GPT-4 API cost 2026: pricing breakdown and how to estimate it

Somewhere between deploying your first GPT-4 API call and receiving your first serious invoice, most teams have the same moment of reckoning: this adds up faster than I thought. The pricing page said $2.00 per million tokens. The bill said otherwise. If that sounds familiar, you are not being naive — you are discovering that gpt 4 api cost is a family of numbers, not a single one, and that the gap between the cheapest and priciest variant in the family is exactly 300x.

This openai gpt-4 api pricing guide covers the GPT-4 API model family: GPT-4, GPT-4 Turbo, GPT-4o, GPT-4o mini, GPT-4.1, GPT-4.1 mini, and GPT-4.1 nano. It is not a general OpenAI pricing guide (we have that), not a per-token calculation walkthrough (we have that too), and not an Azure OpenAI deep-dive. This is the GPT-4 family reference: what each model costs, where legacy pricing is silently draining budgets, and how to close the gap.

What is the GPT-4 API, and why does the family still matter in 2026?

OpenAI launched the original GPT-4 in March 2023. Three years and six model variants later, the GPT-4 API is not a single model. It is a product family that spans an astonishing price range, from $0.10 per million input tokens (GPT-4.1 nano) to $30.00 per million (legacy GPT-4). The family is still the production workhorse for the majority of AI applications built between 2023 and 2026.

Why does it still matter when GPT-5.x exists? Because most production integrations still reference a GPT-4 endpoint. According to McKinsey’s 2025 state of AI report, 88% of organizations report regular AI use in at least one business function, and the majority of those were built on GPT-4-class models. The GPT-4.1 family launched in April 2025 and remains OpenAI’s recommended non-flagship API tier. GPT-5.x is the flagship, GPT-4.1 is the production workhorse.

The finance implication: teams that built on GPT-4 Turbo ($10/$30 per million tokens) in 2024 and have not migrated are now paying 5x more than they need to. The AI ROI case for a migration audit is not subtle.

What is the complete GPT-4 model family pricing in 2026?

Every model in the GPT-4 API family bills on the same token-in, token-out model. You pay separately for input tokens (your prompts, system instructions, injected context) and output tokens (the model’s response). Output costs 2x to 4x more because generation is more compute-intensive than reading.

Here is the full family, with openai gpt-4.1 api pricing and all variants:

Note: Batch API cuts every rate by a flat 50%. Prompt caching saves up to 75% on repeated input blocks (the cached input column above).

Deprecation notes (as of mid 2026):

Three things worth flagging about this table:

What does migrating from GPT-4 Turbo to GPT-4.1 actually save?

Here is the question every engineering manager with a GPT-4 API line item should be asking: what model string is in our model parameter right now? Not what you deployed in 2024 but what is actually running today. If the answer is gpt-4, gpt-4-turbo, or gpt-4o, you have a budget problem that a single pull request can fix.

The GPT-4o api pricing at $2.50/$10.00 per million tokens was competitive when it launched. 

GPT-4.1 is its direct replacement at $2.00/$8.00, 20% cheaper on both input and output, with a nearly 8x larger context window (1M vs 128K tokens), and better instruction-following benchmark scores per OpenAI’s own release notes.

Note: GPT-4o remains callable in the API (it was only removed from the ChatGPT interface in February 2026), but for every team that has not migrated to GPT-4.1, that 20% reduction in AI spend is available on day one, zero engineering effort required beyond changing one string.

For teams still on gpt-4 turbo pricing ($10.00/$30.00 per million tokens), the math is less subtle. Migration to GPT-4.1 saves 80% on input and 73% on output. On a workload running 1 million API calls per month at 2,000 input / 1,000 output tokens per call:

These numbers assume zero change to the product. Same prompts, same workload, same outputs. The only variable is the model string. If your engineering team is spending time on any other cost-reduction initiative before running this migration audit, that is a sequencing problem worth discussing at the next sprint planning.

Find out which GPT-4 model your bill is actually running on CloudZero surfaces AI spend by model, team, and feature, so you can see at a glance whether legacy GPT-4 endpoints are quietly inflating your OpenAI invoice. Toyota, Skyscanner, Grammarly, Duolingo, and Upstart use this to tie every dollar of AI spend back to measurable AI ROI. Average year-one savings: 22%. .

What does migrating from GPT-4 Turbo to GPT-4.1 actually save?

The fastest way to get a wrong number is to multiply your average request size by the listed token price. The second fastest way is to forget that output tokens exist. Here is a more honest framework, the same one finance teams at CloudZero customers use when forecasting new AI features:

1. Measure tokens, not words

A 400-word system prompt is approximately 530 tokens. “System prompt” is already 2 tokens. Your verbose instructions about tone, format, and fallback behavior are all charged every single call. Use OpenAI’s tokenizer at platform.openai.com/tokenizer to measure your actual prompts before estimating, not after.

2. Account for output separately, and be realistic

OpenAI token cost for output is typically 4x the input rate on GPT-4.1 ($8.00 vs $2.00 per million). A customer support reply that averages 200 words is roughly 270 output tokens. Multiply by your daily call volume before committing to a model tier. Unconstrained output, no max_tokens set, is one of the most common sources of bill shock on first deployment.

3. Model the compounding effect of context

Multi-turn conversations and RAG applications inject history and retrieved documents into every prompt. A chat interface that retains the last 10 turns can easily accumulate 8,000 tokens of context before the user types their 11th message. At GPT-4.1 pricing, that single session already costs $0.016 in input tokens before the model responds. See the OpenAI cost per API call guide for the full formula with context accumulation modeled.

4. Apply the real discounts before the final number

Standard pay-as-you-go is the most expensive way to use the GPT-4 API. The Batch API is 50% off all models for non-real-time workloads. Prompt caching is 75% off repeated input blocks. Combining both on a batch document pipeline running GPT-4.1: effective input cost drops from $2.00 to $0.25 per million tokens, an 87.5% reduction. The full optimization playbook covers 14 tactics with worked examples.

Quick reference for common use cases on current 2026 models:

The legacy row is there deliberately. $3,000/month vs $60/month for functionally identical support chatbot behavior. That is a 50x cost difference. The GPT-4.1 nano recommendation for classification is not a quality compromise: for extraction, routing, and tagging tasks, it consistently matches GPT-4 output at 0.3% of the price.

We have seen engineering teams discover this in production, via CloudZero’s per-request cost tracking, and immediately migrate three workloads in a single sprint.

GPT-4 vs GPT-4o vs GPT-4.1: which model do you actually need?

This is the question the pricing table cannot answer by itself. Here is the practical decision framework, organized by what your workload actually does, not by model generation.

You need GPT-4.1 if:

You need GPT-4.1 mini if:

You need GPT-4.1 nano if:

You probably should not use GPT-4o or legacy GPT-4 for new workloads:

GPT-4o at $2.50/$10.00 per million tokens is still callable in the API but was removed from the ChatGPT interface in February 2026. OpenAI gpt-4o api pricing remains available via API, the case for not using it is cost and context, not availability: GPT-4.1 is cheaper, has an 8x larger context window, and benchmarks better.

Legacy GPT-4 ($30/$60) has no caching discount, a tiny 8K context window, and no meaningful quality advantage over GPT-4.1 for any workload we are aware of. If your codebase still contains model: “gpt-4” anywhere, that string is costing you money right now.

What do GPT-4 API costs look like at enterprise scale?

Individual call costs look innocuous. $0.002 per 1,000 input tokens on GPT-4.1. Hard to get alarmed about a fifth of a cent. The place where OpenAI API cost becomes a finance conversation is scale, and the mismatch between what engineering estimated and what the invoice says.

A February 2026 survey of 500 finance leaders by Sapio Research for DoiT found that 79% of organizations experienced AI-related cost overruns in the past 12 months, and only 15% can calculate AI ROI without significant bottlenecks. The pattern is consistent: teams model for current usage, deploy, and then usage grows faster than forecast as features spread through the product.

At $15B+ in managed cloud and AI spend, CloudZero has seen this pattern hundreds of times. A team ships an AI feature, estimates 100K API calls per month, and hits 800K by month three as adoption grows. On GPT-4.1, that is a manageable $1,040/month becoming $8,320/month. On legacy GPT-4, it is $3,000/month becoming $24,000/month, and finance wants to know why the line item grew 8x in a quarter.

The answer to “why did our OpenAI bill 8x?” is almost always a combination of: unmonitored context growth, the wrong model for the task, and missing caching. The answer to avoiding it is granular OpenAI spend tracking that shows cost by feature, by team, and by model before it appears on the invoice. Not after.

How does GPT-4 API cost differ via Azure OpenAI?

Per-token pricing for every GPT-4 family model is identical between Azure OpenAI Service and the direct OpenAI API. GPT-4.1 is $2.00/$8.00 per million tokens on both.

The cost difference is not in the tokens. It is in the infrastructure surrounding them: Azure requires enterprise support plans ($100-$1,000+/month), adds networking and egress overhead, and charges 10% extra for data residency on newer models.

For regulated industries (financial services, healthcare, government) the compliance premium is worth it. For everyone else, the direct OpenAI API is cheaper end-to-end. 

Stop guessing what your GPT-4 API is actually costing you CloudZero’s free cloud and AI cost assessment shows your OpenAI spend broken down by model, team, and feature, before it becomes a quarterly surprise. No code changes required. Most teams find their first optimization opportunity in the first session. You can also take the self guided product tour here.

Frequently asked questions about GPT-4 API cost

Author Spotlight

Lyne Carolyne

Lyne Carolyne has several years of experience in AI and cloud economics and brings that understanding into the content she creates. Outside work, she's an avid explorer.