Skip to main content
The HASP AI API serves models from both Anthropic and OpenAI, routed through the HASP compliance substrate. Every model runs behind the same BAA validation, PHI Guard, credit accounting, and audit chain — the provider you pick changes the underlying weights, not the compliance guarantees. Pass the HASP model identifier in the model field of your request:
The same identifiers work on every surface — native (/v1/ai/chat), Anthropic-compat (/v1/messages), and OpenAI-compat (/v1/chat/completions).

​
Available models

Models are grouped into two access tiers:
  • Standard — enabled by default. Available to any org with an active BAA and the ai:chat scope.
  • Premium — default-off. An org admin must enable the model before requests are accepted (see Premium model access).
Relative cost is the model’s cost multiplier versus Claude Sonnet 4.6 (the credit anchor at 1.0×). It is a display hint for comparing per-token spend — the exact credit rate is the token-proportional formula in Budget Controls. claude-opus-4-7 is the current Opus flagship (extended thinking); claude-opus-4-6 remains available for pinned workloads. gpt-5.3-codex is a code-optimized standard model; gpt-5.5-pro is the highest-cost model in the catalog (extended thinking) and is default-off to contain cost-surprise risk.

​
Default model

If you omit model, the request resolves to your organization’s configured default (Claude Sonnet 4.6 unless an admin has changed it). The default is always drawn from your allowed-model list, so it is never a model you have not enabled.

​
Premium model access

Premium models (claude-opus-4-6, claude-opus-4-7, gpt-5.5, gpt-5.5-pro) are not in the default allowed-model list. An org admin enables them under Settings → AI Workspace → Models in the HASP dashboard. Once enabled, any API key with the ai:chat scope can use them. A request for a model your org has not enabled — or for an unrecognized identifier — fails with:
MODEL_ACCESS_DENIED (HTTP 403) also covers unknown identifiers and models pulled platform-wide during an incident.

​
Choosing a model

For most regulated-industry workloads — clinical documentation assistance, compliance summarization, patient communication drafts — a standard mid-tier model (Sonnet 4.6 or GPT-5.4) delivers the best capability/cost tradeoff. Use the cheapest models (Haiku 4.5, GPT-5.4 mini) for preprocessing, classification, and routing. Reserve premium models for final reasoning steps where accuracy is the primary constraint, and gpt-5.3-codex for code-generation workloads.

​
Deprecation policy

Model versions are deprecated with 6 months advance notice and a 12-month overlap period during which both the old and new versions are available. The model identifier you pass always resolves to a specific model version — HASP does not silently upgrade you to a newer version without your opt-in. When a model version is retired, requests using the retired identifier return:
Subscribe to the Changelog to receive advance notice of upcoming retirements.