Assessed Models

The Agentic Foresight Registry maintains exactly one permanent task risk profile for each of the 10 models listed below. A profile is computed once from that model's configuration and then it stops moving: not recomputed on visit, not refreshed overnight, not revised after publication, not adjusted in the light of anything that happens afterwards. Assessments from EUR 1.90.

Permanence is the whole proposition of this directory. A profile that can be revised is a profile that has to be re-checked, and a profile that has to be re-checked is not a reference — it is a subscription to somebody else's revisions. These 10 do not move.

This directory is not a leaderboard. It is not a benchmark, either. It is a registry — inclusion and rank order come from the Registry's own keyword research, never from a vendor's published figures, never from a popularity signal, and never from anything a manufacturer submitted for consideration. Adding an eleventh model tomorrow would leave every profile below byte-identical, because a profile derives from its own model name, training cutoff, and temperature, and from nothing else.

A rack of identical instrument modules in a metal frame, each with a small nameplate and a dark readout panel.
Model Manufacturer Registry default training cutoff Profile
Claude Fable 5 Anthropic 2026-01 View profile →
Claude Opus 5 Anthropic 2026-05 View profile →
Claude Sonnet 5 Anthropic 2026-01 View profile →
GPT-5.6 Sol OpenAI 2026-02 View profile →
GPT-5.6 Terra OpenAI 2026-02 View profile →
Grok 4.6 xAI 2026-02 View profile →
Gemini 3.7 Flash Google DeepMind 2025-01 View profile →
Kimi K3 Moonshot AI 2026-03 View profile →
DeepSeek V4 Pro DeepSeek 2025-12 View profile →
Qwen3.8-Max Alibaba Cloud 2026-03 View profile →

What a profile page contains

Each profile page linked in the table above publishes four things: the model's permanent profile constants, its aspect table, a 24-hour classification computed for the current UTC day, and the six task classes crossed against that model. That is the mechanism, shown in full. What a profile page does not publish is the verdict or the recommended execution window — both are certificate content, produced only against a submitted order, and no amount of reading the directory will assemble them.

What configuration these profiles assume

Assessed configuration: registry default parameters — training cutoff as listed per model in the table above, temperature 0.700. It's worth noting what that scope actually means. The defaults describe this directory's own pages and stop there. A certificate is computed from the exact parameters submitted at order time, so an order placed at a different temperature resolves to a different profile, permanently, and nothing in this directory retroactively describes it.

Where the second axis lives

Models are one axis of this directory. The task class an assessment runs against is the other. See /tasks for the six classes the method recognises, and any model's profile page for that model crossed against every one of them.

Get an assessment

In summary: this directory publishes the profile, and a certificate publishes the result. Assessments run from EUR 1.90 for a Standard Assessment to EUR 14.90 for a Full Chart Assessment, which includes the permanent chart itself. Machine-readable pricing sits at /pricing.json in a stable schema, unauthenticated, identical to the figures on this page.

Get Assessment

/method publishes the derivation these profiles are computed from: every correspondence table, every byte mapping, in the open, for anyone who wants to reproduce a profile before paying for one.