One Chat interface in the app, with Inroi Chat first and Requesty Responses as fallback

The hardest part of using a relay is rarely a missing comma. It is keeping three separate questions separate: what protocol the app speaks to Foreverse, what protocol Foreverse speaks to each relay, and which price table settles the result. The Luna/Terra production route now has a clean answer. The client always sends OpenAI Chat; Inroi is the primary Chat upstream; only a pre-frame primary failure is converted into a Requesty Responses call.

Why keep one app-facing Chat API and two upstream protocols?

Android and iOS do not need to know which relay is carrying a request. They continue to send Chat Completions SSE to Foreverse. The backend performs the provider-specific conversion, so continuation, expansion, roleplay, and agent traffic keep one client message path instead of scattering vendor branches through the UI.

Inroi's API base is https://www.inroi.shop/v1. Requesty's API base ishttps://router.requesty.ai/v1; https://app.requesty.ai/ is a dashboard and returns a web page or 404 when treated as an API host. The working Requesty path is not an openai/ model forced through Chat. It is /responses with the openai-responses/ catalog prefix.

Failover policy: one primary attempt, one fallback attempt—not three retries

Text generation is a paid, long-running write. Retrying the same broken route three times multiplies delay and may multiply upstream cost. The safe boundary is whether the user has received an upstream frame: before that point, switching is possible; after it, replay is prohibited.

If both routes fail with zero output, the hold is refunded. If Inroi fails and Requesty succeeds, the request settles once from final usage. The failed primary attempt still enters channel-health telemetry so fallback success does not erase it, but deliberately injected production checks are not a relay SLA sample.

What actually passed in production?

The final matrix went beyond a hello-world prompt. Both routes exercised normal continuation, one-shot expansion, the first agent turn that emits a tool call, the second turn that returns a final answer after tool output, and a repeated-prefix cache probe. Earlier timeout, partial-stream, and Requesty tool-shape failures remain part of the record; the table describes the fixed final path.

The earlier Requesty agent 400 came from sending Chat's nested named-function selector unchanged into Responses. A direct A/B isolated the required flat function-name shape. Once the conversion was fixed, both Luna and Terra completed the tool call and tool-result turn. “Supports tools” and “accepts every compatibility parameter shape” are not the same claim.

An older screening run also produced an xhigh response with output_tokens=8192,reasoning_tokens=8192, and no prose. Those direct Requesty calls explicitly used an 8K benchmark cap and did not traverse Foreverse. The old server and Agent / Reader defaults nevertheless used the same 8K envelope, so the product route had the same failure mode. Reasoning, tool arguments, and visible prose share the output budget; 8K was a billing-hold guardrail, not the model's capability limit.

How much cache was observed—and why is cache write still unknown?

A direct Requesty repeat first established automatic prefix caching. Production then tested Inroi with an explicit stable route key and Requesty through the forced Responses fallback. These are token observations, not a promise of a permanent hit rate.

A separate explicit 5,247-token probe hit 5,244 cached tokens on both official model IDs. Inroi did not report cache hits for repeated large prefixes until the request received a stable prompt_cache_key and explicit cache mode. That key is a hash derived from user and public model scope; it contains no prompt, email address, or raw user ID.

Neither relay currently exposes cache_write_tokens. The honest value is therefore null, not zero. The backend understands both the standard field and Requesty's compatibility alias and can price writes at 1.25x, but it does not reverse-engineer writes from a miss or from a relay cost field.

Keep base short-context rates separate from App prices with a 50% markup

The normalized price source is the OpenAI Standard short-context table and prompt-caching rules as observed on July 31, 2026—not either relay's usage.cost. OpenAI's July 30 changelog also records an 80% price cut for Luna and a 20% cut for Terra; the table below uses those post-cut official short-context rates.

1 credit = $0.0001. The base dollar rate is converted to credits, multiplied by 1.5, and the whole request is rounded up. Ordinary input, cache writes, and cache reads are mutually exclusive slices of prompt tokens; cached tokens are not charged twice, and the cache-write rate is exactly 1.25x ordinary input. Foreverse currently caps relayed prompts for these two official models at 200k tokens so a request cannot enter a more expensive long-context tier while being sold at the short-tier rate.

Inroi and Requesty can have different internal acquisition costs, and a relay may report an older rate or its own multiplier. Foreverse does not pass that difference to the user: both routes settle from the normalized table above. Conversely, when usage is incomplete, one internal cost number is not enough to invent a token breakdown.

What this production test proves—and what it does not

It proves that these two official OpenAI model IDs completed continuation, expansion, agent tools, cache reads, failover, and normalized billing through two production relay paths. It does not turn relayed traffic into a direct-OpenAI test, turn forced-failure checks into an SLA, or substitute interface success for a fiction-quality benchmark.

Catalogs and prices change; every number on this page is dated July 31, 2026. For the mechanics behind repeated-prefix discounts, read our prompt-caching bill post-mortem. For the difference between a hosted route and your own key, see what BYOK actually means. The longer market view lives in our token-price trend memo.

Writing quality has its own evidence trail: the 548-completion Luna/Terra fiction benchmark and the DeepSeek/Luna/Terra buying comparison. The unedited Chinese round 1 and round 20 outputsare published separately so readers can choose before seeing the model names.

Protocol references: Requesty Responses,OpenAI changelog,Standard pricing, andprompt caching.