How many requests per dollar?

Xiaomi publishes both a pay-as-you-go API and four subscription tiers. The same model (MiMo V2.5) costs different amounts per token depending on how you buy it. The plan-vs-API gap explains the cross-provider "Best" pills above. Figures are at MiMo V2.5's typical shape (830 fresh + 71,500 cached + 295 output tokens). Token Plan pricing, API pricing.

At list price, Xiaomi's own tiers charge 87–105% of the Xiaomi API rate for MiMo V2.5 — Lite and Standard actually cost more than raw API; only Pro (−6%) and Max (−13%) beat it. OpenCode Go and ClinePass resell the same model at ~17% of API cost (~$60 of usage for ~$10), which is why they top the Req / $1 column. The $5–$88 signup prices are a first-month promo that renews at list.

Z.AI sells GLM-5.3 two ways: the GLM Coding Plan (a flat weekly credit budget) and its API, which OpenCode Go resells at the API list price ($1.40 / $4.40 / $0.26 per M tokens). The plan's credits can be valued against that API rate, which is what the numbers below do — the per-request costs are amortized at GLM-5.3's typical shape (700 fresh + 52,000 cached + 150 output tokens). Z.AI GLM Coding Plan docs.

Go's flat $10 buys ~$15 of GLM-5.3 usage (2× headroom) versus ~$67–$939 for the GLM Coding Plan's $18–$168 — Go's GLM-5.3 monthly cap is just $15 of usage (GLM-5.2 keeps a $60 cap on Go), so the Coding tiers lead per dollar on the new flagship. Two caveats: credits run on a rolling weekly window with a 5-hour burst cap and never roll over, and off-peak hours (Mon–Fri 14:00–18:00 SGT) bill at 50% — roughly doubling requests then — which isn't modeled.

Nous Portal resells OpenRouter models at 0.8× the OpenRouter list price (verified against the current list on /info), billed against a monthly credit grant with a +10% bonus: $20 → $22, $100 → $110, $200 → $220. The exception is GPT-5.6 Luna, where a limited-time -50% promo puts the Portal at half the post-repricing $0.20/$1.20 list (verified 2026-08-08). The example below amortizes each plan's subscription over its grant at GPT-5.6 Luna's typical shape (1,000 fresh + 50,000 cached + 220 output tokens) — a shape where OpenCode Go happens to charge 1× the list. Manage subscription, /info prices, API docs.

Nous Portal resells OpenRouter models at 0.8× list against a monthly grant with a +10% bonus. On most models that roughly cancels out the resellers' bundles — but wherever Go charges above list (GPT-5.6 Luna at 1×), Portal wins outright (~≥100 vs ~1,000 req/$). Caveats: unused credits roll over ($10–$100 by tier), which the monthly math ignores. A figure marked ≥ is a cache floor: the plan publishes no cached-read rate, so cached tokens bill at the input rate and the real figure is that number or higher. A cost marked ≤ is a cache ceiling: the real cost is that number or lower.

QwenCloud publishes two distinct prices for the same models: a PAYG rate and a Token Plan rate where credits are deducted from a single monthly quota on the plan's 30-day subscription cycle (Lite 11,500 · Essential 25,500 · Standard 45,000 · Pro 180,000 credits per cycle — the provider removed the earlier 5-hour and 7-day windows on 2026-09-22, so the table's 5h and Week columns render "—"). The docs only say the Token Plan uses "tiered deduction coefficients by model" and refer you to the subscriber console for the actual numbers — there is no public per-model credit table — so the calculator uses the published PAYG rates as a proxy. The table below shows the worked example for Qwen3.7-Plus on the Lite plan at the Grok shape (1,100 fresh + 71,500 cached + 220 output tokens):

QwenCloud doesn't publish per-token credit costs for the Token Plan, so we estimate them from the public pay-as-you-go rates ($0.006512/req at this shape) — absolute Req / $1 figures (marked ~) carry a ≤2.25× scaling uncertainty (re-measured plan budgets price a credit at $0.000444–$0.000696 across the four tiers, 30–56% below the $0.001 proxy conversion), though relative comparisons hold. The monthly quota is the only published budget — no sub-monthly ceilings remain, unused credits don't roll over, and the preview-model capacity boost isn't modeled.