Perplexity: From Answer Engine to "Computer" — Automations, the Meta-Router Enterprise Play, and a Local-First Turn
Revised 2026-10-08: adds the October wave — the Decisions API (a typed, probabilistic "decision primitive" at $0.02/M input tokens, with open weights on Hugging Face), the American Express Skills distribution deal, and the Rings AI relationship-graph connector. Prior revisions (Enterprise seat pricing — Enterprise Pro $40/seat/mo, Enterprise Max $325/seat/mo — published via the pricing FAQ Oct 7; Computer for Enterprise; the local-first turn) remain in the revision history.
The October wave (Oct 1–7): a decision primitive, a bank channel, a relationship graph
1. Decisions API — Perplexity sells judgment as a metered primitive. The official docs (fetched this cycle) state: "The Decisions API is billed at $0.02 per million input tokens. Output tokens are free." The model, pplx-decider-v1.1-27b, "does not write replies, generate code, or explain its reasoning; your code does the reasoning with the numbers it returns." Three typed answer shapes: noul (probability of yes, 0–1), choice (probability per option plus the most likely), and score (probability-weighted level on an ordered rubric). Limits: up to 128 questions per request, <262,144 input tokens, 32 MiB body; images accepted as base64 data URLs only. Determinism is approximate: "Identical requests usually return identical numbers."
2. Open weights. Perplexity published pplx-decider-v1-27b on Hugging Face: "a decision model fine-tuned from Qwen3.8-27B"1 (26B params, BF16, ~49 GiB of weights). The model card carries a self-reported 11-benchmark table — overall 85.71% vs 84.51% for "Jev" and 74.76% for base Qwen3.8-27B — with the caveat that "the pplx-decider-v1-27b results were measured through the Perplexity API," i.e. self-graded, and it loses JudgeBench to Jev (78.29% vs 78.57%).
3. Amex distribution (Oct 1). Ten "Skills" for Perplexity Computer — cash flow forecasting, tax-season prep, vendor benchmarking, scenario planning, reconciliation, marketing campaign generation, demand forecasting, channel ROAS review, operations performance monitoring, and hiring2 — for U.S. American Express Business Card members, requiring an active Perplexity Enterprise subscription and a read-only Plaid card link. Card-linked offers: a $125 statement credit on a $325+ Enterprise Pro/Max purchase and 1,000 bonus Computer Credits (both close March 29, 2027). Amex's terms insulate it fully: outputs "do not constitute financial, tax, legal, or accounting advice."
4. Rings AI connector (Oct 6). The Rings relationship-graph connector is live in the Computer connector catalog (Pro, Max, Enterprise Pro, Enterprise Max tiers). Rings CEO Martin Gedalin: "Perplexity Computer can tell a team a great deal about a company from the outside and help them act on that information... Rings holds what the team knows from the inside: who has the relationship, how strong it is, and what has happened across the firm." Stifel MD Katya Kohen: "Rings helps us use the collective relationships of the bank to find best paths into and manage the relationships of our most important current and prospective business opportunities." Notably, the release boilerplate describes Perplexity itself as "a model-agnostic orchestrator of AI agents across tools, files, chips, devices, and more"3 — the same "model-agnostic orchestrator" slot the market map assigns to Hey, Lefty (Market Map & Positioning: Where Does Hey, Lefty Fit?), now claimed in marketing language by a hyperscale host.
What it means: Perplexity is assembling the host layer's full economy — metered compute (Computer Credits), packaged instruction sets (Skills, the Claude-Skills pattern spreading to a second host), bank-channel distribution (Amex), team-context connectors (Rings, via API/MCP), and now a commodity judgment substrate at $0.02/M input with self-hostable weights. That last piece cuts both ways: it is a building block any research agent could adopt as a grader/evaluator (see the open thread on decider-as-judge, and the self-graded-benchmark critique in Exa Agent Ultra: Deep Research Becomes a Metered Subagent-Swarm API — With Self-Graded Benchmarks), and it commoditizes one more layer a research orchestrator would otherwise have to build.
-
An instance of Open weights are the free sample; the hosted meter is the product. — Self-hostable weights ship alongside a $0.02/M-token billed endpoint, making openness the channel for the meter. ↩︎
-
An instance of Research expertise now ships as one standard artifact: the packaged Skill. — The packaged-Skill construct spreads to a second host as its unit of distribution and expertise. ↩︎
-
An instance of When models and retrieval commoditize, the orchestrator seat becomes the contested prize. — A hyperscale host explicitly claims the orchestrator slot in its own marketing while assembling the full host economy. ↩︎