// community guide · November 2026

Which model to use for what.

The 9 models on the cluster, their credit rates and a simulator of the real cost for the way you work. No hype: it compares cost, not quality or speed.

9 modelsrates per 1M tokens1 credit ≈ 0.026€ (1,000 pack)source: NaN credits panel

Guide written by @ntizar, a NaN builder, and published as is; we only converted the amounts from dollars to euros, which is how NaN charges. Rates are from November 2026: the current price of each model is at cloud.nan.builders/tokens, and the model ids are in Choose your model.

01 / the guide

Choose by task, not by name

The executive summary. The numbers behind it are below, in the simulator and the rates.

Everyday agentic coding

MiMo V2.6 Flashmimo-v2.6-flash

The sensible default. The cheapest in almost any mix and unbeatable at generating output (14 cr/1M, 2x ratio).

Alternative: GLM 5.3 Flash if your agent rereads the same context many times (high cache).

Cache-heavy flows

GLM 5.3 Flashglm5.3-flash

With 60% cached input it comes to 4.4 cr/1M, the lowest on the cluster. But its cache costs 7.5x MiMo's: it only wins if your cache hit rate is truly high.

Watch out: its published typical cost (1.4 cr) is worse than MiMo's (0.6) despite having the cheapest list input.

Everyday tasks and general chat

MiMo or Qwen 3.8 Flashmimo-v2.6-flash · qwen3.8-flash

Qwen Flash is meant for everyday work, but its typical (1.9 cr) is three times MiMo's (0.6). Unless you prefer Qwen, MiMo stretches the pack further.

Long documents (1M context)

Hy4hy4

The cheapest large model in all three measured scenarios (32/67/98 cr/1M) and the only one built specifically for long documents.

Bonus: its cache (2.10 cr) is the cheapest among the large models.

Hard reasoning

DeepSeek V4-Prodeepseek-v4-pro

106 cr/1M on a balanced mix: 35% less than Qwen Max (163) and less than half of Kimi (242). Its published typical (7.7 cr) is the best of the large models.

Volume and tight budgets

MiMo V2.6 Flashmimo-v2.6-flash

~1,666M tokens per 1,000 credits. The opposite end is Kimi: ~50M. A 33x difference for the same money.

Avoid for volume: Kimi K3.1 Pro (output at 500 cr/1M and the only one that charges cache write separately).

When the Flash falls short

GLM 5.3 fullglm5.3

The quality jump within GLM, but at ~11x the price of the Flash. Save it for what really needs it; its cache (1.05 cr) is as cheap as the Flash's.

Quick rule

Flash firstalways

Always start with a Flash (MiMo or GLM). Move up to Hy4, DeepSeek Pro or Qwen Max only when the task proves it needs it. The large models cost 8x to 33x more per token.

Unsure about your mix?

Ask on Discordthat's where the community lives

Paste the numbers the simulator gives you and compare real usage per task with other builders. The website only handles access; the community is on Discord.

02 / watch real usage

List price isn't everything

Everything above compares prices per token. But you don't pay for tokens: you pay for tasks.

A builder's observation · unverified
“With DeepSeek V4-Flash I noticed it used far more tokens than expected, maybe because of the speed.”
— @ntizar, NaN builder · personal experience, not a systematic measurement

Two models with the same price per token can cost very differently per task: one that answers fast but retries, over-generates or doesn't use the cache spends more tokens for the same result.

  • Measure your real flow: tokens per completed task, not per million. A week of logs is enough to see patterns.
  • Compare cost per session (the simulator below does it), not just the list rate.
  • Published typicals already diverge from the list: GLM 5.3 Flash has the cheapest input and yet its typical (1.4 cr) is worse than MiMo's (0.6) or DeepSeek Flash's (1.1).
  • Cache is the big lever: between 2% (MiMo) and 20% (GLM Flash) of the input price. Structure your prompts to reuse context.

03 / simulator

Your mix, your real cost

Move the percentages to match how you work (default: 40/20/40, typical of agentic coding) and see the cost per million tokens, per session and how many sessions a pack gives you.

// scenarios
modelcost / 1Mcost / sessionsessions / packcache vs input

04 / rates

The 9 models, price by price

Credits and euros per million tokens, at the list price of the 1,000-credit pack (26€, 0.026€/cr; larger packs and the 5% member discount come out cheaper). ZDR: zero data retention. EXT: external provider.

modelinputcacheoutputcache writeout:intypical / 1Mtokens / 1,000 cr
DeepSeek V4-Flash deepseek-v4-flash Fast general-purpose and coding model, 1M context ZDR 10.5 cr0.273€ 0.21 cr0.005€ 42 cr1.09€ — 4x 1.1 cr 930M
GLM 5.3 Flash glm5.3-flash Fast GLM for agents and code ZDR 5.25 cr0.136€ 1.05 cr0.027€ 17.5 cr0.455€ — 3.3x 1.4 cr 706M
MiMo V2.6 Flash mimo-v2.6-flash Xiaomi's fast model, 1M context ZDR EXT 7 cr0.182€ 0.14 cr0.004€ 14 cr0.364€ — 2x 0.6 cr 1,666M
Qwen 3.8 Flash qwen3.8-flash Fast Qwen for everyday tasks, 1M context ZDR 7.5 cr0.195€ 0.8 cr0.021€ 23.5 cr0.611€ — 3.1x 1.9 cr 514M
GLM 5.3 (full) glm5.3 The full GLM 5.3, about 11x the price of the Flash ZDR 49 cr1.27€ 1.05 cr0.027€ 154 cr4€ — 3.1x 11.5 cr 87.3M
Kimi K3.1 Pro kimi-k3.1-pro Moonshot's frontier model for reasoning, agents and code, 1M context ZDR EXT 100 cr2.6€ 10 cr0.26€ 500 cr13€ 100 cr2.6€ 5x 19.9 cr 50.3M
Qwen3.8 Max qwen3.8-max The largest Qwen 3.8, for hard reasoning and agents, 1M context ZDR EXT 100 cr2.6€ 12.5 cr0.325€ 300 cr7.8€ — 3x 20.2 cr 49.5M
DeepSeek V4-Pro deepseek-v4-pro The large DeepSeek for reasoning and code, 1M context ZDR EXT 66 cr1.72€ 2.2 cr0.057€ 198 cr5.15€ — 3x 7.7 cr 130M
Hy4 hy4 Tencent's Hunyuan 4 for agents and long documents, 1M context ZDR EXT 41.7 cr1.08€ 2.1 cr0.055€ 125.05 cr3.25€ — 3x 5.5 cr 181M

Only Kimi has a published cache write rate; for the rest the figure is missing (that's why the simulator warns you if you use it). The "typical" is the one NaN publishes for real mixed usage.

05 / analysis

Three scenarios, three winners

Effective cost in credits per 1M tokens with each mix (no cache write). In violet, the cheapest in the scenario.

model Cache-friendly 60/35/5 Balanced 40/20/40 Output-heavy 25/5/70
DeepSeek V4-Flash 8.47 21.04 32.04
GLM 5.3 Flash 4.39 9.31 13.62
MiMo V2.6 Flash 4.95 8.43 11.56
Qwen 3.8 Flash 5.96 12.56 18.37
GLM 5.3 (full) 37.47 81.41 120.1
Kimi K3.1 Pro 88.5 242 375.5
Qwen3.8 Max 79.38 162.5 235.63
DeepSeek V4-Pro 50.27 106.04 155.21
Hy4 32.01 67.12 98.07
a/ output rules

It costs between 2x (MiMo) and 5x (Kimi) the input. Anything that cuts generated tokens cuts your bill more than any other optimization.

b/ cache isn't free everywhere

MiMo charges cache at 2% of input; GLM 5.3 Flash at 20% (7.5x more expensive). With a high hit rate GLM wins; with a medium one, MiMo.

c/ typical vs list

GLM 5.3 Flash has the cheapest list input, but its published typical (1.4 cr) is worse than MiMo's (0.6) and DeepSeek Flash's (1.1). The list doesn't predict usage.

d/ the jump to GLM full

The full GLM 5.3 costs ~11x its Flash (49 vs 5.25 cr input). There's nothing in between: either Flash or eleven times more.

e/ Kimi, the special case

Output at 500 cr/1M, cache write at 100 cr (the only one that charges it) and a typical of 19.9 cr. Only justified if its reasoning makes the difference on your specific task.

f/ Hy4, the overlooked large model

The cheapest of the large models in all three scenarios (32/67/98), with cache at 2.10 cr. If you need a large model with no specific reason for another, start here.

// limits of this guide

  • It compares cost. It doesn't measure quality, speed or stability: your task decides that.
  • The token mixes are assumptions; your real flow may split very differently (measure, section 02).
  • Only Kimi has a published cache write rate; for the others the figure is missing.
  • Rates from the credits panel as of November 2026; the cluster's models are voted on by the community every quarter and may change.