The budget tier of the AI API market moved again this month. A fresh wave of Chinese flash models — DeepSeek V4.1 Flash, GLM-5.3-Flash, Qwen3.8 Flash, Xiaomi MiMo V2.6 Flash — all landed within pennies of each other, all carry million-token context windows, and several are natively multimodal at a price where vision used to be a paid upgrade. There is even a genuinely free tier now.
This is a spec-and-pricing comparison, not a synthetic benchmark. Every figure comes from the live MeshTok catalog and reflects it on 2026-10-11 — reproducible on the pricing page. We do not publish numbers we cannot back up with the catalog.
The sub-$0.20 field
Every model below costs less than $0.20 per million input tokens through the MeshTok unified API at https://meshtok.com/v1:
| Model id | Context | Input $/M | Output $/M | Notable tags |
|---|---|---|---|---|
bigmodel/glm-5.3-flash | 1,000,000 | $0.12 | $0.42 | vision, tools, fast |
deepseek/deepseek-v4.1-flash | 1,000,000 | $0.15 | $0.60 | vision, reasoning, fast |
deepseek/deepseek-v4-flash | 1,000,000 | $0.15 | $0.45 | fast |
ali/qwen3.8-flash | 1,000,000 | $0.15 | $0.45 | vision, tools, fast |
xiaomi/mimo-v2.6-flash | 1,000,000 | $0.15 | $0.30 | vision, audio, fast |
agnes/agnes-2.0-flash | 512,000 | free | free | fast, tools, coding |
agnes/agnes-3.0-flash | 512,000 | free | free | fast, tools, coding |
deepseek/deepseek-v4-flash-vision-exp | 1,000,000 | $0.15 | $0.60 | vision, fast |
One structural note first: the old budget tier is dead. A year ago, sub-$0.20 models meant 8K–32K context and text-only input. Every paid model in this table has a million-token window, and four of the six ship image input natively. The trade-offs have moved from “how much context do I get” to “which capabilities and which output price.”
Per-model breakdown
GLM-5.3-Flash — the cheapest multimodal model in the catalog ($0.12 / $0.42)
Zhipu’s first natively multimodal GLM-5 model, on a 320B-total / 18B-active sparse architecture. At $0.12 input it is the absolute price floor among paid models, and it beats every rival here on both input and output price while including vision. If you want one number to quote for “the cheapest real model in October 2026,” this is it.
MiMo V2.6 Flash — the output-price king ($0.15 / $0.30)
Xiaomi’s omni-modal MoE has the cheapest output price of any 1M-context model in the catalog — 30 cents per million output tokens, well under GLM’s $0.42. It is also the only model in this field with an audio tag: image, audio and video input at flash price. Workloads that generate more output than they consume input (summarization, report generation, chat) should price MiMo first.
DeepSeek V4.1 Flash — reasoning in the budget tier ($0.15 / $0.60)
The new-architecture DeepSeek preview is the only model in this comparison carrying a reasoning tag, making it the default pick when cheap traffic still needs step-by-step thinking (math checks, structured extraction with validation, lightweight agent loops). The trade: the highest output price in the field ($0.60). Its sibling DeepSeek V4 Flash ($0.15/$0.45) covers the same price band without the reasoning and vision tags — relevant only if you want the previous-generation snapshot.
Qwen3.8 Flash — the balanced all-rounder ($0.15 / $0.45)
Alibaba’s newest flash tier pairs a million-token window with native multimodality and a tools tag, at output pricing between MiMo and DeepSeek. If your routing needs one budget model that handles text, images and tool calls without a specialty bias, this is the least opinionated pick.
The free tier: Agnes 2.0 / 2.5 / 3.0 Flash ($0 / $0)
The catalog now carries three free Agnes models with 512K context, OpenAI-compatible endpoints, and tool-calling/coding tags. Zero cost changes the calculus for prototyping, CI pipelines, classification backfills, and anything where quality tolerance is high. The rational setup for many teams: prototype on Agnes, graduate to GLM-5.3-Flash or DeepSeek V4.1 Flash when the workload becomes revenue-bearing. The 512K window (vs 1M on paid rivals) is the main ceiling.
Cost math on a realistic workload
1M input + 0.25M output tokens per day, 30 days:
| Model | Daily | Monthly |
|---|---|---|
| GLM-5.3-Flash | $0.12 + $0.105 = $0.225 | $6.75 |
| MiMo V2.6 Flash | $0.15 + $0.075 = $0.225 | $6.75 |
| Qwen3.8 Flash | $0.15 + $0.1125 = $0.2625 | $7.88 |
| DeepSeek V4.1 Flash | $0.15 + $0.15 = $0.30 | $9.00 |
| Agnes 3.0 Flash | $0 | $0 |
For context: the same workload on GPT-6 Sol costs $135/month and on GPT-6 Astra $675/month (see our GPT-6 pricing guide). The budget tier runs a full production workload for the price of two coffees.
Which one should you pick?
- Lowest possible bill, output-heavy workload → MiMo V2.6 Flash ($0.30 output, omni-modal)
- Lowest input price with vision included → GLM-5.3-Flash ($0.12 input)
- Cheap traffic that still needs reasoning → DeepSeek V4.1 Flash (only
reasoningtag in the field) - No-opinion default that handles text + image + tools → Qwen3.8 Flash
- Prototyping, CI, backfills, zero budget → Agnes 3.0 Flash (free, 512K)
- Need omni input (audio/video) for pennies → MiMo V2.6 Flash or Qwen3.5 Omni Flash ($0.33/$1.98)
All of them are one model string away — switch freely and let traffic decide:
resp = client.chat.completions.create(
model="bigmodel/glm-5.3-flash", # swap: deepseek/deepseek-v4.1-flash, ali/qwen3.8-flash, xiaomi/mimo-v2.6-flash
messages=[{"role": "user", "content": "Hello"}],
)Prices reflect the MeshTok catalog on 2026-10-11; the live source of truth is always the pricing page.