Perplexity Releases Decider v1.3: 70 Points on the Decision Index and $68 per Million Tasks
Perplexity has released Decider v1.3, a model that decides what an agent should do next: call a tool or ask the user a clarifying question. On the Artificial Analysis decision ranking, the 27-billion-parameter version scored 70 points (sixth place), while the 4-billion-parameter Flash scored 64 (tenth place). A million tasks cost $68 and $34. For cheap agents, Flash is worth comparing with your current model.
AI-processed from X: @ArtificialAnlys; edited by Hamidun NewsAny LLM-based agent constantly faces routine decisions: call a tool or ask the user, pull a field from a document or honestly decline. For these forks, Perplexity has released separate models, Decider v1.3 and Decider Flash v1.3. The independent lab Artificial Analysis tested them on its Decision Index v2.1 ranking, which measures the quality of exactly these kinds of decisions. The larger 27-billion-parameter model scored 70 points and took sixth place. The lighter 4-billion-parameter Flash scored 64 points and took tenth place. According to the lab's estimate, running a million tasks through them costs $68 and $34, respectively.
What Changed Compared with v1.2
The larger Decider v1.3 kept the sixth spot it held with v1.2. The gain is small. On the AA-Paperclip test it ranks fourth with a score of 69, two points better than the previous version. On EvalSafe-O*NET-AA it scores 89 and ranks third, up 1 point.
The smaller model improved more noticeably. Flash v1.3 gained 4 points over v1.2 and rose to tenth place. Two tests account for all of the gain: AA-Symmetry and When2Call, where the score rose by 12 points on each.
When2Call measures a practical skill. The model must understand when to call a function and when to ask a clarifying question or acknowledge that no suitable tool exists. This is where production agents most often make unnecessary or wrong calls, so a 12-point jump for a small model is noteworthy. The source does not explain what the other tests measure; their descriptions should be sought in Artificial Analysis documentation.
Price, Size and Terms
The main argument for these models is cost. Both sit on the Pareto frontier in the capability-price coordinates. That means that within this set of tests, no model delivers the same quality for less or better quality for the same money.
The models' parameters are as follows:
- size: 27 billion parameters for Decider v1.3 and 4 billion for Flash;
- context: 256 thousand tokens, with text and image input;
- input token price: $0.02 per million for v1.3 and $0.01 for Flash;
- license: Apache 2.0;
- access: through Perplexity's own API.
Input token prices look very low, so check them against the price list in your Perplexity account before calculating your budget.
What a Developer Should Do
First. If your agent chooses its next step many times per session, there is no need to send every such decision to the largest general-purpose model. A small specialized model can perform no worse and cost several times less. The 4-billion-parameter Flash suits the role of a fast traffic controller: routing a request to the right tool or deciding whether there is enough data to answer.
Second. An Apache 2.0 license usually means free use, but the source lists only the Perplexity API as the provider. It contains no links to download the weights. It is too early to count on running the model locally; first check whether the weights have been published.
Third. The results come from one lab's tests. They are sufficient for initial screening but do not replace validation on your own data. Take 100–200 real decisions from your agent, for example routing support tickets or extracting fields from contracts. Run them through both Decider versions and through your current model. Compare not only accuracy but also the share of erroneous tool calls.
There are also limitations. Positions in the ranking reflect a snapshot at the time of publication and will shift as new models arrive. The larger model's improvements over v1.2 are modest: 1–2 points on individual tests, with the same sixth place. The real gain went mainly to Flash, which is the one to examine first for anyone trying to save on large agentic pipelines.
- Nov 21, 2026OpenAI Names 4 Reasons GPT-6 Astra Hides Its Reasoning from Monitors
- Nov 21, 2026Perplexity Launches Frontier Research, Plans to Train Its Research Agent on 10,000+ GPUs
- Nov 21, 2026Decision models answer in milliseconds: Jev costs $0.042 per million tokens
- Nov 21, 2026ChatGPT cited Ahrefs pages, but in 43% of answers it recommended other companies' conferences
Key Questions Answered
What Changed Compared with v1.2?
The larger Decider v1.3 kept the sixth spot it held with v1.2. The gain is small. On the AA-Paperclip test it ranks fourth with a score of 69, two points better than the previous version. On EvalSafe-ONET-AA it scores 89 and ranks third, up 1 point. The smaller model improved more noticeably. Flash v1.3 gained 4 points over v1.2 and rose to tenth place. Two tests account for all of the gain: AA-Symmetry and When2Call, where the score rose by 12 points on each. When2Call measures a practical skill.
What are the key details on Price, Size and Terms?
The main argument for these models is cost. Both sit on the Pareto frontier in the capability-price coordinates. That means that within this set of tests, no model delivers the same quality for less or better quality for the same money.
What a Developer Should Do?
First. If your agent chooses its next step many times per session, there is no need to send every such decision to the largest general-purpose model. A small specialized model can perform no worse and cost several times less. The 4-billion-parameter Flash suits the role of a fast traffic controller: routing a request to the right tool or deciding whether there is enough data to answer. Second. An Apache 2.0 license usually means free use, but the source lists only the Perplexity API as the provider.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.