The Cheapest AI Models in 2026 (Real Per-Token Prices, Ranked)

What’s the cheapest AI model I can actually ship on?

Not the cheapest model on a spreadsheet. The cheapest model that can classify a document, call a tool, read an image, or run an agent without making you pay for retries.

That answer has changed quite a lot. The cheapest usable tier is now below $0.50 per 1 million input tokens. At that price, the API bill is no longer the first problem for many products. The real problem is choosing the right model for the task.

I checked the current provider rate cards and compared the models listed on CompanyFabric. The short answer is simple: GLM-5.3-Flash is my first pick for low-cost extraction and routing. DeepSeek V4.1 Flash is more interesting when you need a 1M context window and open-weight model. Gemini Flash-Lite is the safer choice when vision and Google’s ecosystem matter.

How to judge a cheap AI model

The input price is only one part of the bill. Most models charge separately for input and output tokens, and output can cost several times more than input. If your application reads 3,000 tokens and generates 300, input dominates. If it writes 1,000 tokens after reading 1,000, output becomes much more important.

Cached input matters too. Repeating the same system prompt or documentation can be much cheaper when the provider offers a cache-hit rate. Z.ai lists GLM-5.3-Flash at $0.15 per 1 million input tokens, $0.03 for cached input, and $0.50 for output. DeepSeek V4.1 Flash lists separate cache-hit, cache-miss, peak, and off-peak rates.

Then look at the context window. A cheap model that cannot fit your prompt is not cheap. You pay for splitting the task, managing summaries, and handling more requests. Also check tool calling, structured output, vision, maximum output, availability, and whether the model is preview-only.

Finally, calculate the cost of a completed task. A failed agent call followed by two retries can cost more than one call to a stronger model. This is why I would not rank models only by input price—even if that makes a nice headline.

One correction is worth mentioning. Google’s current pricing page lists Gemini 3.1 Flash-Lite at $0.25 input and $1.50 output, and Gemini 3.5 Flash-Lite at $0.30 input and $2.50 output. The same page lists a 1M context window for Gemini 2.5 Flash, but the page does not expose a context length for the two Flash-Lite entries. Do not invent one in your product catalog.

Cheapest AI models ranked

1. GLM-5.3-Flash: cheapest usable model

GLM-5.3-Flash from Z.ai costs $0.15 per 1 million input tokens and $0.50 per 1 million output tokens. It has a 1M context window, up to 131K output, tool calling, and vision. This is the model I would start with for extraction, classification, simple agents, and routing.

It is cheap, but it is not a toy. That distinction matters. You can use the GLM-5.3-Flash API as the default model and send harder requests elsewhere.

The cheapest model is useful only when it can finish the job.

2. Gemini 3.1 Flash-Lite: cheapest Gemini model

Gemini 3.1 Flash-Lite costs $0.25 per 1 million input tokens and $1.50 per 1 million output tokens. Google positions it for high-volume agentic tasks, translation, and simple data processing, and its pricing supports text, image, and video input.

I would use the Gemini 3.1 Flash-Lite API for classification and extraction at volume when you still need vision. The important detail is that its output price is 10 times the GLM input price, so generation-heavy workloads need a real cost estimate.

3. DeepSeek V4.1 Flash: the open-model price-performance reference

DeepSeek V4.1 Flash costs $0.30 per 1 million input tokens on the peak cache-miss rate and $1.20 per 1 million output tokens. It has a 1M context window, tool calls, JSON output, and vision according to DeepSeek’s model table.

The rates are time-dependent. DeepSeek lists off-peak prices of $0.15 input and $0.60 output, while peak prices are $0.30 and $1.20. A gateway without time-of-day billing should not promise you the off-peak number. DeepSeek also says the old deepseek-v4-flash name is retired and requests are served by V4.1 Flash instead.

For long documents, open-weight deployments, and agents where context matters, this is my strongest alternative to GLM. See the DeepSeek V4.1 Flash API.

4. Gemini 3.5 Flash-Lite: cheap input, expensive output

Gemini 3.5 Flash-Lite costs $0.30 per 1 million input tokens and $2.50 per 1 million output tokens. Google describes it as a cost-efficient model for high-volume agentic tasks, translation, and simple data processing.

The input price is identical to DeepSeek’s peak price, but the output price is more than double. I would choose Gemini 3.5 Flash-Lite when its capabilities or Google integration justify the difference—not because the input number looks cheap.

This is also where the ranking becomes subjective. For short extraction tasks, the input rate matters more. For agents that generate a lot of text, DeepSeek should move ahead.

5. Gemini 3 Flash Preview: large context, preview risk

Gemini 3 Flash Preview costs $0.50 per 1 million input tokens and $3.00 per 1 million output tokens. Google lists a 1M context window and makes it available as a preview model.

That price buys a large context and the Flash tier, but preview status changes the calculation. You can test it through the Gemini 3 Flash Preview API, but I would not make it the foundation of a contract-sensitive production product until the model and pricing settle.

Cheapest model from each provider

If you are locked to one vendor, the answer changes. Here is the practical list from the current catalog. Prices are input / output per 1 million tokens.

The cheapest Claude answer is about 7 times the input price of GLM-5.3-Flash. The cheapest OpenAI answer in this catalog is about 33 times higher. That does not make either provider a bad choice. It means vendor lock-in has a direct cost.

When the cheapest model is the wrong choice

A cheap model is the wrong choice when the task is expensive to get wrong.

Use a stronger model for multi-step reasoning, complex coding, sensitive decisions, or long agent loops where one bad tool call creates several retries. The first request may cost more, but the complete workflow can cost less.

I would not use the cheapest model for every stage of an autonomous agent. Use a low-cost model to classify the request and choose a route. Use a stronger model when the request needs planning. Then use the cheaper model again for formatting, extraction, or a final check.

The rule is simple: route by task, not by reputation—and not by the lowest number in a pricing table.

How to use the cheapest LLM API with one key

The annoying part is not only the model price. Every provider gives you a different account, API key, billing system, rate limit, model name, and failure mode. After you add three or four providers, your application starts carrying a small billing department inside it.

A gateway gives you one OpenAI-compatible endpoint and one prepaid balance. With CompanyFabric, you can compare models, see the per-call price before running a request, and route traffic without rewriting every integration. The platform fee is 5% on top-ups, while bring-your-own-key is available at 0% fee according to the brief.

This is where one API for every model becomes useful. You can send a request to the cheapest model that supports the required capability, then skip models that lack vision, tools, context capacity, or the output format you need.

An automatic route should check at least four things before selecting a model:

1. Does the model support the input type—text, image, audio, or video?
2. Does it support tool calls and structured output?
3. Can its context window fit the complete prompt?
4. Is the expected output cost lower after accounting for retries?

That is more useful than a static “cheapest model” label.

FAQ

What is the cheapest AI model right now?

Among the models in this comparison, GLM-5.3-Flash is the cheapest usable option at $0.15 per 1 million input tokens and $0.50 per 1 million output tokens. It also includes a 1M context window, tools, and vision.

How much does 1 million tokens cost?

It depends on the model and whether the tokens are input or output. At GLM-5.3-Flash’s input price, $1 buys about 6.67 million input tokens before any platform fee; output tokens cost more. DeepSeek also separates cache hits, cache misses, peak hours, and off-peak hours.

What is the cheapest Claude model?

The catalog’s cheapest Claude model is Claude Haiku 4.5 at $1 per 1 million input tokens and $5 per 1 million output tokens. That is a higher price than the cheapest models in this list, but vendor-specific features can still make it the right choice.

Is DeepSeek the cheapest AI?

DeepSeek V4.1 Flash is one of the cheapest strong options, but it is not the cheapest on input. GLM-5.3-Flash is lower at $0.15, while DeepSeek’s peak cache-miss rate is $0.30 and its off-peak rate is $0.15.

What is the cheapest AI model with vision?

GLM-5.3-Flash is the cheapest model in this comparison with vision at $0.15 per 1 million input tokens. Gemini 3.1 Flash-Lite is the next practical option when you want Google’s multimodal API at $0.25 input.

Cheapest model for coding?

For simple coding assistance, extraction from repositories, and routing, I would start with GLM-5.3-Flash. For difficult debugging, architecture decisions, or multi-step code changes, pay for a stronger model—the cost of a wrong patch is higher than the token bill.

What’s next?

The cheapest AI model is not the one with the smallest input price. It is the one that completes your task without expensive retries.

Start with GLM-5.3-Flash for extraction and routing, test DeepSeek V4.1 Flash for long-context work, and keep Gemini Flash-Lite available for multimodal workloads. If you want to compare them through one key, try the cheapest AI models ranked.

Thanks for reading! If you have feedback or suggestions, you can find me here: Twitter/X and LinkedIn.