Changelog
Every release, Nov 2025 to today.
Newest first. Any finsen-3.x version remains callable indefinitely — see Models for the
deprecation policy on pre-3.0 versions.
About the names
SwishX names each model generation after a Nobel laureate in Physiology or Medicine, rotating to a new name every two or three versions: Behring (1.x), Ross (2.x), then Finsen (3.x, current).
Finsen 3.2 — Aug 6, 2026
Current. Sharper claim-grounding accuracy and lower multi-model routing latency than 3.1.
- Reduced false-positive rate on the claims-linked stage for comparative claims.
- Faster model-routing decisions when a downstream model is near capacity.
Finsen 3.1 — Jun 2026
Keyframe-first generation, reducing hallucination on rare disease states and off-label claims.
- Introduced
keyframe_idon video generation — reuse or regenerate a still without re-rendering the clip. - Meaningfully fewer off-label visual implications on longer (12–16s) generations.
Finsen 3.0 — May 2026
Full reasoning pipeline shipped: brief decoded, context harnessed, prompts layered, claims linked, model routed, output verified — MLR-ready by construction, not by a separate review pass afterward.
- First version where request/response shape is considered stable going forward.
- Output verification added as a distinct post-render stage (see Compliance & verification).
Ross 2.5 — Mar 2026
Claim Extractor and Audience Classifier split into dedicated reasoning stages, instead of being handled implicitly inside context harnessing.
- Claim-linking rejections now include which specific claim triggered them.
Ross 2.2 — Feb 2026
Compliance Verifier added — automated ISI and fair-balance checks before render, not just at claim-linking time.
- First version to reject a request for fair-balance reasons pre-generation, not just unlinked claims.
Ross 2.0 — Jan 2026
Multi-model routing across Veo, Kling, and Runway; scene-level model selection instead of a single fixed generator.
routed_modelfirst appears in the response shape.
Behring 1.5 — Dec 2025
Introduced the Context Harness — brief, dossier, and regulatory ruleset merged before generation, rather than the brief being sent to the generator on its own.
Behring 1.0 — Nov 2025
Text-to-video orchestration, single-pass prompting, no compliance layer. The first version of the API — a single generation call with no claim-linking or verification stages.