# Cheapest LLM API in 2026: 12 Open-Weight Models Compared

The cheapest LLM API depends on the model. We priced 12 open-weight models, from Qwen and DeepSeek to Kimi and GLM, across the makers, DeepInfra and SayGM, with the cost of 100M tokens for each and which model fits which job.

_Source: https://saygm.com/blog/cheapest-llm-api-in-2026-12-open-weight-models-compared_
he cheapest LLM API in 2026 depends on the model. SayGM is the cheapest paid route for six of the 12 open-weight models it carries, including Qwen3.8 27B at $9 per 100 million tokens and Kimi K3 at about half what any major provider charges. On the other six, DeepInfra is cheaper, often on a promotional rate, and SayGM's price is within a few dollars or buys you a sealed route nobody can read.

The table below puts every open-weight text model on SayGM next to its maker's own price and the cheapest major provider, with the cost of 100M tokens for each. Under it, each model family has a short note and a link to its full pricing breakdown.

## Table of Contents

- [Cheapest LLM API by model](#cheapest-llm-api-by-model)
- [Which model should you use?](#which-model-should-you-use)
- [Qwen3.8 27B: $9 per 100M tokens](#qwen38-27b-9-per-100m-tokens)
- [DeepSeek V4.1 Flash: close to a tie](#deepseek-v41-flash-close-to-a-tie)
- [Kimi: half the major-provider price](#kimi-half-the-major-provider-price)
- [GLM: where DeepInfra leads](#glm-where-deepinfra-leads)
- [How to read LLM API prices](#how-to-read-llm-api-prices)
- [Cheapest LLM API questions](#cheapest-llm-api-questions)
- [About SayGM](#about-saygm)

## Cheapest LLM API by model

![Best priced LLM API Provider for each open sourced model](https://assets.saygm.com/blog/2026/10/850917ab-2ce0-4aa9-a4b4-29c07542c707-cheapest-llm-provider-per-model-2160x3894.png)

Prices are per million tokens, input / output, checked 14 October 2026. Each model section further down breaks the same figures out by provider. The maker's price is what the lab charges on its own API. The cheapest major provider is the lowest price among the large hosts listed on [OpenRouter](https://openrouter.ai/models): DeepInfra, Together, Fireworks, Novita, Cloudflare, Groq, Cerebras, Amazon Bedrock, Google and the makers themselves. SayGM's price is its cheapest route for that model. "Sealed" means SayGM runs that route inside a hardware-attested enclave where nobody can read the prompt. The last column is the cost of 100M tokens at an 80/20 input-to-output mix.

| Model | Maker's price | Cheapest major provider | SayGM | Sealed on SayGM | SayGM per 100M |
|---|---|---|---|---|---|
| Qwen3.8 27B | $0.43 / $2.55 (Alibaba) | $0.15 / $1.88 (DeepInfra, promo) | **$0.045 / $0.27** | No | **$9** |
| DeepSeek V4.1 Flash | $0.30 / $1.20 | $0.14 / $0.42 (DeepInfra, promo) | $0.13 / $0.54 | Yes, $0.27 / $1.08 | $21 |
| Kimi K3 | $3.00 / $15.00 | $2.85 / $14.25 (DeepInfra) | **$1.50 / $7.50** | Yes, $2.85 / $14.25 | **$270** |
| Kimi K2.6 | $0.95 / $4.00 | $0.75 / $3.50 (DeepInfra) | **$0.48 / $2.71** | Yes | **$92** |
| GLM 5.3 | $1.40 / $4.40 | $0.56 / $2.50 (DeepInfra, promo) | $0.78 / $2.46 | Yes, $1.26 / $3.96 | $112 |
| GLM 5.3 Flash | $0.15 / $0.50 | $0.075 / $0.25 (DeepInfra, promo) | $0.13 / $0.44 | Yes, $0.14 / $0.45 | $19 |
| Gemma 4 31B | Free tier (AI Studio) | $0.09 / $0.34 (DeepInfra) | $0.11 / $0.35 | Yes | $16 |
| gpt-oss-20b | Not sold by OpenAI | $0.03 / $0.14 (DeepInfra) | $0.062 / $0.27 | No | $10 |
| Mistral Nemo | $0.15 / $0.15 | $0.029 / $0.03 (DeepInfra) | $0.023 / $0.093 | Yes | $4 |
| Nemotron 3 Nano Omni 30B | Not sold by NVIDIA | Free variant only (OpenRouter) | **$0.023 / $0.093** | Yes | **$4** |
| MiMo V2.6 Pro Ultraspeed | $4.35 / $8.70 (Xiaomi) | Xiaomi only | **$3.92 / $7.83** | Yes, same price | **$470** |
| Ornith 1.5 397B | No public list price | Not listed | **$0.70 / $2.20** | Yes, same price | **$100** |

Bold marks the models where SayGM is the cheapest route of the three. DeepInfra is the cheapest major provider on every model it serves here. Sealed means nobody can read the prompt, including the host and SayGM, so on a sealed row the fair comparison is another sealed route; our [guide to private inference](https://saygm.com/blog/what-is-private-inference) explains how that's verified. "Promo" marks a provider rate OpenRouter currently shows as discounted; those tend to change without notice. Live SayGM rates for every model are on the [pricing page](https://saygm.com/#pricing). Each model's full pricing post was written earlier, so its figures reflect prices on the date at the top of that post. The table above is current as of 14 October 2026. For closed models, see our [OpenAI API pricing](https://saygm.com/blog/openai-api-pricing) and [Gemini API pricing](https://saygm.com/blog/gemini-api-pricing) breakdowns.

## Which model should you use?

![Grid showing which cheap Open Source LLM model is best for which task](https://assets.saygm.com/blog/2026/10/0e7c1b82-0506-4460-8bff-81fa591a50dc-which-cheap-llm-to-use-selection-2400x2848.png)

| Model | Use it for | Cheapest per 100M |
|---|---|---|
| Qwen3.8 27B | Everyday coding and general work at the lowest price that still holds up; takes text, images and video | $9 |
| DeepSeek V4.1 Flash | High-volume steps: extraction, classification, routing, summarising, the cheap calls inside an agent loop | $20 |
| Kimi K3 | Hard agentic and coding work where you'd otherwise reach for a frontier model | $270 |
| Kimi K2.6 | Coding agents on a budget, and the cheapest sealed Kimi | $92 |
| GLM 5.3 | Coding, especially if you also use Z.ai's Coding Plan | $95 |
| GLM 5.3 Flash | Fast, cheap GLM for simple tasks | $11 |
| Gemma 4 31B | Image understanding on a small budget, privately | $14 |
| gpt-oss-20b | Small tool-calling jobs | $5 |
| Mistral Nemo | Very cheap, simple text tasks | $3 |

The last column is the cheapest paid route from any provider in the main table, at the same 80/20 mix.

## Can you get privacy at the same price?

On a standard route, whoever runs the GPU can read your prompt. On a sealed route the model runs inside a hardware-attested enclave, so nobody can, including the host and SayGM. On several models that costs nothing extra, or close to it:

- **Kimi K2.6:** the sealed route is $92 per 100M tokens, cheaper than any standard route we found.
- **MiMo V2.6 Pro Ultraspeed and Ornith 1.5:** sealed and standard are the same price.
- **GLM 5.3 Flash, Mistral Nemo and Gemma 4:** sealed costs $1 to $2 more per 100M tokens than the cheapest standard route.

On DeepSeek V4.1 Flash, GLM 5.3 and Kimi K3, the sealed route costs about twice the cheapest standard one. Whether that's worth paying depends on what's in your prompts. Our [guide to private inference](https://saygm.com/blog/what-is-private-inference) explains how the attestation works.

## Qwen3.8 27B: $9 per 100M tokens

Qwen3.8 27B is the cheapest capable model on this page, at $0.045 input and $0.27 output on SayGM, or $9 per 100M tokens. That's less than a fifth of DeepInfra's promotional rate. It's a 27B dense model with Apache 2.0 weights. DeepInfra charges $1.88 per million output tokens against SayGM's $0.27, which is where most of the gap comes from. Our [Qwen3.8 27B pricing breakdown](https://saygm.com/blog/qwen-pricing) compares every provider for it.

| Route | Input / output | Per 100M |
|---|---|---|
| **SayGM** | $0.045 / $0.27 | $9 |
| DeepInfra (promo) | $0.15 / $1.88 | $50 |
| Alibaba Cloud (maker) | $0.43 / $2.55 | $85 |
| Novita | $0.42 / $3.00 | $94 |
| Cloudflare | $0.45 / $3.20 | $100 |
| Cerebras | $0.99 / $1.49 | $109 |

## DeepSeek V4.1 Flash: close to a tie

DeepSeek V4.1 Flash is close to a tie: $21 per 100M tokens on SayGM against $20 on DeepInfra's promotional rate, and both are well under DeepSeek's own $48. DeepSeek is also the one maker in this table that changes its own price by time of day. The [DeepSeek pricing post](https://saygm.com/blog/deepseek-pricing) covers peak and off-peak rates and which version to run.

| Route | Input / output | Per 100M |
|---|---|---|
| DeepInfra (promo) | $0.14 / $0.42 | $20 |
| **SayGM** | $0.13 / $0.54 | $21 |
| Cloudflare | $0.10 / $0.90 | $26 |
| Novita (promo) | $0.20 / $0.78 | $31 |
| SayGM sealed | $0.27 / $1.08 | $43 |
| DeepSeek (maker) | $0.30 / $1.20 | $48 |
| Together | $0.30 / $1.20 | $48 |
| Fireworks | $0.30 / $1.20 | $48 |

## Kimi: half the major-provider price

Kimi K3 shows the biggest gap on the page. SayGM is $1.50 input and $7.50 output, against $3.00 and $15.00 from Moonshot directly and $2.85 and $14.25 from DeepInfra, so 100M tokens costs $270 instead of $513 or more. Kimi K2.6 runs only on a sealed route on SayGM and is still the cheapest of the three. Our [Kimi K2.6 and K3 pricing comparison](https://saygm.com/blog/kimi-api-pricing) has the full provider list.

| Model | Route | Input / output | Per 100M |
|---|---|---|---|
| Kimi K3 | **SayGM** | $1.50 / $7.50 | $270 |
| Kimi K3 | SayGM sealed | $2.85 / $14.25 | $513 |
| Kimi K3 | DeepInfra | $2.85 / $14.25 | $513 |
| Kimi K3 | Moonshot AI (maker) | $3.00 / $15.00 | $540 |
| Kimi K3 | Together | $3.00 / $15.00 | $540 |
| Kimi K3 | Fireworks | $3.00 / $15.00 | $540 |
| Kimi K3 | Amazon Bedrock | $3.00 / $15.00 | $540 |
| Kimi K2.6 | **SayGM sealed** | $0.48 / $2.71 | $92 |
| Kimi K2.6 | DeepInfra | $0.75 / $3.50 | $130 |
| Kimi K2.6 | Novita | $0.80 / $3.40 | $132 |
| Kimi K2.6 | Moonshot AI (maker) | $0.95 / $4.00 | $156 |
| Kimi K2.6 | Cloudflare | $0.95 / $4.00 | $156 |

## GLM: where DeepInfra leads

GLM is where DeepInfra leads. Its promotional FP4 rate for GLM 5.3 is $95 per 100M tokens against SayGM's $112, and GLM 5.3 Flash is $11 against $19. Both SayGM routes are well under Z.ai's own price of $200 and $22. The [GLM 5.3 pricing post](https://saygm.com/blog/glm-pricing) also covers Z.ai's Coding Plan, which can beat per-token pricing for steady coding use.

| Model | Route | Input / output | Per 100M |
|---|---|---|---|
| GLM 5.3 | DeepInfra FP4 (promo) | $0.56 / $2.50 | $95 |
| GLM 5.3 | Novita (promo) | $0.70 / $2.20 | $100 |
| GLM 5.3 | SayGM | $0.78 / $2.46 | $112 |
| GLM 5.3 | Alibaba Cloud | $1.19 / $3.74 | $170 |
| GLM 5.3 | SayGM sealed | $1.26 / $3.96 | $180 |
| GLM 5.3 | Z.ai (maker) | $1.40 / $4.40 | $200 |
| GLM 5.3 | Together / Fireworks | $1.40 / $4.40 | $200 |
| GLM 5.3 Flash | DeepInfra FP4 (promo) | $0.075 / $0.25 | $11 |
| GLM 5.3 Flash | Novita (promo) | $0.084 / $0.28 | $12 |
| GLM 5.3 Flash | SayGM | $0.13 / $0.44 | $19 |
| GLM 5.3 Flash | SayGM sealed | $0.14 / $0.45 | $20 |
| GLM 5.3 Flash | Z.ai (maker) | $0.15 / $0.50 | $22 |
| GLM 5.3 Flash | Together / Fireworks | $0.15 / $0.50 | $22 |

## How to read LLM API prices

The per-token rate in a table is a starting point, and a handful of things move the real bill.

Cache hits come first. Most providers charge a fraction of the input rate for tokens they've seen recently, and agents and coding tools resend the same context on every turn, so a provider's cache-read price can matter more than its headline input price.

Your own input-to-output mix matters nearly as much. Some providers price output at 6 times input and others at 40 times, so a route with a tiny input price and an expensive output price can turn out to be the most expensive one once you price a real workload. That's why every total on this page uses the same 80/20 mix.

Then there's precision. Several of the cheapest routes serve models at FP4, a heavily compressed format; DeepInfra's GLM 5.3 rate above is one. Compressed models are cheaper to run and usually close in quality, though not identical, so test on your own prompts before switching.

Routers also add a fee on top of the provider's rate. OpenRouter charges 5.5% on card credit purchases and 5% on crypto, which is enough to flip close comparisons like DeepSeek V4.1 Flash. SayGM charges no credit fee.

Finally, promotions don't last. Four of the cheapest major-provider prices in this table are currently discounted, so check them again before committing a budget to one.

## Cheapest LLM API questions

### What is the cheapest LLM API right now?

For a capable general model, Qwen3.8 27B on SayGM at $0.045 input and $0.27 output, or $9 per 100M tokens. For tiny models, Mistral Nemo and Nemotron 3 Nano Omni are both about $4 per 100M tokens on SayGM, and gpt-oss-20b is about $5 on DeepInfra.

### What is the cheapest AI API for coding?

Qwen3.8 27B is the cheapest strong coding model in this table; on Alibaba's own evaluations it scores above Claude Opus 4.6 Max on SWE-bench Pro, as covered in our [Qwen pricing post](https://saygm.com/blog/qwen-pricing). For heavier agentic coding, Kimi K3 on SayGM is half the price of the next cheapest major provider.

### Is there a free LLM API?

Google offers Gemma 4 on a rate-limited free tier in AI Studio, and OpenRouter lists free variants of some models, including Nemotron 3 Nano Omni. Free endpoints suit testing rather than anything you're shipping, because they're rate-limited and can disappear without notice.

### Which is the cheapest LLM API provider overall?

There isn't one across every model. DeepInfra is the cheapest major provider on every model it serves in this table. Of the nine models both carry, SayGM is cheaper on three: Qwen3.8 27B (by more than five times), Kimi K3 and Kimi K2.6. On the three DeepInfra doesn't serve, SayGM is the only paid route for Nemotron 3 Nano Omni and the only route for Ornith 1.5, and under Xiaomi's own price for MiMo. Pick by model, then compare totals at your own input-to-output mix.

Every SayGM price on this page is live in the [model catalog](https://saygm.com/models). Point your existing OpenAI client at `api.saygm.com/v1`, pick a model, and say gm to AI.

## About SayGM

SayGM is a drop-in inference gateway for teams who don't want to just take a company's word that their prompts are private. Route to a confidential model and your prompt is sealed inside hardware-verified execution, where not even SayGM can see it. That's not a policy, it's provable. Swap in your existing OpenAI, Anthropic, or Gemini code and you're covered in minutes, at transparent, published rates with no hidden markup.

Say gm to AI at [saygm.com](https://saygm.com).

[Website](https://saygm.com) | [Twitter](https://x.com/say_gm_) | [Discord](https://discord.com/channels/799672011265015819/1343950080465698836) | [Blog](https://saygm.com/blog) | [Medium](https://medium.com/@say_gm_) | [Docs](https://docs.saygm.com/)