Open Source LLM API Pricing: Kimi, GLM, Qwen and DeepSeek (2026)

Open-weight model APIs cost a fraction of frontier models. GLM-5.3-Flash runs at $0.10 per million input tokens against Claude Opus 5 at $5.00, roughly 50 times cheaper, while scoring 57 on the Artificial Analysis index against the frontier's 60. But running these models yourself is usually more expensive than either, unless you already own the hardware.

Key takeaways

• GLM-5.3-Flash is $0.10/1M input, the cheapest capable open model, and it is MIT licensed.
• Kimi K3 is $2.30 blended, DeepSeek V4 Pro $0.70, Qwen3.8-27B $0.40 (Artificial Analysis blended rates).
• Claude Opus 5 is $5 in / $25 out; GPT-5.6-sol $4/$20; Gemini 3.7 Flash $0.75/$3.75 on promotional pricing.
• A $20 monthly subscription buys roughly 200 million input tokens at open-model API rates, so for normal chat use the subscription is not the expensive option.
• Self-hosting frontier-class open models is not cheap: Kimi K3 is 1.56TB of weights and GLM-5.3 needs around 8 GPUs.

Open-weight API pricing, verified September 2026

These are blended rates from Artificial Analysis, which weights input and output together:

Kimi's own listed rates are $3.00 input, $0.30 cached, $15.00 output per million tokens, with a 1M token context window.

Closed frontier model pricing

For comparison, verified on each vendor's own documentation:

Worth noting the small closed models undercut most open ones: GPT-5.6-luna at $0.20 input and Gemini 2.5 Flash-Lite at $0.10 are competitive with GLM-5.3-Flash. The open-weight advantage is largest at the capable end, not the cheap end.

Discounts that change the maths: Anthropic's Batch API is 50% off both directions, and cached input reads are 0.1x the base rate. OpenAI cached input is 10% of standard, with Batch and Flex around 50% off. Google's Batch is also 50% off. If your workload tolerates delay or repeats context, these matter more than switching models.

What this actually costs you

Abstract per-token pricing is hard to reason about, so here is the arithmetic in a form you can redo.

A $20 monthly subscription (ChatGPT Plus, or Claude Pro at $17 annually) against GLM-5.3-Flash at $0.10 per million input tokens buys you roughly 200 million input tokens.

For context, a million tokens is about 750,000 words. Two hundred million tokens is more text than most people process in a year of heavy chat use.

The honest conclusion: for individual conversational use, the subscription is not the expensive part. You are paying for convenience, apps, web browsing and multimodal features, not for compute you could get cheaper elsewhere.

Open-model APIs win decisively when you are running programmatic or batch workloads: classification at volume, document processing, agent loops, or anything where a script makes thousands of calls. That is where a 50x price difference compounds into real money.

What about running them yourself?

This is where expectations usually break.

The models scoring near frontier are enormous. Kimi K3 is 1.56TB of weights. GLM-5.3 needs roughly 8 GPUs, at 756GB in FP8 or 1.5TB in BF16. DeepSeek V4 Pro is 893GB. None of these run on a workstation, let alone a laptop.

What does run locally is the smaller tier:

So there are three tiers, not two: cheap open APIs, capable local models that are a step below frontier, and datacenter-only open models you can technically download but not run.

Self-hosting beats an API only when you already own idle GPUs, or when data residency requires it. Buying hardware to save on API costs almost never pays back for an individual.

The licence trap in API pricing

Cheap API access does not always mean permissive licensing, and the two get conflated constantly.

GLM-5.3-Flash is MIT, verified on its model card. GLM-5.3 is not: its HuggingFace licence tag reads other, a custom Z.ai licence. Several outlets describe both as MIT.

Kimi K3 is open-weight with a revenue gate. Its own LICENSE states that if you operate a Model-as-a-Service business and your group revenue exceeds $20 million over any consecutive 12 months, you must enter a separate agreement with Moonshot AI before using the software.

For most readers neither restriction bites. But if you are building a product on these models, read the licence rather than the marketing.

Genuinely unrestricted: Qwen3.8-27B and Gemma 4 are Apache-2.0, DeepSeek V4 Pro and GLM-5.3-Flash are MIT, Muse Glimmer is Apache-2.0.

Which should you actually use?

For chat and everyday work: a $20 subscription. The arithmetic above shows you are not overpaying, and you get browsing, apps and multimodal features an API does not include.

For programmatic volume: GLM-5.3-Flash at $0.10 if you want capability per dollar, or Gemini 2.5 Flash-Lite at $0.10 if you prefer a major vendor's infrastructure.

For agentic and coding workloads: GLM-5.3 at $0.90 or DeepSeek V4 Pro at $0.70, both of which score well on terminal and software-engineering benchmarks.

For privacy or offline work: run Qwen3.8-27B or Muse Glimmer locally. Our guide to running AI locally with Ollama covers the setup.

For building a product: check the licence first. Apache-2.0 and MIT models remove the question entirely.

The verdict

Open-weight APIs are genuinely cheap, and at the capable end they are dramatically cheaper than frontier models for similar benchmark scores.

But the framing that open models save you money is only true for the right workload. If you are chatting, a subscription is fine and probably better. If you are running scripts at volume, the 50x gap is real and worth taking. And if you were planning to self-host something frontier-class, check the weight sizes before budgeting: most of these need a datacenter, not a desktop.

If you would rather run a model on your own GPU, the best local coding model by VRAM covers what fits at 8GB, 16GB and 24GB.

For a full roundup of which models are worth using and their licences, see the best open source LLMs in 2026.

Related reading: how to run AI locally with Ollama and local AI versus paid subscriptions.