GPT API Pricing Explained: Cost Comparison of 13 Models per Million Tokens
Prices across 13 GPT-series models vary by up to 100x; choosing the right tier and using caching well are the keys to controlling costs.
Based on the pricing list from uplinkswitch.com, the article lays out the input, cached input, and output prices for 13 GPT-series models and compares them side by side with OpenAI's official list prices. The list shows that between the cheapest GPT-6 Luna and the most expensive GPT-6 Astra, the input price gap is 48x and the output price gap is 100x. Calculated across chat-turn and long-document scenarios, the per-request cost difference between Astra and Luna is about 90x, and model tier directly determines project sustainability. Cached input is about 10% of regular input, saving 90% of costs in repeated-prefix scenarios; the GPT-4o series has no separate cached rate, amplifying its disadvantage in long-context reuse. The article recommends "using the cheapest model that can complete the task" and taking a rational view of third-party platforms' 70%-lower quotes.
OpenAI's API pricing has always been token-based, but the price gap between different models can span two orders of magnitude. A recent pricing list from uplinkswitch.com lays out the input, cached input, and output prices for 13 GPT-series models, compares them side by side with OpenAI's official list prices, and works through the math for two specific requests.
How GPT API Pricing Works
OpenAI bills API calls per million tokens across three tiers: input, cached input, and output. Every model's output price is higher than its input price, reflecting the fact that generating tokens takes more compute than processing them. GPT-4o and GPT-4o mini have no separate cached rate—repeated prefixes are billed at the input price. The cost of a request is simply the token count multiplied by the corresponding model's rate, with no seat fees or subscription fees.
This structure means that when choosing a model, what really matters is "price per million tokens multiplied by business volume," not the absolute cost of a single call. For high-frequency, low-difficulty tasks, the annualized gap between a cheap model and an expensive one can reach dozens of times.
Tokens and Context Caching
A token is the basic unit of text a model processes, roughly a word fragment: in English, one token corresponds to about 0.75 words, while in Chinese, depending on the tokenization method, one character may correspond to 1 to 2 tokens. Per-million-token billing means multiplying the input and output token counts by the corresponding tier's unit price.
Cached input relies on the "prompt caching" mechanism: when multiple requests begin with the same prefix, the server can reuse the computation from the previous request for that prefix instead of reprocessing it, billing at a lower rate. Caching typically has a minimum length threshold (e.g., 1024 tokens) and a lifetime (from minutes to hours), and the cache is invalidated as soon as the prefix changes. So the biggest caching gains come from long system prompts, fixed knowledge bases, and the unchanging history portions of multi-turn conversations.
Full Price Overview of 13 Models
The list covers the full range of GPT models, from flagship to lightweight. The cheapest is GPT-6 Luna, at $0.0625 input and $0.15 output per million tokens; the most expensive is GPT-6 Astra, at $3.00 input and $15.00 output—a 48x gap in input price and a 100x gap in output price.
The middle tiers are densely packed. GPT-6.1 Sol and GPT-6 Sol are both $0.60 input and $3.00 output; GPT-5.5 and GPT-5.4 are both $1.50 input and $9.00 output; GPT-5.6 Sol and GPT-5.6 Terra share the $1.20 input tier, with outputs of $3.00 and $3.60 respectively. GPT-5.3-Codex has a special positioning, at $1.05 input and $8.40 output, clearly priced separately for coding scenarios.
Interestingly, the list presents both its own platform prices and OpenAI's official list prices side by side. Take GPT-6 Astra: the official list price is $10.00 input and $50.00 output, while the list price is about 30% of that—70% lower. GPT-6.1 Sol is likewise 70% lower, and GPT-6 Luna is 37% lower on input and 70% lower on output. This side-by-side presentation makes the channel price gap easy to see, but note that these are third-party platform quotes; actual costs also need to factor in stability, latency, and compliance.
The Difference Between Third-Party Resale and Official List Prices
The "official list price" shown alongside in the list refers to the per-million-token price OpenAI charges developers directly; the "list price" comes from third-party aggregation or resale platforms. Such platforms typically obtain discounts through bulk purchasing, quota pooling, or regional price differences, then resell below the official price to earn the spread. The price gap itself isn't mysterious, but note: third-party quotes may fluctuate with exchange rates, promotions, or remaining quota and are not long-term commitments; requests routed through them introduce extra latency and single-point-of-failure risk, and data processing may not be directly governed by OpenAI's terms. For businesses with high compliance requirements, direct official access remains the baseline option.
Cost Calculation for Two Real Requests
To turn abstract unit prices into tangible numbers, the list designs two typical scenarios: a chat turn (2,000 input + 500 output tokens) and a long-document request (50,000 input + 2,000 output tokens), both calculated at non-cached prices.
| Model | Chat turn (official price) | Chat turn (list price) | Long document (official price) | Long document (list price) |
|---|---|---|---|---|
| GPT-6 Astra | $0.045 | $0.0135 | $0.60 | $0.18 |
| GPT-6.1 Sol | $0.009 | $0.0027 | $0.12 | $0.036 |
| GPT-6 Luna | $0.0005 | $0.0002 | $0.006 | $0.0034 |
A few observations: a normal chat turn on GPT-6 Astra costs 4.5 cents officially, while GPT-6 Luna costs just 0.05 cents—a 90x difference. In the long-document scenario, Astra costs $0.60 per request versus just 0.6 cents for Luna. If you process thousands of long-document requests per day, model choice directly determines whether a project is sustainable.
The Real Impact of Cached Input
When a request repeats a prefix the API has already seen—such as a long system prompt or the same document—the response marks those tokens as cached and bills them at the cached rate. For GPT-6 Astra, for example, cached input is $0.30 versus $3.00 for regular input—90% cheaper. For scenarios that repeatedly carry the same knowledge base or instruction set, the cost reduction from caching is more immediate than switching models.
Models without a cached rate (GPT-4o and GPT-4o mini) bill repeated prefixes at the input price, meaning their real cost disadvantage is amplified in long-context reuse scenarios.
Cache Hit Rate as a Cost Lever
The extent of cache savings depends on the hit rate, not just the cached unit price. If 80% of a request's input hits the cache, the effective input cost is roughly "0.8 × cached price + 0.2 × regular price"; when the hit rate is low, savings shrink dramatically. Hit rate is affected by prefix stability, request intervals (must be within the cache lifetime), and concurrency patterns. Engineering-wise, you can improve the hit rate by putting invariant content first and variable content last to stabilize the prefix. For the GPT-4o series, which has no separate cached rate, this optimization space doesn't exist, so long-context reuse scenarios should prioritize models that support caching.
Usage Selection and Access Methods
The list's selection advice is straightforward: use the cheapest model that can complete the task. Luna suits high-concurrency short tasks like classification and extraction; Sol covers most coding and writing scenarios; Astra is reserved for cases where "a harder answer is worth the price." It's recommended to run the same prompt across all three tiers for comparison rather than choosing by intuition.
On the access side, all listed GPT model IDs can be called through the same OpenAI-compatible SDK—just point base_url to the provider's address and fill in the key. Billing uses a prepaid credit system with a minimum top-up of $10, and each top-up locks in the price for 30 days, with no subscription model.
A Few Takeaways
The value of this pricing list lies not just in the numbers themselves but in showing the structural differences in API costs: within the same family, the input price gap can reach 48x and the output price gap 100x. Many teams focus only on model capability leaderboards when selecting models, overlooking the match between task difficulty and model tier—using a flagship model for classification and extraction can cost dozens of times what's necessary.
On the other hand, the gap between third-party platforms and official list prices (commonly 70% lower in the list) needs to be viewed rationally. Behind the price advantage lie service stability, data compliance, and long-term sustainability; production environments shouldn't look at unit price alone. The truly reliable savings come from two places: choosing the right model tier and making good use of caching.
One-line Summary
Prices across 13 GPT-series models vary by up to 100x; choosing the right tier and using caching well are the keys to controlling costs.
Paragraph Summary
Based on the pricing list from uplinkswitch.com, the article lays out the input, cached input, and output prices for 13 GPT-series models and compares them side by side with OpenAI's official list prices. The list shows that between the cheapest GPT-6 Luna and the most expensive GPT-6 Astra, the input price gap is 48x and the output price gap is 100x. Calculated across chat-turn and long-document scenarios, the per-request cost difference between Astra and Luna is about 90x, and model tier directly determines project sustainability. Cached input is about 10% of regular input, saving 90% of costs in repeated-prefix scenarios; the GPT-4o series has no separate cached rate, amplifying its disadvantage in long-context reuse. The article recommends "using the cheapest model that can complete the task" and taking a rational view of third-party platforms' 70%-lower quotes.
Related articles
Why Does the Tesla Roadster Need to Fly? A Deep Dive into Musk's Wild Demo
Tesla's Roadster launch showed wall-climbing, upside-down driving, and flying stunts—but The Vergecast notes the demo car may lack a battery, the Rocket version is invite-only and not street-legal, and the website terms don't guarantee delivery. After nine years of delays, why does this car still need to exist?
AI Is Not a Replacement: Human-AI Collaboration Through the Lens of Factory Root Cause Analysis
Starting from factory root cause analysis, this article explains how AI accelerates evidence chain construction rather than replacing human experts, exploring the pragmatic implementation logic and business value of 'AI supporting people rather than replacing them.'
AI Agent Decision Auditing: Building an Evidence Chain in Regulated Workflows
Starting from a Reddit discussion, this article analyzes three core issues in AI agent decision auditing within regulated workflows: what fields audit records should contain, how to handle feedback latency for erroneous decisions, and how to continuously calibrate LLM judge consistency with human judgment.