Parallel Decisions API
⚠️ Why not a model
Hosted service that retrieves live web evidence and answers typed questions; Parallel publishes no decision model, weights or card behind it.
Technical specs
- Base LLMs
- undisclosed
- Decision types
- boolean, choice, score
- Features
- hosted-api, web-grounded, sources-returned, predefined-answers, batched-questions, mcp-tool, public-docs, free-tier
- License
- Proprietary
Links
The Parallel Decisions API takes context plus named questions, searches the live web, and returns typed answers with probabilities and the sources it used. Parallel announced it on X on 27 Oct 2026 (13:02 BRT); the blog post is dated 25 Oct.
It follows the System One question shape that Jev popularized (yes/no, pick one of fixed options, score on an ordered rubric), but it is not a decision model. Parallel’s pitch is the opposite trade-off: models like Jev answer only from their input and training data, while this API looks things up first.
| Endpoint | Decisions quickstart |
| Price | $0.001 per request ($1 per 1,000); 5,000 free requests |
| Also available as | web_decisions tool in the Parallel MCP server, free with no key (rate-limited) |
| Model behind it | not disclosed |
Vendor claims (UnverifiedClaim). The blog’s example answers come back in roughly 0.5 to 0.9 seconds and it claims “10x lower cost than a standalone model”. Those are Parallel’s own demos, not a public benchmark. At $1,000 per million calls it is also far pricier than hosted decision models; one reply called it five times the most expensive one they had seen.
Fit / anti-fit
Fits questions whose answer depends on recent facts (company size, a policy’s current status, news). Does not fit hot-path routing or guardrails where milliseconds and cost per call matter; use a decision model there.
