Vision used to be a paid upgrade. In the October 2026 catalog it is a budget-tier feature: five models under $0.16/M input read images natively, one of the best omni-modal options is free, and even the premium Western picks have fallen to a fraction of last year’s vision pricing. If your workload touches screenshots, documents, UI verification, or product photos, the price of seeing has collapsed.
This is a catalog-driven ranking — every price and tag below comes from the live MeshTok catalog on 2026-10-26 (verifiable on the pricing page). We group by price tier and flag which models add audio or video on top of images.
The vision field at a glance
All models below carry the vision tag (native image input) and are callable through one OpenAI-compatible endpoint at https://meshtok.com/v1:
| Model id | Context | Input $/M | Output $/M | Extras |
|---|---|---|---|---|
agnes/agnes-2.0-flash | 512K | free | free | agentic, tools |
bigmodel/glm-5.3-flash | 1M | $0.12 | $0.42 | tools, fast |
xiaomi/mimo-v2.6-flash | 1M | $0.15 | $0.30 | audio input |
ali/qwen3.8-flash | 1M | $0.15 | $0.45 | tools, fast |
deepseek/deepseek-v4.1-flash | 1M | $0.15 | $0.60 | reasoning |
deepseek/deepseek-v4-flash-vision-exp | 1M | $0.15 | $0.60 | experimental vision line |
ali/qwen3.5-omni-flash | 256K | $0.33 | $1.98 | audio + video input |
google/gemini-3.8-flash | 1M | $0.75 | $3.75 | coding, reasoning, tools |
ali/qwen3.5-omni-plus | 256K | $1.04 | $5.95 | audio + video, omni flagship |
stepfun/step-5-preview | 256K | $1.04 | $2.98 | video input, coding flagship |
anthropic/claude-sonnet-5.5 | 1M | $2.00 | $10.00 | reasoning, tools |
anthropic/claude-opus-5.5 | 1M | $4.00 | $20.00 | flagship, reasoning, tools |
openai/gpt-6-astra | 400K | $10.00 | $50.00 | OpenAI flagship |
Budget tier: vision at $0.12–$0.15/M
This is the tier that did not exist a year ago, and it is where most image-volume workloads should live:
- GLM-5.3-Flash ($0.12/$0.42) — the cheapest vision-enabled model in the catalog. Zhipu’s first natively multimodal GLM-5 (320B-total/18B-active sparse) reads images at the same price tier where text-only models sat last year. For screenshot classification, receipt parsing, and content moderation at scale, this is the price floor.
- MiMo V2.6 Flash ($0.15/$0.30) — the only budget model with an
audiotag: images and sound in, at the catalog’s cheapest output price. Voice-note transcription plus screenshot understanding in one model is a genuinely useful combo for support-tooling. - Qwen3.8 Flash ($0.15/$0.45) — Alibaba’s balanced flash: images plus tool-calling, 1M window. The pick when your vision calls also need to trigger function calls.
- DeepSeek V4.1 Flash ($0.15/$0.60) — adds a
reasoningtag to vision: the choice when images need thinking (chart interpretation, diagram QA, math from photos). - Agnes 2.0 Flash (free) — zero dollars, 512K window, image understanding included. For prototyping vision pipelines, CI smoke tests, and backfills, free beats cheap.
Mid tier: the omni-modal step up ($0.33–$1.04/M)
- Qwen3.5 Omni Flash ($0.33/$1.98) — the cheapest true omni model: text, image, audio and audio-video input, built for real-time interaction. Voice assistants and media-understanding pipelines start here.
- Gemini 3.8 Flash ($0.75/$3.75) — Google’s strongest Flash, pairing vision with
codingandreasoningtags on a 1M window. The pick for mixed workloads where the same model must read screenshots and write code from them. - Qwen3.5 Omni Plus ($1.04/$5.95) — the omni flagship tier: highest-quality audio-video understanding when the Flash variant is not enough.
- Step 5 Preview ($1.04/$2.98) — natively takes video input alongside text and images, with a coding-flagship tag — a niche but real edge for video-to-code and media workflows.
Premium tier: vision attached to flagships ($2.00–$10.00/M)
Claude Sonnet 5.5 ($2.00/$10.00) is the screenshot-debugging champion: vision plus reasoning plus tools on a 1M window, and the cheapest way to get vision inside a top-tier agentic stack. Claude Opus 5.5 ($4.00/$20.00) and GPT-6 Astra ($10.00/$50.00) are for the hard end — long agent runs where the model must read the screen, plan, and act for hours. For pure image understanding, both are overkill; their vision is a bonus, not the reason to pay.
What a month of vision traffic costs
Reference workload: 1M input + 0.25M output tokens daily, 30 days (image tokens bill as input — this is where vision costs concentrate):
| Model | Monthly cost | Vision tier |
|---|---|---|
| Agnes 2.0 Flash | $0 | free |
| GLM-5.3-Flash | $6.75 | budget |
| MiMo V2.6 Flash | $6.75 | budget + audio |
| Qwen3.8 Flash | $7.88 | budget + tools |
| DeepSeek V4.1 Flash | $9.00 | budget + reasoning |
| Qwen3.5 Omni Flash | $24.75 | omni |
| Gemini 3.8 Flash | $50.63 | mid + coding |
| Step 5 Preview | $53.55 | mid + video |
| Claude Sonnet 5.5 | $135.00 | premium |
| Claude Opus 5.5 | $270.00 | premium flagship |
| GPT-6 Astra | $675.00 | OpenAI flagship |
The spread is stark: a full month of image-heavy traffic costs $6.75 at the budget floor or $675 at the top of the market — a 100× range for the same fundamental job.
Picking by workload
- Screenshot classification, receipts, moderation at scale → GLM-5.3-Flash ($0.12 input floor)
- Images + voice notes (support tooling) → MiMo V2.6 Flash
- Charts, diagrams, math from photos → DeepSeek V4.1 Flash (vision + reasoning)
- Voice assistants, audio-video understanding → Qwen3.5 Omni Flash → Omni Plus for quality
- Read the screen, then write code → Gemini 3.8 Flash or Step 5 Preview
- Agentic loops with screenshots → Claude Sonnet 5.5 (then Opus 5.5 for the hard end)
- Zero-budget prototyping → Agnes 2.0 Flash
Try it with one request
import base64
from openai import OpenAI
client = OpenAI(
base_url="https://meshtok.com/v1",
api_key="sk-your-MeshTok-key",
)
img = base64.b64encode(open("screenshot.png", "rb").read()).decode()
resp = client.chat.completions.create(
model="bigmodel/glm-5.3-flash", # or mimo-v2.6-flash, gemini-3.8-flash, claude-sonnet-5.5
messages=[{
"role": "user",
"content": [
{"type": "text", "text": "Describe every UI element in this screenshot."},
{"type": "image_url", "image_url": {"url": f"data:image/png;base64,{img}"}},
],
}],
)Prices reflect the MeshTok catalog on 2026-10-26; live figures on the pricing page, per-model detail on each model page. For the text-only budget tier, see Best Cheap AI APIs Under $0.20/M.