Model pricing · 2026-11-12 · 321AI engineering notes · get a key
Quick answer: The GPT-5.x family spans flagship reasoning models and cheaper high-volume variants, all quoted on the same token-per-1M basis. This guide shows how to read those quotes, estimate real workloads, and pick the tier that earns its keep — using live market floor prices rather than list-price folklore.
A GPT quote has three numbers that matter: input price per 1M tokens, output price per 1M tokens, and the cache-read rate. For chat-shaped traffic the split is roughly balanced. For reasoning-shaped traffic it is not — reasoning models spend tokens thinking before they answer, and thinking tokens bill as output. If your workload is agentic or multi-step, estimate output generously or your budget will be fiction.
Reasoning spend is predictable in shape even when not in magnitude: agent loops multiply cost (turns × context re-sent), and thinking-heavy prompts push output per answer well past the visible answer length. Two habits keep estimates honest — log output tokens separately per step from day one, and budget with a 2-3× output multiplier on multi-step flows until your own logs say otherwise. The usage logs itemize tokens per call, so the multiplier corrects itself within a week of real data.
Every model below is quoted by multiple independent suppliers on 321AI; the table shows the market floor per model and how many lanes bid on it, rendered from the live catalog at page build:
| Model | Market floor price | Channels |
|---|---|---|
| gpt-5.6-sol | $0.0075 in · $0.0375 out / 1M tok | 639 |
| gpt-6-astra | $0.065988 in · $0.32994 out / 1M tok | 1020 |
Market floor across competing suppliers, rendered live at page build (data date 2026-11-20); per-supplier quotes on the marketplace. Machine-readable: GET https://321ai.cc/api/public/model-prices.
Per-supplier quotes are on the OpenAI-family page; the whole catalog lives at /en/models/.
Three common shapes, computed from the live floor at page build. The volumes are illustrative — swap in your own:
| Workload | Model (live floor) | Volume | Estimated cost |
|---|---|---|---|
| Support bot, busy day | gpt-5.6-sol | 5M in / 1M out | $0.0750 |
| Agent loop with tools | gpt-5.6-sol | 3M in / 0.9M out | $0.0563 |
| Batch classification | gpt-5.6-sol | 1M in / 0.1M out | $0.0112 |
Computed at page build from live market floor prices (data date 2026-11-20): cost = input tokens × input price + output tokens × output price. Platform fee is 0 and failures are never billed, so this is the whole bill. Volumes are illustrative — re-run with your own from the usage logs.
| Flow | Dominant cost | Planning rule |
|---|---|---|
| Chat support | Balanced in/out | Estimate 5:1 input:output, then correct from logs |
| Agent / tool loops | Output + re-sent context | Budget 2-3× output; cache the loop's stable prefix |
| Batch classification | Input | Cheap tier first; verify quality on a sample |
The workable procedure: prototype on the flagship tier and log tokens per task; find the stable, high-volume flows whose quality survives a cheaper tier; move those flows and keep the hard cases on the flagship. The comparison pages price any two models head-to-head on live data, and the success-rate column on the marketplace tells you whether a cheaper lane is also a reliable one. Token choice is per key — see the FAQ for the group semantics.
One more input belongs in the decision: the guaranteed token group. If the flow serves customers and model authenticity matters commercially, its merchant-supplied lanes are part of the tier decision, not an afterthought — price, stability and supply transparency are three columns of the same choice, and the setup guides show how to wire different keys to different flows.
On a single-source plan, a GPT price is a wall. On the marketplace it is a board: suppliers bid per model, the floor updates as they do, and conditional routes re-pick the best lane within 5 minutes. Platform fee is 0 and failed calls refund automatically, so the floor you read is the bill you get. The pricing page at /en/pricing shows the current board in full.
A final note on timing: GPT pricing is the most watched board in the market, which means it moves the most. Quotes you screenshot today are stale in both directions — a cheaper lane can appear mid-week, and a flagship can firm up. Anything downstream of GPT pricing (budgets, unit economics, a pricing page of your own) should read the public API at build time or daily, never from a hardcoded constant.
Reasoning tokens bill as output, and multi-step agent loops re-send context every turn. Log tokens per step from the usage logs and re-run the estimate with output-heavy ratios — the usage logs itemize per call.
They are market floors from competing suppliers, which move with the market and are frequently below any static list. The public API exposes the current numbers — judge with live data, not cached screenshots.
Whatever your logged token data justifies: keep hard reasoning on the flagship, move stable high-volume flows to a cheaper tier once quality holds. The stable token group keeps stability prioritized on production keys.
No. GPT models run on the OpenAI protocol natively: set base_url to https://321ai.cc/v1 and existing SDKs work unchanged.
Live prices: /en/models/ · billing: /en/pricing · setup: /en/guides/ · FAQ