The LLM landscape has fractured into a hyper-competitive niche: decision models. In just nine days, three major players launched APIs that return probabilities rather than prose, triggering a price war that has already driven costs down by 5x. OpenAI’s new Decisions API charges $0.10 per million input tokens, Microsoft sits in the middle at $0.042, and Perplexity has undercut both with its pplx-decider-v1.1-27b at just $0.02 per million. All three services bill $0 for output, signaling a shift where the value is in the structured inference, not the token generation.

The Economics of Small Models

Perplexity’s ability to charge $0.02 per million tokens stems from two verifiable factors: model size and licensing. The pplx-decider-v1.1-27b is a 27B parameter model fine-tuned from Qwen3.8-27b, significantly smaller than the undisclosed frontier-scale model likely powering OpenAI’s gpt-6-luna endpoint. Furthermore, Perplexity released the weights under the Apache 2.0 license, allowing developers to self-host the model on hardware with roughly 49 GiB of VRAM. This open-source approach places a hard ceiling on pricing, as anyone with local hardware can run the model for the cost of electricity, whereas OpenAI’s closed ecosystem has no such constraint.

Benchmark Hype vs. Independent Reality

While the pricing is transparent, the performance claims remain murky. Perplexity’s internal benchmarks show their model achieving 85.71% accuracy compared to TypeSafe’s Jev at 84.51% across an 11-benchmark panel. However, these figures come directly from vendor announcements, not independent reruns. The community-driven Decision Index leaderboard on Hugging Face tells a different story, showing Perplexity at 62.75 versus Jev at 60.11, though the scoring formula changed between editions, making direct comparison difficult. An independent evaluation by SAS found the model had 0.95 accuracy on hand-written cases but dropped to 0.80 on public datasets, highlighting that cheaper per token does not automatically mean cheaper per correct decision.

Latency and Infrastructure Trade-offs

Performance isn't just about accuracy; it's about speed at scale. Perplexity publishes latency figures showing responses under 2 seconds for short inputs, but this degrades significantly with context size, reaching approximately 23 seconds near the 262k token ceiling. OpenAI claims their API is "10x faster than the Responses API" but provides no benchmark data to back this up. For developers prototyping applications, the choice often comes down to infrastructure convenience versus raw cost. If you are already embedded in the OpenAI ecosystem, the $0.10 price tag buys you a unified billing account and SDK, a premium that may be worth paying at prototype scale but becomes unsustainable at high volume.

Key Takeaways

  • Perplexity offers the lowest price ($0.02/M) and the only open-source weights (Apache 2.0) among the major new entrants.
  • OpenAI charges a 5x premium ($0.10/M) for its closed gpt-6-luna model, leveraging existing developer infrastructure lock-in.
  • Independent benchmarks are scarce; current accuracy comparisons are largely vendor-provided marketing claims.
  • Self-hosting Perplexity's model requires ~49 GiB VRAM, offering a cost ceiling that closed models lack.

The Bottom Line

This is the unbundling of the LLM stack in real time. Classification, extraction, and moderation are no longer just tasks for a generalist chatbot; they are specialized products with their own supply chains. Expect the price war to continue until the market settles on a clear accuracy-per-dollar winner, or until open-weight models completely commoditize the inference layer.