Release v0.15.0 - 2026-09-05

Highlights

Search current GPU cluster offers with deterministic filters and ranking across active live-pricing sources.
Ask Enzan for exact live GPU offer rates and instance-aware cluster totals, including source and freshness context.

Enzan Chat can answer GPU cluster pricing and training-fit questions from current offer rows when live pricing is enabled for chat. Cluster comparisons and switch-option quotes are supported without naming a provider or referring to a current provider; they do not provide provider exclusion or migration support.
Cluster estimates account for how many GPUs are included in each buyable instance, preventing an instance price from being presented as a per-GPU price.
Enzan uses fresh offers to answer pricing questions. Usable older offers include a freshness caveat, while unavailable current data produces a retry prompt instead of an estimated price.

Added:
- POST /v1/enzan/pricing/search for filtering and deterministically ranking active live GPU offers.
- Live-pricing, refresh, cluster-comparison, and switch-option actions in Enzan Chat responses.
Changed:
- Live GPU offer documentation now distinguishes buyable instance counts from GPU counts and identifies hourly rates as per-instance prices.
- Refresh-request rejection logs are now written asynchronously on a best-effort basis, with repeated rejections coalesced. A rejected refresh request is no longer guaranteed a corresponding log entry.

Added searchPricing with typed live-offer filters, ranked results, freshness, and source summaries.
Added the four live-pricing chat action variants.

Added search_pricing with native request, filter, result, and response models.

Added SearchPricing with typed ranking modes, filters, results, freshness, and source summaries.

Added enzan.pricing_search for current live GPU offer discovery and deterministic ranking.

The search endpoint and SDK search methods are additive. Existing refresh responses and SDK error handling are unchanged; refresh-log consumers must not assume one log entry per rejected request.
For live GPU offers whose type ends in -<count>x, the suffix is the GPU count per buyable instance. Cluster-size fields count instances, and hourly prices are per instance.

Updated API, SDK, and MCP documentation for live-offer search and chat grounding.