The 20 most used models on MoeRouter and what each is for (November 2026)
The 20 models that used the most tokens on MoeRouter over the last 7 days, with what each is good for, whether to move off its predecessor, and picks by use. Every new model links to a review that gathers independent tests.
In short
- The ranking is by token usage on MoeRouter over the 7 days to November 18, 2026. It gives ranks only, not volumes.
- Of the top 20, eight are OpenAI's and four Google's; Anthropic, DeepSeek and Qwen have two each, Zhipu and Moonshot one each.
- The models at the top are not the strongest but the ones that are good enough and cheap: GPT-6.1 Sol is first, and the inexpensive DeepSeek V4.1 Flash and GLM-5.3-Flash are both in the top four.
- By use: GPT-6.1 Sol or Claude Opus 5.5 for coding, Claude Fable 5.2 for the hardest problems, the Flash tier and GPT-6 Luna for volume.
On this page
The ranking#
gpt-6.1-soldeepseek-v4.1-flash-1017gpt-6.2-solglm-5.3-flashgpt-6-astragpt-6.2-astraclaude-opus-5-5gpt-6-lunaqwen3.8-flashclaude-fable-5-2kimi-k3.1-progemini-4-hydrogengemini-3.8-flashgpt-6-solgemini-4-argongemini-4-carbondeepseek-v4.1-flashgpt-6.1-terragpt-5.6-lunaqwen3.8-max-0902
Ranked by token usage on MoeRouter over the last 7 days (2026-11-18).
How it is counted: each model's token usage on MoeRouter over the last 7 days, highest first. We publish ranks only, not volumes or shares, and list only the models the model square offers publicly.
What stands out#
- The top is good enough and cheap. First place is GPT-6.1 Sol, not the stronger Astra or Claude Fable 5.2. The inexpensive DeepSeek V4.1 Flash and GLM-5.3-Flash are both in the top four.
- New models climb fast. GPT-6.2 Sol and Astra launched on November 17; with the count closing on the 18th they were already 3rd and 6th.
- Old models are still used. GPT-6 Sol, GPT-5.6 Luna and Gemini 3.8 Flash all have successors and are still in the top 20: some projects need a fixed version, and some older models are cheaper here.
- By maker: OpenAI eight, Google four, Anthropic, DeepSeek and Qwen two each, Zhipu and Moonshot one each.
Picks by use#
| For | Try first | Review |
|---|---|---|
| Daily coding in Codex | gpt-6.1-sol |
GPT-6.1 Sol and Terra |
| Daily coding in Claude Code | claude-opus-5-5 |
Claude Opus 5.5 |
| Problems other models cannot get through | claude-fable-5-2 |
Claude Fable 5.2 |
| Hard problems where cost and speed matter more | gpt-6.2-astra |
GPT-6.2 Sol and Astra |
| A cheap workhorse for code | deepseek-v4.1-flash |
DeepSeek V4.1 Flash |
| Chinese Q&A, documents and multimodal work on a budget | glm-5.3-flash, qwen3.8-flash |
GLM-5.3-Flash, Qwen3.8 |
| Reading images, audio, video and long documents | gemini-4-hydrogen |
Gemini 4 |
| Flagship-class coding without a flagship price | kimi-k3.1-pro |
Kimi K3.1 Pro |
| Bulk classification, extraction, rewriting | gpt-6-luna, gpt-6.1-terra |
See below |
The reference prices of a few common choices, side by side:
| Model | Group | Input | Output | Cache read |
|---|---|---|---|---|
gpt-6.1-sol | default | CN¥0.324≈ US$0.0486 | CN¥1.62≈ US$0.243 | CN¥0.0162≈ US$0.00243 |
claude-opus-5-5 | ccmax-cheap | CN¥4.8≈ US$0.72 | CN¥24≈ US$3.6 | CN¥0.24≈ US$0.036 |
default | CN¥5.2≈ US$0.78 | CN¥26≈ US$3.9 | CN¥0.26≈ US$0.039 | |
stable | CN¥5.2≈ US$0.78 | CN¥26≈ US$3.9 | CN¥0.26≈ US$0.039 | |
deepseek-v4.1-flash | default | CN¥0.1≈ US$0.015 | CN¥0.4≈ US$0.06 | CN¥0.002≈ US$0.0003 |
glm-5.3-flash | default | CN¥0.0072≈ US$0.00108 | CN¥0.0252≈ US$0.00378 | CN¥0.00207≈ US$0.000311 |
gpt-6-luna | default | CN¥0.018≈ US$0.0027 | CN¥0.09≈ US$0.0135 | CN¥0.0018≈ US$0.00027 |
Reference prices per 1M tokens in Chinese yuan (CNY), the currency accounts are charged in, from the model square (). Dynamically priced models settle at the route that served each request; the model square is authoritative. The same model takes other routes in another group and costs differently there; a key's group is chosen when the key is created. About groups Amounts in USD are estimates, converted at 1 CNY = 0.15001 USD ().
One by one#
OpenAI: eight#
gpt-6.1-solis first by usage: the everyday model in Codex, cheap, and able to continue long sessions. See the GPT-6.1 Sol and Terra review.gpt-6.2-solandgpt-6.2-astraare the generation released on November 17. Astra is second on general intelligence; Sol lists at 6.1 Sol's price and is 3 points lower on an independent coding-agent index. See the GPT-6.2 review.gpt-6-astrais the previous top tier. 6.2 Astra scores no lower and costs less per token: use 6.2 for new projects.gpt-6-solis the version before 6.1 Sol, kept for projects that need a fixed version; new projects should use 6.1 Sol.gpt-6.1-terrais the half-price fast tier, for short tasks such as forms, extraction and computer use. It cannot continue a long Codex session that carries encrypted reasoning.gpt-6-lunais OpenAI's cheapest light model, listing at $0.10 input and $0.50 output per 1M tokens[1]. On the night it launched, the Bilibili creator GenJi noted that its prices were lower even than DeepSeek's[3]. It suits large-scale classification, extraction and simple sub-tasks; a common layering is Luna for the cheap preliminary steps, Sol for the everyday work, and an upgrade only on real failure[1].gpt-5.6-lunais the previous light model. GPT-6 Luna has the lower reference price on MoeRouter, so switch unless there is a reason not to.
Anthropic: two#
claude-opus-5-5is the default choice for Claude Code and the model reviewers call their daily driver. See the Opus 5.5 review.claude-fable-5-2is the strongest and the dearest, for hard problems. It is not in the default group: it needs a key of the stable or ccmax-cheap group. See the Fable 5.2 review.
Google: four#
gemini-4-hydrogen,gemini-4-argonandgemini-4-carbonare the three Gemini 4 models, for fast execution, balance and the hardest work. On MoeRouter they cost far less in the google-proxy group than in the other two. See the Gemini 4 review.gemini-3.8-flashis the generation before Hydrogen. Google's promotional price for it runs until December 31, 2026, after which it matches Hydrogen's[2]. In MoeRouter's default group it is much cheaper than Hydrogen at present, which is why fixed pipelines still use it; Hydrogen is clearly the stronger coder.
Chinese models: six#
deepseek-v4.1-flash-1017anddeepseek-v4.1-flash: the cheap workhorse for code, second by usage. The name ending in1017is the pinned build. See the DeepSeek V4.1 Flash review.glm-5.3-flash: the lowest reference price of all, natively multimodal, on the slow side. See the GLM-5.3-Flash review.qwen3.8-flashandqwen3.8-max-0902: Alibaba's light tier and flagship. See the Qwen3.8 review.kimi-k3.1-pro: Moonshot's flagship, with vendor scores close to Claude Opus 5.5 and divided hands-on opinion. See the Kimi K3.1 Pro review.
A layered set-up that saves money#
The ranking shows what most people do: a cheap model by default, and an upgrade for the hard tasks.
- First layer:
deepseek-v4.1-flash,glm-5.3-flash,gpt-6-luna, for volume that can be checked automatically. - Second layer:
gpt-6.1-sol,gemini-4-hydrogen, for daily coding and agent work. - Third layer:
claude-opus-5-5,gpt-6.2-astra,kimi-k3.1-pro, for what the second layer keeps failing. - Fourth layer:
claude-fable-5-2,gemini-4-carbon, for the hardest problems only.
The higher the layer, the more tokens and money a task usually takes. Sending the failed tasks upward, instead of running everything on the strongest model, is the easiest money to save. Billing has more ways to spend less. The docs are in Chinese for now.
Frequently asked questions#
How is the usage ranking counted?#
By each model's token usage on MoeRouter over the last 7 days, input and output together, highest first. The ranking in this post is as of November 18, 2026.
Why ranks and not volumes?#
Volumes and shares are business figures we do not publish. Ranks are enough to answer what people actually use.
Does high usage mean a model is good?#
No. It means the model's balance of price and ability fits most tasks. Claude Fable 5.2, the strongest, is 10th because it is expensive and used only on hard problems; cheap models handle the volume and so use the most tokens.
Which model is the cheapest?#
On this post's data the lowest reference price is glm-5.3-flash, then gpt-6-luna. Prices change: the table in this post and the model square are authoritative.
Where should a newcomer start?#
With gpt-6.1-sol in Codex, claude-opus-5-5 in Claude Code, and deepseek-v4.1-flash for scripts and batch jobs. The quick start creates a key.
Can one key call all 20 models?#
A key of the default group can call nearly all of them. The exception is claude-fable-5-2, which is only in the stable and ccmax-cheap groups; and Gemini 4 is cheaper in the google-proxy group. You can create several keys in different groups; see about groups.
Sources
The test results and figures quoted in this post come from the public sources below and belong to their authors; each quotation is marked where it appears.
- GPT-6.1 Sol正式发布:价格、跑分、Codex与GPT-6 Astra、Claude Opus 5.5怎么选 (in Chinese)
- Gemini三个模型怎么选:Hydrogen、Argon、Carbon适用场景梳理 (in Chinese)
- Opus 5.5、GPT-6 Sol首发实测,Claude再次统治世界?! (video, in Chinese)
Disclosure: MoeRouter sells API access to the models discussed here. The test results in this post come from the third-party sources cited; prices and endpoint details come from MoeRouter's own data.