Methodology

How the numbers are made

The build fetches OpenRouter's models API and keeps labs deepseek, qwen, moonshotai, z-ai, minimax, bytedance, bytedance-seed, tencent, stepfun, xiaomi, and baidu. It drops ids that contain a colon, and it drops zero-price rows. The snapshot on this build is 2026-10-16T04:09:43.607Z.

Blended price

Blended price per million tokens is 75% of the input price plus 25% of the output price. "X% cheaper" is one minus that blend divided by the comparator's blend, rounded to the nearest percent. A negative number is printed as "dearer".

Comparators

The Opus comparator is the newest OpenRouter id matching a base Claude Opus version, ignoring batch, pro, and preview suffixes. The Sol comparator is the newest base GPT Sol id, same rule. This build resolved Anthropic: Claude Opus 5.5 at $4.00/$20.00 and OpenAI: GPT-6.1 Sol at $2.00/$10.00. Articles do not type those version strings. The template prints the name from the snapshot, so the next build can move on without a rewrite.

List price versus the lab

OpenRouter's model-level price is often the cheapest host, or a host that is simply listed first. For GLM-5.3 and Kimi K3 we also stored the endpoints payload on 16 Oct 2026. Z.AI's own host matches Z.AI's docs at $1.40/$4.40. Moonshot's own host matches Moonshot's docs at $3/$15. The headline list prices do not. Guides that quote either model show both.

Benchmarks

• LMArena style-control ratings from the Hugging Face dataset lmarena-ai/leaderboard-dataset, config text_style_control, split latest, publish date on the rows we kept: 15 Oct 2026. These are Bradley-Terry ratings, not Elo.
• SWE-bench Verified resolved percentages from the SWE-bench site's leaderboards.json, fetched 16 Oct 2026. Harnesses differ. The current Opus and Sol aliases are not rows on that board.
• Terminal-Bench 2.1 figures are copied from Moonshot's Kimi K3 model card, with the card's harness note. We did not re-run the bench.
• Artificial Analysis needs an API key we do not have, so it is not a chart.

Task costs

The calculator multiplies an illustrative token mix by the price. The mixes are round numbers, not a trace from a real session. Cache-write fees are left out. If a model has no cache-read price, cached tokens are billed as input, and the table says so.

What we refuse to fill in

If a source did not load, the sentence says so. We do not invent a free-tier size, a ban, or a benchmark. The corrections log is where a wrong sentence goes after we ship it.