Perplexity has released the weights of its two newest decision models, Decider v1.3 and Decider Flash v1.3, under the Apache license, and in its research post of 12 November it says Decider read 1 billion tokens across 7,000 of its own agent sessions in one day and identified flaws that led to 14 fixes to its agent security rules. OpenRouter is serving both models free until 19 November, its chief executive Alex Atallah wrote on X.
At Perplexity’s list price of $0.02 per million input tokens, reading 1 billion tokens with the larger model costs $20, or $10 with Flash, by our arithmetic. That works out to about 143,000 tokens per agent session and about 71 million tokens per rule fix.
Business Pill · Verification Cost
An evergreen Business Engineer explainer on why work that is cheap to make is not always cheap to check. It teaches the general idea only and does not discuss Perplexity.
The key insight: As we read the post, the number to repeat is the ratio: about 71 million tokens of log reading per security rule fix, by our arithmetic, at a list cost of $10 to $20 for the whole 1 billion tokens.
Decider v1.3 is built on Qwen3.8-27B, and Flash v1.3 on Qwen3.5-4B, distilled from the larger model, according to the Hugging Face model cards. The post says both are available on Hugging Face under Apache license, through OpenRouter and through the Perplexity Decisions API.
A decision model does not write text. The post says “A decision model is a general-purpose classifier.” It takes an input and a set of questions, and returns a probability for each possible answer: yes or no, one of several options, or a level on a scored scale. Version 1.3 adds locate, which pulls exact text spans out of the input along with their character positions.
Perplexity’s pricing page lists pplx-decider-v1.3 at $0.02 per million input tokens and pplx-decider-flash-v1.3 at $0.01, with output tokens free and no per-request fee. OpenRouter’s Perplexity page lists both v1.3 models as free, at $0 per million input and output tokens. Its listings for the earlier Decider V1.2 and Flash V1.2 show $0.02 and $0.01.
Cost of reading 1 billion input tokens with Perplexity’s Decider v1.3 models at the list prices on Perplexity’s pricing page ($0.02 and $0.01 per million input tokens), our arithmetic, and on OpenRouter’s free listing, which OpenRouter’s CEO Alex Atallah wrote runs until 19 November. Perplexity says its one-day security run read 1 billion tokens.
What Was Posted on X
Decision Models are the future of agentic security. They can process billions of tokens of telemetry, identify security issues, and help you patch fixes at a scale and coststructure not possible before. We’ve open sourced our models and made it free on Openrouter for a week.
Aravind Srinivas, Perplexity’s cofounder and CEO, posted this at 04:08 UTC on 13 November 2026, quoting Alex Atallah’s post below. It says Perplexity has open-sourced its decision models and made them free on OpenRouter for a week. The piece quotes it as his view on agentic security and checks the release details against Perplexity’s research post.
Neat example of how Perplexity used Decider internally
In one day it read 1B tokens across 7,000 agent sessions and found flaws that led to 14 fixes to their agent security rules.
Both models are Apache-licensed and free on OpenRouter until Nov 19
Alex Atallah, OpenRouter’s CEO and co-founder, posted this at 03:16 UTC on 13 November 2026, quoting Perplexity’s developer account. It summarises the one-day security run from Perplexity’s post (1 billion tokens, 7,000 agent sessions, 14 rule fixes) and gives 19 November as the end of the free OpenRouter window, the date the piece uses.
How Perplexity Uses It on Agent Security
The post describes the setup. Perplexity records the activity of its agents and runs rule-based checks on it in real time. Offline, Decider reads samples of those logs: where no rule fired, it checks whether the rules missed something, and where an alarm was raised, it checks whether the alarm was false.
Records where the rules failed either way go to an agent for triage, which proposes a fix to the rules. A member of Perplexity’s security team approves each fix, the post says. Of the 14 rule fixes from the one-day run, 6 were approved and shipped within two days.
Replayed on a week of logs, the post says, the fixes cut the busiest rule’s volume by 52%, and the widened rules caught “about 40% more legitimate actions to investigate”. By our arithmetic, 6 of 14 is 43% of the proposed fixes shipped, or about 167 million tokens read per shipped fix.
Perplexity reports its own evaluations. Across 63 classification tasks, Decider v1.3 averaged 91.6% accuracy, 6.54 points above Perplexity’s production classifiers. Across 72 retrieval tasks it scored a mean nDCG@10 of 83.99, 1.81 points above Qwen3-Reranker-8B.
On one million requests from the first 18 days of Perplexity’s Decision API, Decider v1.3 agreed with frontier-model judges 93.98% of the time, against 93.68% for v1.2 and 93.35% for TypeSafe AI’s Jev 1.14, the post says. When options were reordered or relabelled, v1.3 changed its answer on 0.27% of questions, against 2.69% to 4.84% for Jev 1.14.
When we read OpenRouter’s Perplexity page on 13 November, it showed 1.48 billion tokens processed on Decider V1.3 (free) and 689 million on Decider Flash V1.3 (free), both listed on 12 November, against 488 billion on Decider V1.2. In October we covered Perplexity’s Decider for Mac, which runs the 4B model on a laptop.
The Structural Read
As we read the security section, Decider is used as a second check on rules, not as the real-time guard: the rules run live, Decider reads samples of the logs offline, and a person on the security team approves each fix, the post says.
As we read the prices, the free OpenRouter week and the Apache-licensed weights lower the cost of trying the same setup, while Perplexity’s own API keeps billing $0.02 and $0.01 per million input tokens, per its pricing page.
Aravind Srinivas, Perplexity cofounder and CEO, on X, 13 November 2026 (his view)
“Decision Models are the future of agentic security.”
Three Implications
FOR SECURITY TEAMS RUNNING AGENTS Perplexity describes a loop in which a decision model checks rule hits and misses on sampled logs, an agent proposes rule fixes and a person approves them; 6 of its 14 fixes shipped within two days.
FOR THE COST OF MONITORING At list price, 1 billion input tokens cost $20 on Decider v1.3 and $10 on Flash v1.3, by our arithmetic, about 143,000 tokens per agent session in Perplexity’s run.
FOR DEVELOPERS Both v1.3 models are free on OpenRouter until 19 November, per Atallah, and the weights are on Hugging Face under Apache 2.0, per the model cards.
The framework says: “As companies move from chatbots to thousands of autonomous agentic workflows running 24/7, the line item that used to be a rounding error becomes the budget.”
As we read it, Perplexity’s run shows the monitoring side of that line item: 7,000 agent sessions produced 1 billion tokens of logs in a day, and checking them with a decision model lists at $10 to $20. The framework predates the release and does not discuss Decider.
Security rule fixes from Perplexity’s one-day Decider run, from its research post of 12 November 2026: 6 of 14 were approved and shipped within two days. The 8 not shipped in that time is our arithmetic.
What Is Not Established
The post does not say which model size ran the security job or what the run cost; the $10 and $20 figures are our arithmetic at list price. It does not describe the flaws found. The benchmark figures are Perplexity’s own, and we did not test the models.
The 19 November end of the free OpenRouter window comes from Atallah’s post; Srinivas wrote “free on Openrouter for a week”. We did not contact Perplexity or OpenRouter for this piece.
This Week In The AI Supercycle
The Great Convergence of Capital, Power, and Geopolitics
This week’s issue: for two years the AI build-out was priced as a supply story, and last week as a credit story. This week it was priced as a permission story, and the answer arrived as a cost, not a ban.
Perplexity says one day of Decider reading 1 billion tokens of its agent logs led to 14 security rule fixes, 6 of them shipped within two days, and it has released the models under the Apache license. At its own list prices that volume of reading costs $10 to $20, by our arithmetic, and on OpenRouter it is free until 19 November.
A note on sourcing. The model release, the security setup and its results, and the benchmark figures come from Perplexity’s research post of 12 November 2026, which we read in full. Prices come from Perplexity’s API pricing page and OpenRouter’s model pages, read on 13 November; base models and licence from the Hugging Face model cards. The end date of the free OpenRouter window comes from OpenRouter CEO Alex Atallah’s post on X; Aravind Srinivas’s view is quoted from his post. Per-session, per-fix and list-price cost figures are our arithmetic. Perplexity and OpenRouter were not contacted for this piece.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
