AI providers / Governed choice
Hyperoru routes each task to an explicit role, measures its cost, and keeps provider expansion behind evidence, privacy, and reliability gates.
Current configuration
The production configuration uses provider-pinned OpenAI model identifiers through LLMGateway. Catalogue presence never means automatic Hyperoru support.
Configured production models are different from provider options under evaluation.
OpenAI
openai/gpt-5.6-luna$1.00 in · $0.10 cached · $6.00 out
Routing, classification, formatting, and citation verification
Chosen for high-frequency routing and verification where latency and disciplined structure matter more than maximal depth.
OpenAI
openai/gpt-5.6-terra$2.50 in · $0.25 cached · $15.00 out
Architecture analysis, security review, and remediation proposals
Chosen for the security and architecture steps that need deeper reasoning over a deliberately bounded evidence set.
OpenAI
openai/gpt-5.6-sol$5.00 in · $0.50 cached · $30.00 out
Explicitly approved reviews that need the highest reasoning budget
Reserved for explicitly approved work where the expected decision quality justifies the largest reasoning and cost envelope.
Provider names and marks reflect the live LLMGateway catalogue. Hyperoru support is a separate, reviewed production decision.
View catalogueCurrent price evidence
The active registry follows LLMGateway's published GPT-5.6 list rates. Try a workload shape, then compare the economics with the quality and authorization boundary required for the task.
Explore model list-price economics before retries, tools, embeddings, or scanner infrastructure. Production cost uses measured provider responses when available.
Illustrative list-price estimate
$5.7000
estimated model list price per month
Input / 1M
$2.50
Output / 1M
$15.00
With credit fee
$5.9850
LLMGateway documents a 5% platform fee when purchasing gateway credits and a free gateway path for bring-your-own provider keys. Rates can change. This calculator is explanatory, not a quote.
Control plane
Provider flexibility is useful only when evidence access, retention, attribution, and action authority stay consistent.
Production model identifiers include the provider so routing is explicit.
The configured gateway stores request metadata for usage analysis, not message or tool payloads.
Tokens, provider cost, latency, cache state, and usage category are attributed to the workspace action.
Models receive selected evidence and typed tools, never ambient shell or unrestricted repository access.
Provider and retention details are linked to current first-party documentation. Read LLMGateway retention documentation
Evaluation contract
A useful answer is not enough. Production models must preserve evidence quality, access control, predictable failure states, and billing-grade attribution.
Can the model distinguish facts, hypotheses, contradictions, and missing coverage without inventing certainty?
Does it call only the narrow tool required, preserve tenant scope, and return the expected typed artifact?
Does it resist prompt injection in repository content and stay inside the configured action boundary?
Do provider retention, region, logging, and private endpoint options satisfy the workspace policy?
Do latency, failure rate, citation verification, and replay behavior hold under realistic review workloads?
Are uncached and cached tokens, provider cost, specialist work, and retries attributable to the exact action?
From catalogue to production
Availability, support, and production use are three different states. The lifecycle makes that distinction visible.
A provider or model is added to the evaluation catalogue, never directly to a production route.
Retention, regions, endpoint controls, model capabilities, and failure behavior are documented.
The model runs a versioned evidence set with citation, isolation, quality, latency, and cost gates.
A named role, pinned identifier, budget, and fallback policy are reviewed before activation.
Evidence IDs remain mandatory for factual claims.
Repository text remains untrusted input, never instruction.
Tools stay typed, narrow, tenant-scoped, and observable.
Consequential actions still require explicit authorization.
Retention and source scope stay visible to the reviewer.
Tokens, retries, cache use, and cost stay attributable.
We measure evidence fidelity, citation validity, decision usefulness, replayability, isolation, latency, and cost before expanding the model set.