Pricing

Included on every plan, including the free one

Feature-gated pricing forces an evaluation decision before you have the information to make it. You pick a tier, build against what it exposes, discover in production that the capability you actually needed lives two tiers up, and rewrite. That wastes your quarter and usually costs us the account anyway. Better that you find out on day one.

• Retrieval modes
• Story grouping
• Exact filters
• Signal ranges
• Entity resolution
• Place filters
• Event classes
• Relationships
• Time series
• Article and story lookup
• Publisher profiles
• Provenance
• Watches and webhooks
• Language coverage
• MCP server
• Credits and previews

How to estimate what you will spend

Every call costs credits from one published table, and its price is known before it runs: add estimate=true to any request, or estimate to any tool call, and the answer is its credit range, with nothing run and nothing charged. Every response states what the call cost as credits_charged.

Estimate from distinct queries rather than from users. A monitoring workload issues a predictable number of scheduled calls regardless of how many people read the output, so its spend is easy to forecast a year ahead. An interactive workload scales with sessions instead, and the usual surprise there is retries and ranking: an agent that reformulates a hybrid query three times before it answers has spent 9 credits, and that multiplier is invisible in testing because you were watching one query at a time. Passing max_credits on each call caps it.

Watches spend credits the same way. Each run costs its query's credits, 1 for a plain poll, whether or not it finds anything: a 15-minute watch is about 2,900 credits in a 30-day month, an hourly one 720, and each webhook delivery attempt adds 1. Two hourly watches on the free plan would need 1,440 credits a month against an allocation of 1,000, and they pause once it is spent. Checking usage with account_status is free.

Credits past your plan's allocation are billed per 1,000 rather than calls failing closed: $1.00 on Build, $0.60 on Scale and $0.40 on Archive, charged in whole thousands. Each is at or above that tier's own included rate, so bursting is never cheaper than buying the allocation it exceeds. An alerting pipeline that stops at its allocation is worse than useless on the one day something actually happens. If you would rather it stopped, set a hard ceiling: a call that could cost more than the credits left is then refused before it runs.

What the rest of the market charges

Only some vendors publish prices at all. These are the published figures, captured from their own pricing pages on 2026-07-29 — where a row says pricing is not published, that is not an omission on our part.

Prices change. If a figure here disagrees with a vendor's own page, theirs is correct and this table is stale — tell us and we will re-measure.

Why we publish the numbers

Quote-only pricing costs you a week to discover something you could have read in ten seconds. It also increasingly means being left out of the decision entirely: when someone asks an assistant to compare news APIs, the vendors with a public price table are the ones that appear in the answer, and the ones without are described as "contact sales" and skipped. Better to be compared on the number than absent from the comparison.

Beyond the published tiers, pricing depends mostly on archive depth and sustained query volume, and the contact form routes those to a person rather than into a sequence.