Skip to content
Get in touch
The ModelariumAll Worlds
Choose a different job

Classify or route content

What must be sorted into useful, explainable categories?

PrivacyNon-sensitive inputsDeploymentHosted service allowedScaleOccasional to moderate volumeLatencyInteractive, not real-timeBudgetBalanced quality and costReliabilityHuman review for consequential output
Refine these assumptions in the Calibration Room

3 candidates. No automatic winner.

These are dated starting points under the disclosed defaults. Evaluate these approaches against your workload.

Quality & reliability

GPT-5.6 Terra

Inside: document media pipeline

LifecyclecurrentVerified2026-08-27Evidence1 provider source + RS interpretationStatuscurrent

A balanced hosted candidate for nuanced labels and explanations.

TradeoffHigh-volume routing needs unit-cost and abstention testing.

Why?
  • supports classify
  • accepts text
  • fits quality-reliability
Go deeper

This approach joins the release gpt-5.6-terra to the deployment offer openai-gpt-5-6-terra-api. The recommendation is dated RS interpretation, not a provider ranking.

Technical details

Modalities: text, image. Access: proprietary. Evidence class: rs-interpretation.

Evaluation plan

  1. Build a stratified labeled set.
  2. Measure per-class errors and abstentions.
  3. Review consequential false routes.
Inspect the model recordProvider source

Speed & cost

Gemini 3.7 Flash

Inside: document media pipeline

LifecyclecurrentVerified2026-08-27Evidence1 provider source + RS interpretationStatuscurrent

A high-throughput multimodal option for routing large mixed-input queues.

TradeoffEfficient volume is valuable only if class-specific errors remain acceptable.

Why?
  • supports classify
  • accepts text
  • fits speed-cost
Go deeper

This approach joins the release gemini-3.7-flash to the deployment offer google-gemini-3-7-flash-api. The recommendation is dated RS interpretation, not a provider ranking.

Technical details

Modalities: text, image, audio, video. Access: proprietary. Evidence class: rs-interpretation.

Evaluation plan

  1. Use the same labeled set.
  2. Measure throughput and per-class recall.
  3. Inspect error clusters.
Inspect the model recordProvider source

Control & locality

Qwen 3.8-27B

Inside: document media pipeline

LifecyclecurrentVerified2026-08-27Evidence1 provider source + RS interpretationStatuscurrent

A local option for routing content that cannot leave the environment.

TradeoffLocal capacity and evaluation ownership grow with queue volume.

Why?
  • supports classify
  • accepts text
  • fits control-locality
Go deeper

This approach joins the release Qwen/Qwen3.8-27B to the deployment offer qwen-3-8-27b-local. The recommendation is dated RS interpretation, not a provider ranking.

Technical details

Modalities: text, image. Access: open-weight. Evidence class: rs-interpretation.

Evaluation plan

  1. Run the labeled set locally.
  2. Compare error clusters with hosted options.
  3. Measure sustained throughput.
Inspect the model recordProvider source
See what did not enter this shortlist

No additional approaches were evaluated for this job.