How to read the Q4 2026 model list: naming, tiers and three pricing traps

These names are no longer version numbers

Open any relay's model list and you see names like these:

GPT-6 Astra / GPT-5.6 Sol / GPT-5.6 Terra / GPT-5.6 Luna
Claude Opus 5 / Claude Sonnet 5 / Claude Haiku 4.5 / Claude Fable 5.1
GLM-5.3 / GLM-5.3-Flash / GLM-4.7
Qwen3.7-Max / Qwen3.7-Plus / Qwen3.7-Flash
Kimi K3 / Kimi K2.7 Code

Four entirely different things are mixed together here:

1. Generation (6, 5.6, 5.3, 3.7, K3): the training generation;
2. Product codename (Astra, Sol, Terra, Luna, Fable): positioning within one generation;
3. Tier suffix (Max, Plus, Flash, mini, Lite, Air): the trade-off between capability and price;
4. Task marker (Code, Pro, Thinking): versions tuned for a specific job.

The key judgement: a bigger number does not mean a stronger model. GPT-5.6 Luna has a longer number than GPT-5.6, yet Luna is the lightest tier of that generation. GLM-5.3-Flash carries exactly the same number as GLM-5.3 and is an order of magnitude weaker.

Trap one: reading a tier suffix as an upgrade

"5.6 now has Luna, so is it stronger than 5.6?"

No. The suffix splits one generation into tiers, usually three:

Pick the tier first, then compare prices. Comparing across tiers is comparing economy with first class: "ten times cheaper" is numerically true and analytically meaningless.

Trap two: looking only at input price

Model pricing has two numbers, input and output, and output is typically four to five times input. Some models also list a separate cache price.

Look at this set (USD per million tokens):

That ratio is itself a signal. A low ratio (3.0x) suggests a model built for output-heavy work such as writing and generation. A high ratio (6.0x) suits input-heavy work such as long-document analysis and code review.

Real cost must be computed on one typical request:

cost per request = input tokens / 1,000,000 * input price + output tokens / 1,000,000 * output price

Worked example: a request that reads an 8,000-word document and writes a 500-word summary uses roughly 12,000 input tokens and 800 output tokens.

• GPT-6 Astra: 12000 10 / 1e6 + 800 50 / 1e6 = 0.12 + 0.04 = $0.16
• GPT-5.6 Sol: 12000 4 / 1e6 + 800 20 / 1e6 = 0.048 + 0.016 = $0.064
• DeepSeek V4 Pro: 12000 0.66 / 1e6 + 800 1.98 / 1e6 = 0.0079 + 0.0016 = $0.0095

The most expensive is 17 times the cheapest. Judging by input price alone you would expect about 15 times — which looks close, until you multiply by volume: at 10,000 requests a day, the annual difference is roughly $550,000.

Conclusion: unit price is where selection starts, not where it ends. Convert it onto your own token mix.

Trap three: long context is not free

Growing context from 128K to 1M is real progress, but it brings two problems people miss:

1. Tiered pricing: some vendors bill input above a threshold (say 200K) at a higher unit price. The list shows the base rate, so long-input pricing has to be checked separately;
2. Attention decay: inside a 1M context, information in the middle is used correctly far less often than information at the start or end. That is not a pricing issue but an effectiveness one — fitting it in is not the same as using it.

Do this: before relying on long context, run a middle-of-document retrieval test. Put the key fact in the exact centre, then ask a question that cannot be answered without it. If it answers correctly, then talk about price.

The correct order for tier-based selection

1. Decide the tier: does this task need flagship, workhorse or lightweight? Judge by the cost of getting it wrong, not by benchmark scores.
2. Compare prices within that tier, using the per-request formula on your own token mix.
3. Check the output-to-input ratio, and match it to your workload's shape.
4. Look at benchmarks last. Differences within a tier are usually small, while prices can differ threefold.

Doing it the other way round — read the leaderboard, then hunt for the cheapest — is how most people go wrong, because leaderboards mostly rank flagship models while 90% of your traffic belongs on the workhorse tier.

How to use this site's model library

The library groups models by vendor and labels each with a tier, with prices normalised to USD per million tokens. Three suggestions:

• Browse by vendor rather than by the overall ranking: each vendor's three tiers are a designed set, and cross-vendor comparison requires aligning tiers first;
• Read the price note and verification date: Chinese vendors publish in renminbi and this site converts at a fixed rate, so exchange-rate movement introduces a few percent of drift;
• Never treat an aggregator's price as the official price: aggregators sell their own regional price list, which can differ from the vendor's site by several times. The two sets cannot be mixed.

Summary

1. Model names mix generations, codenames, tier suffixes and task markers; a bigger number is not a stronger model.
2. Choose the tier before comparing prices; cross-tier comparison is meaningless.
3. Input price alone misleads: output is usually four to five times input, so compute a typical request.
4. A low output-to-input ratio means generation-heavy, a high one means analysis-heavy; match it to your workload.
5. A 1M context is not a usable 1M: long input may be tier-priced and mid-context recall decays.
6. Ninety percent of traffic belongs on the workhorse tier; do not make global decisions from a flagship leaderboard.