What's new

Stamp AI is now the default check: every claim is checked against screened live-web evidence in seconds, with audited citations, and its accuracy is published on the Benchmarks page: on two public benchmarks of real-world claims professional fact-checkers had settled (AVeriTeC and LIAR2), its high-confidence verdicts were right 97% of the time. The multi-provider panel becomes the Research Panel, our enterprise-grade depth; Model Fact-Check (Tier 2) is no longer offered in the app but still runs through the API. Pricing: Stamp AI is priced per claim, and the per-check fee now scales with the document (10 credits a claim, at most 100), so short texts and API calls cost far less. New opt-in verdict re-use: when you re-check a draft, unchanged claims your organization verified in the last 24 hours keep their verdict for free, labelled with when they were verified. Confidence now shows as a band (high, moderate, low) instead of a percentage, reports name the exact model version that voted, and printed certificates state plainly what a verdict is and isn't.

Privacy Policy update: model calls are still never routed to a provider that trains on your content. By default, they may now reach providers that keep prompts for a limited period solely for safety and abuse monitoring, as several of the largest providers require. Zero data retention is available to organizations under a separate agreement. Knowledge Bank embeddings now carry the same no-training guarantee as every other model call.

Verification layer: an independent system now checks the pipeline's own work on every check. It confirms that each extracted claim says what your document says, that gathered evidence is on-topic and carries no instructions aimed at the models, and that citations support the verdict they're cited for. Confidence scores now also account for whether each model's reasoning backs its vote. The model lineup is refreshed (Claude Opus 5 and Sonnet 5, Gemini 3.8 Flash, GPT-6 Astra, and new GPT-5.6 Sol and GLM 5.3 Flash options) at unchanged prices, and API estimates now flag which claims benefit from live research.

Claim extraction rebuilt: coverage now reaches the facts in appositives, asides, and subordinate clauses; every extracted claim anchors to the exact passage it came from; and extraction is much cheaper to run, so estimates stay fast.

Public report sharing: make any report public and share a read-only link. Also new About, Security, Methodology, Benchmarks, Changelog, and Blog pages, plus a public model catalog and API docs.

Organizations: invite your team, assign per-module roles and permissions, and share a pooled credit balance.

Fact-checking quality round: more resilient claim evaluations, a claim-fidelity audit on every check, and broader extraction coverage.

Terms of Service and Privacy Policy published, with data-export and account-deletion requests; full product redesign; deterministic claim extraction so identical documents get identical quotes.

New experimental Rapid Check tier: a single fast model checks each claim against the live web, at a flat price.

Panel-weighted pricing: your per-claim rate now reflects the exact models on your panel, quoted before you submit.

Trust upgrade: source-authority ratings, a citation-faithfulness audit, the 7-point verdict scale, and deep claim research.

First release: three-tier fact-checking (Spell & Grammar, Model Fact-Check, Live Research) with multi-model panels, judged consensus, real-time report streaming, and credit-based pricing.