Cost comparison · 2026-11-12 · 321AI engineering notes · get a key
Quick answer: Cost per token tells you nothing by itself; cost per task tells you what a model really costs. This guide fixes three everyday workloads — code review, long-form drafting, translation — and prices them across DeepSeek, GPT and Claude lanes using live market floor prices.
A task has a token footprint: what you send and what you expect back. We fix the footprints below (they are illustrative — measure your own from usage logs), then apply each family’s live floor price. The formula is always the same:
task_cost = (input_tokens x input_price + output_tokens x output_price) / 1,000,000
Why these three tasks: they cover the three volume shapes. Code review is input-heavy (context in, verdicts out), drafting is output-heavy, and translation is symmetric — so together they expose how each family prices all three shapes rather than favoring one. Any task you add later will be a blend of these shapes, which makes the pattern easy to extrapolate.
| Task | Token footprint (illustrative) | What drives the volume |
|---|---|---|
| Code review | 40K in / 3K out | Diff + context in, findings out |
| Long-form draft | 12K in / 6K out | Brief + style refs in, article out |
| Translation | 20K in / 20K out | Mirror ratio: text in, translation out |
Rendered from the live catalog at page build — market floor across competing suppliers, with supplier counts:
| Model | Market floor price | Channels |
|---|---|---|
| claude-sonnet-5 | $0.033492 in · $0.16746 out / 1M tok | 501 |
| gpt-5.6-sol | $0.0075 in · $0.0375 out / 1M tok | 639 |
| deepseek-v4.1-flash-1017 | <$0.0001 in · $0.00027 out / 1M tok | 278 |
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.
The same three tasks applied to each family’s current floor:
| Workload | Model (live floor) | Volume | Estimated cost |
|---|---|---|---|
| Code review — Claude lane | claude-sonnet-5 | 40K in / 3K out | $0.0018 |
| Code review — GPT lane | gpt-5.6-sol | 40K in / 3K out | $0.0004 |
| Code review — DeepSeek lane | deepseek-v4.1-flash-1017 | 40K in / 3K out | $0.0000 |
| Drafting — Claude lane | claude-sonnet-5 | 12K in / 6K out | $0.0014 |
| Drafting — GPT lane | gpt-5.6-sol | 12K in / 6K out | $0.0003 |
| Drafting — DeepSeek lane | deepseek-v4.1-flash-1017 | 12K in / 6K out | $0.0000 |
| Translation — Claude lane | claude-sonnet-5 | 20K in / 20K out | $0.0040 |
| Translation — GPT lane | gpt-5.6-sol | 20K in / 20K out | $0.0009 |
| Translation — DeepSeek lane | deepseek-v4.1-flash-1017 | 20K in / 20K out | $0.0000 |
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.
The tables above use our illustrative footprints; yours will differ. Building a private version is an afternoon: pick five tasks that actually recur in your product, log their real input/output tokens from usage logs, run each against two candidate models, and score outputs with a rubric you can defend. Price the winners with the public API and you have a cost-per-task table for your product, not for the average product — which is the only one that pays your bills. The setup guides connect any client you want to benchmark.
Cost is one axis. Quality is not in the table — it belongs to your eval on your tasks, and the honest procedure is to run the task on two lanes and compare outputs yourself. Reliability is on the board though: stability and P50 latency per supplier are public on the marketplace, so a cheap lane with a poor stability is visible as exactly that. Head-to-head model pairs have their own pages under /en/compare/ with live prices on both sides.
Practical pattern: use the cheap lane where the task is high-volume and verifiable (translation drafts, classification, first-pass review), keep the premium lane where errors are expensive (final review, hard reasoning), and let conditional routes re-pick within 5 minutes as floors move. Switching a model costs one line in the model parameter — the calculator helps you budget before you commit.
Depends on the task footprint and the live board — the tables above are computed from current floors at page build. Run your own footprint through the public price API for a current answer rather than trusting any static ranking.
Check its upstream self-check records on the marketplace first. For user-facing traffic, the stable token group prioritizes stability; the default group is the price-first lane.
No. All models run behind one endpoint: base_url https://321ai.cc/v1 for the OpenAI protocol. Switching a task between families is a model-name change, nothing else.
The usage logs itemize tokens per call. Sum input and output across a representative batch, then reuse the formula in this article with live prices from the public API.
Live prices: /en/models/ · billing: /en/pricing · setup: /en/guides/ · FAQ