GuidesOneMux

Diagnose Hy3 API Incidents With Cost Signals You Can Reproduce

A practical Hy3 cost-checking guide for API incidents: compare equivalent token units, account for cached input, and use request and failure signals to investigate unexpected spend.

AI

Summary · TL;DR

A reliable Hy3 cost estimate starts by matching the price snapshot and billing units to the workload. This guide shows how to calculate an example bill and investigate cost spikes alongside observable API failure signals.

Key Takeaways

  • Hy3 token prices differ across the supplied dated source and OneMux’s published catalogue; do not combine their rates in one calculation.
  • Calculate input, cached input, and output separately, then multiply each by its own rate.
  • For incident diagnosis, compare token usage and cost per successful request with request volume, errors, retries, and latency.
  • Treat a cost change as a signal to investigate, not proof of its cause.
On this page (6)

Start with comparable prices—not a blended headline

When an API bill jumps during an incident, the useful question is not simply “What does Hy3 cost?” It is: Which request pattern changed, and what would that pattern cost at a clearly dated rate? Separate input, cached input, and output; then compare the result with observable request and failure signals.

OneMux publishes access to Hy3 through a unified API. The figures below are a published price snapshot captured on 2026-11-16, not a live quote or benchmark. Prices are per million tokens. The OneMux catalogue lists the published price ratio as 0.26; that ratio is reproduced as listed, not recalculated into a general savings claim. See the Hy3 model page and OneMux pricing for current published details.

Price sourceInput per 1M tokensOutput per 1M tokensCached input per 1M tokens
OneMux catalogue snapshot, 2026-11-16$0.03848$0.1508$0.0091
Official prices listed in OneMux catalogue snapshot, 2026-11-16$0.148$0.58$0.035

The supplied LLMLearner Hy3 pricing page gives a different rate card, dated 2026-11-13: $0.132 per million standard input tokens, $0.528 per million output tokens, and $0.033 per million cached input tokens. These values do not match the later OneMux catalogue snapshot. They may reflect different published rate data or conditions; the supplied information does not establish why. Do not combine rates from the two sources. For the example below, use only the 2026-11-16 catalogue figures. Check the applicable provider and route pricing before using an estimate for a live budget.

Reproducible example: an assistant workload

Assume 10,000 requests in a month, with 8,000 input tokens and 2,000 output tokens per request. Assume 80% of input tokens qualify for the listed cached-input rate. This is a cost illustration, not a measured workload or promise of cache hit rate.

  1. Total input: 10,000 × 8,000 = 80 million tokens. Of these, 64 million are assumed cached and 16 million uncached.
  2. Total output: 10,000 × 2,000 = 20 million tokens.
  3. Multiply each token total by its own per-million rate:
ComponentOneMux catalogue calculationEstimated cost
Cached input64 × $0.0091$0.5824
Uncached input16 × $0.03848$0.61568
Output20 × $0.1508$3.016
Total$0.5824 + $0.61568 + $3.016$4.21408

At these assumptions, the estimate is about $4.21 per month, or about $0.000421 per request. Using the official rates in the same catalogue snapshot, the corresponding calculation is 64 × $0.035 + 16 × $0.148 + 20 × $0.58 = $16.208, about $16.21. These are like-for-like workload calculations using two separate price columns, not a claim about actual invoiced cost. If no input is billed at the cached rate, the OneMux estimate becomes 80 × $0.03848 + 20 × $0.1508 = $6.1008. Actual cache eligibility and billing conditions matter.

Use cost as one incident signal

A rising estimate alone does not tell you whether the cause is traffic, longer prompts, larger responses, retries, or a change in cache behavior. Compare a stable interval before the incident with the incident interval, using the same model route and price basis.

Track these signals by time window and, where available, endpoint or request class:

  • Request count and completed-request count.
  • Input and output tokens per request, plus total tokens by direction.
  • Cached versus uncached input, if the API reports it.
  • Success rate, HTTP status codes, timeouts, and latency.
  • Retry count and repeated requests per user action.
  • Estimated cost per request and per successful request.

A useful diagnosis is a pattern, not a single metric. For example, if request volume is stable but output tokens per successful request rise, investigate longer generations or changed application instructions. If failed requests and retries rise together, check whether the client is resending requests and whether failed attempts still consume billable tokens; the supplied sources do not specify billing for every failure case. If uncached input rises while request shape remains similar, check cache eligibility and whether prompts are changing. These are investigation leads, not proof of cause.

A practical incident checklist

  1. Fix the comparison window. Use equal-duration periods and record the price snapshot or provider rate card in force.
  2. Separate token directions. Reconcile input and output totals independently; output has a different published rate.
  3. Check cache assumptions. Do not classify all input as cached just because a prompt is repeated. Use reported cache usage where available; otherwise show estimates with and without caching.
  4. Reconcile requests and failures. Compare attempts, successes, retries, and token totals. Confirm how failed or cancelled requests are billed with the applicable provider documentation.
  5. Repeat the arithmetic. Divide token totals by one million, multiply by the matching per-million rate, and add components. Keep units consistent.
  6. Mark unknowns. If tier rules, cache eligibility, or failure billing are undocumented for your route, label them unknown rather than assuming zero cost.

OneMux provides model access, routing, keys, spend visibility, and credit top-ups through one OpenAI-compatible API. Compatibility with a particular request pattern depends on documented endpoint and parameter support; this article does not claim that a specific integration or incident workflow has been tested. For a real investigation, use the telemetry your application and API expose, and verify the current applicable prices before drawing a bill-level conclusion.

Bottom line

For Hy3 incident diagnosis, make the estimate reproducible: record the date and rate source, count input, cached input, and output separately, and compare cost with request success, retries, and token usage. A careful estimate narrows the investigation; observable failure signals help determine what to inspect next.

Published price snapshot

Captured 2026-11-16T20:00:06.085Z from the OneMux published catalogue. Prices retain website billing units; this is not a live benchmark. Tiered, cached or time-based billing may differ. Unknown values are not zero.

ModelOfficial inputOfficial outputOneMux inputOneMux outputPublished price ratio
Hy3$0.148$0.58$0.03848$0.15080.26

Current OneMux prices. Compare workload costs only when input/output and billing units match; see each model page for tier conditions.

FAQ

What is the Hy3 price per million tokens?

The OneMux catalogue snapshot captured 2026-11-16 lists $0.148 per million input tokens and $0.58 per million output tokens as official prices, and $0.03848 input and $0.1508 output through OneMux. It lists cached input at $0.035 official and $0.0091 through OneMux. A separate supplied LLMLearner source dated 2026-11-13 reports different rates, so use one dated source consistently and verify applicable current pricing.

How do I estimate Hy3 cost for an API workload?

Count input, cached input, and output tokens separately. Divide each total by one million, multiply by its matching per-million price, and add the results. State request volume, tokens per request, cache assumptions, and the price snapshot.

Can a Hy3 cost spike identify the cause of an API incident?

Not by itself. Compare cost with request volume, input and output tokens per request, cached-input usage, success rate, errors, retries, and latency. These signals can guide investigation but do not prove a cause.

Does cached input mean all repeated prompts receive the cached rate?

No. A repeated prompt does not by itself establish that tokens qualify for cached-input pricing. Check reported cache usage and the applicable route conditions; if unknown, calculate both cached and uncached scenarios.

Related articles