Cheapest AI APIs in 2026: Price vs Capability

Summary

The cheapest AI API in 2026 depends on what you mean by “cheap.” A low input-token rate can look impressive and still lose on the final bill if the model generates long outputs, needs repeated retries, misses caches, or requires a higher reasoning setting to complete the task.

For general-purpose paid APIs from major providers, the current raw price floor in our shortlist is Gemini 2.5 Flash-Lite at $0.10 per million input tokens and $0.40 per million output tokens. Mistral Small 4 is close at $0.15/$0.60, while OpenAI's GPT-5.6 Luna costs $0.20/$1.20 and offers a much stronger current independent capability signal than the ultra-cheap baseline models.

For teams that need more intelligence without moving into flagship pricing, Gemini 3.7 Flash is our best mainstream price-performance pick at its current introductory $0.75/$3.75 rate through December 31, 2026. DeepSeek's V4 family can be even cheaper for schedulable workloads because its off-peak rates are 50% below peak.

Pricing was rechecked on August 31, 2026. This guide excludes free-tier quotas, temporary signup credits, specialized classifiers/embeddings/OCR, and third-party router markups so the comparisons remain useful for production planning.

Quick Take

• Lowest raw price: Gemini 2.5 Flash-Lite at $0.10 input / $0.40 output per 1M tokens.
• Best low-cost intelligence: GPT-5.6 Luna at $0.20/$1.20 with a 1.05M-token context window.
• Best mainstream value: Gemini 3.7 Flash at $0.75/$3.75 through December 31, 2026.
• Best schedulable economics: DeepSeek V4 Flash at $0.22/$0.66 off-peak.

The price table that actually matters

The table below uses each provider's current standard paid rate unless a condition is explicitly labeled.

Why it matters: the cheapest row is not automatically the best value. A model that finishes in one pass can cost less than a cheaper model that needs several attempts.

A realistic monthly example: 1B input + 100M output tokens

To make the rates tangible, we calculated an illustrative workload with 1 billion uncached input tokens and 100 million output tokens per month. This is not a forecast of anyone's bill; it is identical arithmetic applied across providers.

The spread is enormous: the same token volume ranges from $140 to $15,000 before caching, batch discounts, tool fees, retries, regional pricing or long-context surcharges.

But this is exactly where a simple price ranking becomes dangerous. Those models do not provide the same capability, latency, reasoning depth, output behavior or reliability.

Best cheap API for raw volume: Gemini 2.5 Flash-Lite

If the task is straightforward classification, extraction, rewriting, routing, translation, tagging or other high-volume processing, Gemini 2.5 Flash-Lite sets the lowest mainstream paid token floor in this comparison.

Google charges $0.10 input / $0.40 output, and its Batch API halves those token rates to $0.05/$0.20. On our 1B-input + 100M-output scenario, that would reduce the illustrative token bill from $140 to $70 for work that can tolerate asynchronous processing.

The trade-off is capability. This is an older efficiency model, not a frontier reasoning model. Use it when the work is easy enough that a stronger model would be wasted.

Best low-cost intelligence: GPT-5.6 Luna

GPT-5.6 Luna is the most interesting model in the low-cost tier because its price is close to the raw-cost leaders while its current capability evidence sits in a very different class.

OpenAI now charges $0.20 input / $1.20 output, with cached input at $0.02. Luna has a 1.05M-token context window, 128K maximum output, reasoning controls, image input and the same broad Responses API tool surface as the rest of the GPT-5.6 family.

Artificial Analysis' GPT-5.6 release evaluation scored Luna at 51 on its Intelligence Index, while reporting roughly 80% lower cost per task than GPT-5.6 Sol in that evaluation. That is not proof that Luna beats every cheap model, but it is strong evidence that “nano-tier price” no longer means “toy model.”

The caveat is long context. Once input exceeds 272K tokens, OpenAI applies 2× input and 1.5× output pricing to the full request.

Best mainstream price-performance: Gemini 3.7 Flash

For teams that need stronger agentic and reasoning capability without premium-model economics, Gemini 3.7 Flash is our current default value pick.

Google's introductory rate is $0.75 input / $3.75 output per 1M tokens through December 31, 2026. The model has a 1M-token context window and supports text, image, audio and video input.

Independent evidence is unusually favorable for the price. Artificial Analysis scored Gemini 3.7 Flash at 56 at high reasoning and measured roughly 324 output tokens per second. It also placed the model on its intelligence-versus-time Pareto frontier.

That makes Gemini 3.7 Flash more than “cheap Gemini.” It is a credible production option for workloads where both quality and throughput matter.

The recommendation is time-sensitive: Google's published standard rate doubles on January 1, 2027 unless the company changes the schedule.

Best if you can schedule workloads: DeepSeek V4 Flash and V4 Pro

DeepSeek prices V4 Flash and V4 Pro differently depending on the time of day.

For V4 Flash, cache-miss input/output costs are $0.22/$0.66 off-peak and $0.44/$1.32 at peak. Cache-hit input falls as low as $0.007 off-peak.

For V4 Pro, the corresponding rates are $0.66/$1.98 off-peak and $1.32/$3.96 at peak.

This pricing structure creates an obvious opportunity for CI jobs, background agents, data processing, evaluation runs and batch-like workloads that do not need to execute immediately.

DeepSeek also supports a 1M context window and up to 384K output across the V4 lineup. V4 Pro is the better fit when the workload requires more reasoning; V4 Flash is the economic choice for simpler or higher-volume tasks.

Mistral Small 4 is the quiet raw-cost contender

Mistral Small 4 deserves a place in cost discussions because its standard API rate is only $0.15 input / $0.60 output, and Mistral offers 50% lower batch pricing plus up to a 90% cached-input discount for repeated prompts.

Artificial Analysis gives the reasoning version a much lower intelligence score than today's frontier models, so we would not treat its price as evidence that it can replace Luna, Gemini 3.7 Flash or a premium model on difficult tasks.

Its value is different: inexpensive multimodal processing, good throughput, open weights and a predictable low-cost API for workloads that do not need frontier reasoning.

When $2/$6 is actually a good deal

At first glance Grok 4.6 and Qwen3.8-Max at roughly $2/$6 look expensive next to Luna or DeepSeek. In the frontier-adjacent tier, however, they can be highly competitive.

Grok 4.6 targets coding, agentic tasks and knowledge work, with a 500K context window and a $0.50 cached-input rate below 200K prompt tokens. The catch is important: at 200K prompt tokens or more, its rates double to $4/$12.

Qwen3.8-Max offers a 1M context window and strong current web/full-stack evidence. Alibaba's international list rate is $2/$6 in Singapore, while some global regional deployments list lower base pricing. Region therefore belongs in any production cost model.

These models make sense when a workload needs substantially more capability than the cheapest tier but cannot justify Opus/Fable economics.

Caching and batch can matter more than the model's list price

Three cost levers routinely change the ranking:

• Caching: repeated system prompts, repository context or documents can cut input cost dramatically. Luna falls from $0.20 to $0.02 for cached input; Sonnet 5 from $2 to $0.20; Grok 4.6 from $2 to $0.50 below its long-context threshold.
• Batch: Google, OpenAI, Anthropic and Mistral all offer material asynchronous discounts on supported workloads. OpenAI and Anthropic describe 50% batch discounts; Google's published batch rates are also generally half standard token rates for the models shown here.
• Scheduling: DeepSeek's off-peak pricing is 50% below peak with no need to change model.

A production team should therefore model effective cost, not just copy the two numbers from a pricing page.

Choose by workload, not provider

For a serious production system, the best architecture may use several of these at once. Easy traffic can go to a cheap model; uncertain or high-value requests can escalate to a stronger one.

Methodology and limitations

This guide compares direct provider API list prices in USD per million tokens as of August 31, 2026. We prioritize current first-party pricing pages and model documentation.

We do not include free quotas in the ranking because they do not scale predictably. We also exclude specialized classifier, embedding, transcription, OCR and image-generation pricing from the main table because their units and workloads are not comparable to general-purpose text/multimodal LLM calls.

The monthly example assumes uncached input unless stated otherwise and does not include tool-call fees, cache storage, retries, regional taxes, enterprise discounts or the extra token consumption produced by different reasoning settings.

Most importantly, token price is not cost per completed task. Models can use different numbers of reasoning/output tokens and achieve different success rates on the same job.

What to watch next

The biggest near-term pricing trigger is January 1, 2027, when Google's published Gemini 3.7 Flash introductory rates are scheduled to double.

OpenAI also describes GPT-5.6 Sol's current lower rate as promotional and says it is available at least through November 21, 2026. DeepSeek's peak/off-peak scheme and Alibaba's regional pricing can also change the optimal routing strategy without any model release.

That is why we recommend checking the live AI World Scope Price Tracker before committing a high-volume workload.

AI World Scope Take

The API price war has moved beyond “who has the cheapest model.”

The more useful question is now: what is the least expensive model that clears the quality threshold for this specific workload?

At the raw floor, Gemini 2.5 Flash-Lite and Mistral Small 4 are exceptionally cheap. For general-purpose intelligence, GPT-5.6 Luna changes the economics. For stronger high-throughput reasoning, Gemini 3.7 Flash is difficult to ignore at its current promotional rate. And for workloads that can move in time, DeepSeek's off-peak model introduces scheduling as a new cost-control lever.

The winning production strategy is therefore increasingly model routing: cheap models for easy work, stronger models for hard work, and aggressive use of caching, batch processing and workload scheduling around both.