Closed hosted models
Often provide strong general capability and integrated tools, but require scrutiny of data retention, contractual controls and provider dependence.
Thomas Cole, The Titan’s Goblet, 1833
ResourceThe model landscape.
The model landscape.
A dated snapshot, not a permanent ranking.
AI systems, work organisation and occupational health.
Rankings change quickly and cannot select a model on their own. This page compares capabilities, limits, cost and conditions of use without confusing novelty with occupational relevance.
In 30 seconds
Essential points
- Every comparison must state its date, provider and exact model version.
- Latency, privacy, tool access, stability and total workflow cost matter as much as a public score.
- Specialised occupational-health performance requires dedicated validation rather than inference from general benchmarks.
EstablishedEmergingAnalysis
For the detailed current table, refer to the French model landscape, which is updated more frequently.
Market
A dated snapshot
Public model leaderboards are useful for orientation but often mix different prompts, tools, dates and evaluation conditions. Fast product updates can make a ranking obsolete within weeks. Preserve the model identifier, provider, access route and evaluation date in every internal record.
Open-weight models
Can support local hosting and greater control, but shift responsibility toward infrastructure, security, maintenance and evaluation.
Specialised systems
Retrieval, rules or smaller domain models may outperform a general chatbot on a narrow, well-defined workflow.
Selection
Criteria beyond headline scores
A model selection process should include quality, critical error types, answer stability, refusal behaviour, latency, price, context handling, tool permissions, logging, geographic hosting and the ability to control model updates.
Reliability
Repeat difficult cases and inspect systematic error patterns rather than relying only on average accuracy.
Privacy and security
Assess data roles, retention, reuse for training, subprocessors, transfers and access controls.
Total cost
Include integration, human verification, monitoring, incident handling and skill maintenance, not only token price.
Domain fit
Specialised validation remains necessary
A model that performs well on broad reasoning or medical exams may still fail on French occupational-health law, institutional roles or nuanced workplace scenarios. Domain-specific tests should reflect the intended use and the severity of possible errors.
Screen broadly
Use public benchmarks to identify plausible candidates.
Test locally
Evaluate the exact workflow, data format, prompt, tools and human review process.
Monitor drift
Re-test after provider updates, prompt changes, retrieval changes or new regulations.
Key references
- See the detailed and dated French model landscape.
- NIST AI RMF and provider documentation should be used alongside local testing.