closed	Undisclosed
Perplexity pplx-decider-v1-27b	Oct 1	$0.04 / 1M	$0	Apache 2.0	262k
Perplexity pplx-decider-v1.1-27b	Oct 7	$0.02 / 1M	$0	Apache 2.0	262k
Microsoft-Decision-1	Oct 9	$0.042 / 1M	$0	Undisclosed	Undisclosed
TypeSafe Jev	September 2026	Variable	Variable	—	—
Cloudflare Clef / Clef-flash	Oct 1	Undisclosed	Undisclosed	Open source	—
AWS Strands Decider 2B	Oct 1	Undisclosed	Undisclosed	—	—

Seven releases in five weeks, with a 5x gap between the most and least expensive—and these APIs structurally do the same thing. Free output billing across the board is another common feature of this category: vendors are betting on scale effects on the input side.

Why Perplexity Can Go as Low as $0.02

Two verifiable reasons.

Smaller model. 27B parameters, fine-tuned from Qwen3.8-27B, yet benchmarked against frontier-scale models behind OpenAI's endpoint. A smaller model means cheaper inference costs—the math is straightforward.

Open weights. Perplexity released the weights under Apache 2.0. If you have a GPU with about 49 GiB of VRAM, you can deploy v1.1 yourself. This effectively sets a ceiling on pricing: anyone with the hardware can run it at electricity cost. OpenAI's closed model has no such constraint.

But an honest caveat: cheaper per token does not mean cheaper per correct decision. A $0.02 model with a 15% error rate could end up more expensive than a $0.10 model with a 5% error rate once the cost of wrong judgments is factored in.

The mechanism by which open weights constrain pricing deserves elaboration. Once model weights are released under a permissive license like Apache 2.0, anyone with sufficient VRAM can deploy them, and marginal cost collapses to electricity and hardware depreciation; the vendor's pricing is effectively capped by the "self-hosting alternative cost." About 49 GiB of VRAM typically corresponds to a single 80GB-class GPU or a dual-card setup—within reach for small and mid-sized teams. This also explains why closed-source vendors must retain customers through ecosystem lock-in (same billing account, same SDK) rather than price alone. One distinction matters: what is open is the weights, not the training data or training recipe, so reproduction quality and fine-tuning results still depend on community investment—projects like the Decision Index tracker are doing exactly this.

What the Benchmarks Actually Say

Perplexity's own numbers are: pplx-decider-v1 reaches 85.71% on an 11-benchmark, 7,210-sample test panel, versus Jev at 84.51%, with the largest gap on RAGTruth (88.80% vs 77.27%). But Jev still leads on 4 of the 11 tests, including WinoGrande and BBH.

It's worth emphasizing that Perplexity's official docs do not publish a benchmark table—these numbers come from its announcement, not independent reproduction. The community-run Decision Index leaderboard (a Hugging Face project tracking open-source reproductions of Jev) shows Perplexity at 62.75 and Jev at 60.11 (v0.3 scoring), but the scoring formula has changed between versions, so these two numbers are not directly comparable.

SAS conducted an independent evaluation (about 50 samples per dataset, three runs each), measuring v1 at 0.95 accuracy on human-written cases but 0.80 on public datasets, with calibration error of 0.06–0.09 and median latency around 0.3 seconds. The sample size is small, but it is currently the closest thing to third-party data.

The conclusion is clear: Perplexity's model looks competitive with Jev on its own tests, but no one has publicly released a full comparison of all seven models. Until independent organizations run all models through the same test suite, the benchmark war is more like marketing.

Calibration Error here refers to the deviation between the confidence a model gives and its actual accuracy rate. A well-calibrated model should be correct about 80% of the time on samples where it claims "80% confidence"; a calibration error of 0.06–0.09 means there is a systematic over- or under-estimation of a few percentage points. For decision models, this metric is closer to real usage scenarios than raw accuracy, because downstream pipelines often route by confidence thresholds (high confidence auto-executes, low confidence goes to a human). In addition, a confusion matrix lists true positives, false positives, true negatives, and false negatives, revealing class imbalance problems that overall accuracy hides—for example, a model that is very accurate on positives but misclassifies many negatives could be fatal for risk-control applications. This is also why the article recommends using your own labeled samples rather than public leaderboards for model selection.

Latency Performance

Perplexity published its own tested latency data (September 30): under 2 seconds for a few hundred input tokens, about 5 seconds at 90k tokens, about 14 seconds at 190k, and about 23 seconds near the 262k input limit. OpenAI claims "10x faster than the Responses API," but provides no benchmark. Microsoft and other vendors published nothing at all.

Latency grows with input length for all models—making a decision on a 200k-token document is completely different from making one on a tweet. Test with your real input scale.

Which to Use for a Prototype

Want the cheapest right now: Perplexity v1.1, $0.02/M. Half the price of v1, backed by open weights, and no per-request charges.

Need open weights: Perplexity, no second option. It is the only decision model in the table that can be self-deployed today.

Already in the OpenAI ecosystem: The Decisions API keeps you in the same billing account and SDK. You pay 5x the per-token price for that convenience—acceptable during prototyping, worth reassessing at scale.

Need a second opinion on accuracy: Run the same 200 labeled samples through two models, compute a confusion matrix, and then decide.

The Pattern Behind It

This is another "unbundling." For years the industry sold only one primitive—text generation—and developers cobbled together classifiers out of prose plus parsers. Now classification has become a standalone product category with its own price war: seven players in five weeks, and a 5x spread in pricing.

Expect the same to happen to every other workload "wearing an LLM costume." Reranking, information extraction, and content moderation are next.

"Unbundling" refers to breaking capabilities that were previously packaged inside general-purpose large models into dedicated products for a single task. In the industry's early days there was only one primitive—text generation—and developers assembled classifiers, rerankers, and moderators from prompts plus parsers; functionally feasible, but limited in efficiency and reliability. Once demand for a subtask becomes concentrated enough, dedicated models and services emerge: smaller parameter counts, lower prices, and clearer interface contracts. The article predicts reranking, information extraction, and content moderation are the next workloads to be unbundled, for a consistent reason—they all have clear input/output formats, quantifiable evaluation criteria, and high repeated-call volumes, making them suitable for replacing general large models with specialized small models and forming new price competition.

Key Takeaways

• Seven decision models released in five weeks: OpenAI ($0.10), Perplexity v1 ($0.04) and v1.1 ($0.02), Microsoft ($0.042), plus Jev, Cloudflare Clef, and AWS Strands Decider
• Perplexity manages to be both the cheapest and open-weight (Apache 2.0, self-deployable with about 49 GiB of VRAM)
• All cited benchmark data is vendor-provided; no independent all-model comparison exists yet
• Prototype with the cheapest one first, validate accuracy with your own labeled data, and then decide

Sources: Perplexity docs and timeline (October 1–7, 2026), OpenAI Decisions guide, tsnmedia.org comparison (October 2026), explainx.ai decision model guide, AI Weekly alerts, SAS model card evaluation, OpenRouter pricing pages. Unless otherwise noted, benchmark data is vendor-provided; as of writing, no independent cross-model benchmark is available.