How Expensive Is Each OpenAI Model? (2026 Pricing Comparison)
Updated 27 October 2026. OpenAI released GPT-6.1 Terra on 26 October: $1 input / $0.05 cached input / $5 output per million tokens, with a 1,050,000-token context window and 128,000 maximum output. It is a fast option for visual reasoning, computer use and natural writing. GPT-6.1 Sol and GPT-6 Astra remain better candidates for the most complex coding; Terra does not replace every frontier model. See OpenAI’s GPT-6.1 Terra announcement and our current ChatGPT model guide.
OpenAI now offers models at radically different price points. Standard short-context input pricing ranges from $0.05 per million tokens for GPT-5 nano to $30 per million tokens for GPT-5.5 Pro and GPT-5.4 Pro. Output pricing has an even wider range, from $0.40 to $180 per million tokens.
OpenAI's most capable model is now GPT-6 Astra, priced at $10 input and $50 output per million tokens, with $1 cached input and $12.50 cache writes. It has a 1,050,000-token context window, 128,000 maximum output tokens and a knowledge cutoff of 30 April 2026.
OpenAI added GPT-6 Sol and Luna on 22 September, then GPT-6.1 Sol on 29 September. The current Sol and Luna prices are:
- GPT-6.1 Sol at $2 input, $0.10 cached input, $2.50 cache write and $10 output per million tokens.
- GPT-6 Luna at $0.10 input, $0.01 cached input, $0.125 cache write and $0.50 output per million tokens.
Both follow the same long-prompt rule as Astra. Above 272,000 input tokens, Sol bills at $4 input, $0.20 cached input, $5 cache write and $15 output; Luna bills at $0.20, $0.02, $0.25 and $0.75. For most production work, GPT-6.1 Sol now replaces GPT-5.6 Sol as the sensible flagship and GPT-6 Luna replaces GPT-5.6 Luna as the volume tier, at half the price in each case.
The GPT-5.6 family below it is divided into three models:
- GPT-5.6 Sol for the most difficult professional work.
- GPT-5.6 Terra for a balance of intelligence and cost.
- GPT-5.6 Luna for cost-sensitive, high-volume workloads.
OpenAI reduced the API prices of Terra and Luna on 30 July 2026 and cut GPT-5.6 Sol to a promotional rate of $4 input and $20 output per million tokens on 21 August 2026, available at least until 21 November 2026. The GPT-6.1 Terra and Sol rows were checked on 27 October 2026; older and specialist rows retain their earlier source dates. Recheck those rates before budgeting. It covers API usage rather than ChatGPT subscription plans. If you want capability rather than cost, see our best ChatGPT model guide and the GPT-5.6 vs Claude Fable 5 deep dive.
How OpenAI model pricing works
Most OpenAI language models have separate prices for:
- Input tokens: The instructions, documents, conversation history and data sent to the model.
- Cached input tokens: Previously processed prompt content that the model can reuse at a reduced rate.
- Cache writes: Cache-write and retention charges depend on model and mode. Terra writes cost $1.25 per million for 30 minutes or $2.50 for 12 hours; its $0.05 reads are 5% of fresh-input price, not 10%.
- Output tokens: Everything generated by the model, including internal reasoning tokens that may not appear in the visible answer.
That final point matters. A request that returns 500 visible words may consume more billable output tokens when the model performs substantial internal reasoning. OpenAI confirms that reasoning tokens are charged as output tokens.
A realistic OpenAI model cost comparison
Price per million tokens can be difficult to translate into an operational budget. The estimates below cover the cost of processing 1,000 typical requests, where each request contains 2,000 uncached input tokens and 500 billable output tokens. Across 1,000 requests, that produces 2 million input tokens and 500,000 output tokens.
The calculation is:
Estimated cost = (2 × input price) + (0.5 × output price)
This scenario does not include caching, tools, long-context pricing, image inputs or additional reasoning beyond the assumed 500 billable output tokens.
GPT-6.1 Terra, Sol, Astra, Luna and previous-generation pricing
The following are OpenAI's standard short-context API rates, priced per one million tokens.
| Model | Input | Cached input | Output | Estimated cost for 1,000 requests |
|---|---|---|---|---|
| GPT-6 Astra | $10.00 | $1.00 | $50.00 | $45.00 |
| GPT-6.1 Terra | $1.00 | $0.05 | $5.00 | $4.50 |
| GPT-6.1 Sol | $2.00 | $0.10 | $10.00 | $9.00 |
| GPT-6 Luna | $0.10 | $0.01 | $0.50 | $0.45 |
| GPT-5.6 Sol (promotional) | $4.00 | $0.40 | $20.00 | $18.00 |
| GPT-5.6 Terra | $2.00 | $0.20 | $12.00 | $10.00 |
| GPT-5.6 Luna | $0.20 | $0.02 | $1.20 | $1.00 |
| GPT-5.5 | $5.00 | $0.50 | $30.00 | $25.00 |
| GPT-5.5 Pro | $30.00 | Not listed | $180.00 | $150.00 |
| GPT-5.4 | $2.50 | $0.25 | $15.00 | $12.50 |
| GPT-5.4 mini | $0.75 | $0.075 | $4.50 | $3.75 |
| GPT-5.4 nano | $0.20 | $0.02 | $1.25 | $1.03 |
| GPT-5.4 Pro | $30.00 | Not listed | $180.00 | $150.00 |
GPT-6 Astra is five times the token price of GPT-6.1 Sol, for the same 1,050,000-token context window and the same 128,000-token maximum output. The premium buys capability, not additional context capacity, so Astra should be reserved for work where a better answer is worth five times Sol’s fresh-input and output token bill.
Astra also carries a long-context penalty. Prompts above 272,000 input tokens are charged at twice the normal input and cache rates and 1.5 times the normal output rate for the entire request, so a single oversized prompt reprices the whole call. Chunk large document sets rather than sending them in one request.
Within the historical GPT-5.6 rows, GPT-5.6 Luna is one tenth of GPT-5.6 Terra’s cost, and GPT-5.6 Terra is approximately 44 percent less expensive than GPT-5.6 Sol at its promotional rate. GPT-5.6 Luna is also marginally cheaper than GPT-5.4 nano because both have the same input price, while Luna has a slightly lower output rate.
GPT-5.6 Sol is now cheaper than GPT-5.5 at $4 input and $20 output, against GPT-5.5's $5 and $30. GPT-5.5 is therefore difficult to justify for a new deployment unless an existing system has been specifically tested and optimised around its behaviour. Sol's promotional rate runs at least until 21 November 2026, so long-term forecasts should also model a return to $5 and $30.
The older Pro models are particularly expensive. In this comparison, GPT-5.5 Pro and GPT-5.4 Pro cost 15 times more than GPT-5.6 Terra and 150 times more than GPT-5.6 Luna.
OpenAI's current GPT-5.6 guidance uses pro mode on the selected Sol, Terra or Luna model rather than providing a separate gpt-5.6-pro model. Higher reasoning settings and pro mode can increase total output-token consumption, so the final cost still depends on how much model work each request requires.
Older and specialist text models
OpenAI continues to list several previous-generation models. Some remain useful for compatibility, specialised coding or non-reasoning behaviour, but these are compatibility references rather than the current GPT-6.1 shortlist.
| Model | Input | Cached input | Output | Estimated cost for 1,000 requests |
|---|---|---|---|---|
| GPT-5.3 Codex | $1.75 | $0.175 | $14.00 | $10.50 |
| GPT-5.2 | $1.75 | $0.175 | $14.00 | $10.50 |
| GPT-5.2 Pro | $21.00 | Not listed | $168.00 | $126.00 |
| GPT-5.1 | $1.25 | $0.125 | $10.00 | $7.50 |
| GPT-5 | $1.25 | $0.125 | $10.00 | $7.50 |
| GPT-5 mini | $0.25 | $0.025 | $2.00 | $1.50 |
| GPT-5 nano | $0.05 | $0.005 | $0.40 | $0.30 |
| GPT-5 Pro | $15.00 | Not listed | $120.00 | $90.00 |
| o3-pro | $20.00 | Not listed | $80.00 | $80.00 |
| o3 | $2.00 | $0.50 | $8.00 | $8.00 |
| GPT-4.1 | $2.00 | $0.50 | $8.00 | $8.00 |
| GPT-4.1 mini | $0.40 | $0.10 | $1.60 | $1.60 |
| GPT-4o | $2.50 | $1.25 | $10.00 | $10.00 |
| GPT-4o mini | $0.15 | $0.075 | $0.60 | $0.60 |
| Chat Latest | $5.00 | $0.50 | $30.00 | $25.00 |
GPT-5.3 Codex is a specialist agentic coding model rather than a general-purpose low-cost model. GPT-5.2, GPT-5.1 and the original GPT-5 remain available, but the current shortlist for new applications is Terra, Sol, Astra and Luna rather than these older defaults.
GPT-5 nano is the cheapest listed text model in this comparison. It would cost approximately $0.30 for the example workload, compared with $1.00 for GPT-5.6 Luna. However, GPT-6 Luna is our current-generation starting point for simple high-volume workloads. The lower price of GPT-5 nano therefore needs to be weighed against model quality, reliability and the potential cost of failed or repeated work.
GPT-4o mini also remains inexpensive at approximately $0.60 for the example workload. It can still make sense where an existing workflow has been validated against it, although it is now an older-generation model.
OpenAI describes Chat Latest as the model currently used for ChatGPT's Instant experience, but a changing chat alias is not the same as selecting a tested GPT-6.1 production model. Its underlying snapshot can change, which also makes it less suitable for applications that require predictable model behaviour.
How much do OpenAI's audio, image and embedding models cost?
These models use different pricing units, so they cannot be placed directly into the text-model chart.
| Workload | Model | Standard pricing | Practical recommendation |
|---|---|---|---|
| Realtime voice | GPT-Realtime-2.1 | Text: $4 input and $24 output per 1M tokens. Audio: $32 input and $64 output | Use for more advanced voice agents |
| Lower-cost realtime voice | GPT-Realtime-2.1 mini | Text: $0.60 input and $2.40 output. Audio: $10 input and $20 output | Best starting point for cost-conscious voice agents |
| Live translation | GPT-Realtime-Translate | $0.034 per minute | Use for streaming speech translation |
| Live transcription | GPT Live Transcribe | $0.017 per minute | Use when low latency is required |
| File transcription | GPT Transcribe | $0.0045 per minute | Current accuracy-focused transcription model |
| Lower-cost transcription | GPT-4o mini Transcribe | Approximately $0.003 per minute | Lowest estimated transcription cost |
| Text to speech | TTS-1 | $15 per 1M characters | Lower-cost speech generation |
| Higher-quality text to speech | TTS-1 HD | $30 per 1M characters | Use when its quality improvement is measurable |
| Image generation | GPT Image 2 | Text input: $5 per 1M tokens. Image input: $8. Image output: $30 | Current recommended image model |
| Embeddings | text-embedding-3-small | $0.02 per 1M input tokens | Default choice for most retrieval systems |
| Higher-quality embeddings | text-embedding-3-large | $0.13 per 1M input tokens | Use when retrieval evaluations justify the higher price |
| Moderation | omni-moderation | Free | Use for applicable text and image safety checks |
Audio, image and text tokens represent different quantities and should not be treated as equivalent units. GPT Image 2 is token-priced, so its final per-image cost changes according to image dimensions, quality, edits and the amount of image output generated.
The large embedding model costs 6.5 times more than text-embedding-3-small. The higher price should therefore be supported by a measurable improvement in retrieval relevance, search quality or downstream task accuracy. OpenAI lists the older ada embedding model at $0.10 per million tokens, making it five times more expensive than text-embedding-3-small.
OpenAI's open-weight gpt-oss-20b and gpt-oss-120b models do not have direct OpenAI API list prices. Their costs depend on local hardware, cloud infrastructure or the third-party provider used to host them.
The official pricing page still shows rates for Sora 2 and Sora 2 Pro, but OpenAI's current model catalogue marks both as deprecated. They should not form the basis of a new long-term API dependency without checking current replacement guidance.
The costs that are easy to overlook
1. Output often costs more than input
Output tokens cost six times more than uncached input tokens on GPT-5.6 Terra and Luna, and five times more on Sol at its promotional rate.
A system that sends short prompts and requests long reports can therefore cost more than a document-processing system that sends large inputs but returns compact structured data.
This becomes more important with reasoning models because invisible reasoning tokens are included in output billing. Limiting reasoning effort and setting sensible output limits can have a meaningful effect on total spend.
2. Long context has a higher rate
The listed GPT-6.1, GPT-6 and GPT-5.6 models have separate short-context and long-context rates. In the long-context tier, input pricing doubles and output pricing increases by 50 percent:
| Model | Short input | Long input | Short output | Long output |
|---|---|---|---|---|
| GPT-6.1 Terra | $1.00 | $2.00 | $5.00 | $7.50 |
| GPT-6.1 Sol | $2.00 | $4.00 | $10.00 | $15.00 |
| GPT-6 Astra | $10.00 | $20.00 | $50.00 | $75.00 |
| GPT-6 Luna | $0.10 | $0.20 | $0.50 | $0.75 |
| GPT-5.6 Sol | $4.00 | $8.00 | $20.00 | $30.00 |
| GPT-5.6 Terra | $2.00 | $4.00 | $12.00 | $18.00 |
| GPT-5.6 Luna | $0.20 | $0.40 | $1.20 | $1.80 |
Applications that repeatedly send large document collections or extensive conversation histories need to account for this higher tier.
3. Cache writes now have a price
For GPT-6.1 Terra, Standard cache writes cost $1.25 per million tokens for 30-minute retention or $2.50 for 12 hours. Reads cost $0.05, 5% of fresh input. Do not carry the older GPT-5.6 ten-percent cache rule over to Terra or GPT-6.1 Sol.
Caching is most valuable when a long, stable prompt prefix will be reused across multiple requests. Frequently changing user data should generally appear after the stable instructions so that the reusable portion can remain consistent.
4. Processing mode changes the price
OpenAI offers several processing modes:
- Standard is the normal synchronous price.
- Batch provides a 50 percent discount for asynchronous jobs that can complete within 24 hours.
- Flex uses Batch-level token pricing in exchange for slower processing and occasional resource unavailability.
- Fast mode charges a premium for faster processing.
Batch or Flex can significantly reduce the cost of enrichment, classification, evaluation, document extraction and other work that does not need an immediate response.
5. Model pricing does not include every tool cost
Web search currently costs $10 per 1,000 calls for supported model workflows, with search content tokens potentially charged separately. File search, containers and other hosted tools also have their own rates.
A model that appears inexpensive at the token level can therefore become more expensive when every request triggers searches, code execution or several agentic tool calls. This is one reason we design agents inside controlled workflows rather than letting every step call a frontier model.
6. Regional processing can add 10 percent
OpenAI applies a 10 percent uplift to eligible regional data-processing endpoints for models released on or after 5 March 2026. Organisations requiring regional processing should include this in their cost forecast.
Which OpenAI model should you use?
Best everyday starting point: GPT-6.1 Terra
Terra costs $4.50 for 1,000 example requests: 2 million fresh input tokens × $1 plus 0.5 million billable output tokens × $5. That is half Sol’s $9 and one-tenth Astra’s $45 at identical token volumes. It is an attractive starting point for writing, visual reasoning and computer-use agents; actual savings depend on completion rate and tokens used.
Batch/Flex would reduce that fixed-token Terra example to $2.25; Fast doubles it to $9. Above 272K input tokens, the full request uses $2 input / $0.10 cached / $7.50 output. Tool charges and review time are additional.
Best starting point for demanding coding: GPT-6.1 Sol
GPT-6.1 Sol is the stronger starting point for demanding coding and agent tasks that Terra does not handle well. It costs approximately $9 for the example workload, less than GPT-5.6 Terra ($10) and half of GPT-5.6 Sol ($18), while delivering current-generation GPT-6 reasoning with a 1,050,000-token context window. Keep prompts under 272,000 input tokens to avoid the long-context rate.
Sol is suitable for general assistants, document processing, research, operational workflows, sales support, structured extraction and most business automation. Existing GPT-5.6 Terra and Sol workloads should be re-evaluated on GPT-6.1 Sol.
Best current-generation budget model: GPT-6 Luna
GPT-6 Luna is the best starting point for high-volume classification, extraction, routing, categorisation and simple transformation tasks. The example workload costs approximately $0.45, less than half of GPT-5.6 Luna ($1.00).
GPT-5 nano remains marginally cheaper at list price, but belongs to an older generation. Luna is the safer general recommendation where the system needs low price, current capabilities and a 1M+ context window.
Cheapest model by list price: GPT-5 nano
GPT-5 nano produces the lowest estimated cost in the comparison, at $0.30 for 1,000 typical requests. It can be appropriate for tightly defined classification, summarisation and extraction tasks where evaluations demonstrate an acceptable error rate. The savings disappear quickly when weaker outputs create retries, human corrections or downstream workflow failures.
Best premium model: GPT-6 Astra
Astra should be reserved for work where improved reasoning creates measurable commercial value: difficult analysis, high-value coding, complex professional decisions and long-horizon agents. It costs approximately $45 per 1,000 example requests.
Routing every request to Astra would cost 100 times more than GPT-6 Luna in the comparison scenario. A controlled escalation rule is far more economical than using Astra as the universal default. See our GPT-6 Astra vs Sol vs Luna guide.
Best model for agentic coding: GPT-5.3 Codex
GPT-5.3 Codex is OpenAI's specialist model for agentic coding. Its example cost is $10.50 per 1,000 requests, close to GPT-6.1 Sol's $9. Its value comes from the coding specialisation rather than being the cheapest general-purpose model. It should be evaluated on long-horizon software-engineering work, repository changes and coding agents.
Best embedding model for most systems: text-embedding-3-small
At $0.02 per million tokens, text-embedding-3-small is the sensible default for retrieval, semantic search and clustering. The large model costs 6.5 times more. Use it only when retrieval evaluations show that the additional embedding quality materially improves the final result.
Best value realtime voice model: GPT-Realtime-2.1 mini
The mini model reduces text input pricing from $4 to $0.60 per million tokens and audio input pricing from $32 to $10. Start with the mini model for routine voice agents. Move selected use cases to the full GPT-Realtime-2.1 model when instruction following, tool use or voice-agent reliability produces a measurable gain.
The recommended production setup
A single-model strategy is rarely the most cost-effective approach.
| Request type | Recommended model | Example cost per 1,000 requests |
|---|---|---|
| Classification, extraction and routing | GPT-6 Luna | $0.45 |
| Everyday writing, visual and computer-use work | GPT-6.1 Terra | $4.50 |
| Demanding coding and agent work | GPT-6.1 Sol | $9.00 |
| Difficult, high-value exceptions | GPT-6 Astra | $45.00 |
| Agentic software engineering | GPT-5.3 Codex | $10.50 |
| Semantic search and retrieval | text-embedding-3-small | $0.02 per 1M input tokens |
| Realtime voice | GPT-Realtime-2.1 mini | Modality-dependent |
The strongest architecture uses GPT-6 Luna for predictable high-volume tasks, GPT-6.1 Terra for everyday work, GPT-6.1 Sol for demanding coding and agents, and GPT-6 Astra behind an escalation rule for requests that require more reasoning. This structure gives the system access to frontier intelligence without paying frontier prices for every classification, extraction or administrative action. It is the same principle behind our five-layer stack: put judgement where it pays, and keep everything else cheap and deterministic.
Final recommendation
- Evaluate GPT-6.1 Terra first for everyday writing, visual reasoning and computer-use applications.
- Use GPT-6.1 Sol for demanding coding and agents that do not meet their quality bar on Terra.
- GPT-6 Luna is the best current-generation model for high-volume, cost-sensitive work.
- GPT-5 nano has the lowest list price, but should only be selected after task-specific evaluations.
- GPT-6 Astra should be reserved for difficult work where better reasoning has measurable value.
- GPT-5.3 Codex is the specialist choice for agentic coding.
The main budgeting mistake is comparing input prices alone. Output length, reasoning tokens, long-context requests, cache behaviour, processing mode and tool calls can have a larger effect on the final bill than the headline model rate.
Want the same cost discipline applied to your own stack? Compare capabilities in our AI model comparison tool, or see how we design and build model-routed automation systems.
Sources checked 27 October 2026: OpenAI’s GPT-6.1 Terra announcement, Terra API specifications and OpenAI API pricing. Benchmark claims below are OpenAI-reported, not independent Stob.AI measurements.