OpenAI API Pricing 2026: Real Cost Per Million Tokens
GPT-6 Astra, Sol, Terra and Luna pricing explained with real workload costs, batch and caching math, and where GPT-4o and o3 fit now.
What's Changed Since April 2026
Last updated: November 2026. This guide was written around GPT-4o, GPT-4.1 and the o-series. That generation has been replaced. Based on OpenAI's pricing page and release notes:
- GPT-6 arrived in September. GPT-6 Astra is the flagship, with GPT-6 Sol and Luna following on September 22 and GPT-6.1 Sol and Terra later (September 29 and October 26).
- Sol and Luna were cut 50% versus the GPT-5.6 generation that preceded them.
- The o-series is legacy. o4-mini left ChatGPT in February and o3 left ChatGPT on August 26 (the API was unaffected). No new o-models have shipped.
- Sora is gone (app in April, API in September), and new image models GPT-Image-2.6 Flare and Sunburst launched on November 4.
- Voice moved to GPT-Live and GPT-Realtime-2.1.
The slug of this page still says GPT-4 for SEO continuity, but everything below reflects current pricing.
OpenAI API Pricing in 2026: The Complete Model-by-Model Cost Table
OpenAI runs one of the largest commercial APIs in the industry, with text, image, voice and tool pricing. The foundation of billing is the token: roughly 4 characters of English, or about 0.75 words. Every call has input tokens (your system prompt, user message and context) and output tokens (what the model generates). Output always costs more, and across the GPT-6 family it is exactly 5x input.
Here is the current text pricing as of November 2026, for prompts under 272K tokens:
| Model | Input (per 1M tokens) | Cached input | Output (per 1M tokens) | Primary Use Case |
|---|---|---|---|---|
| GPT-6 Astra | $10.00 | $1.00 | $50.00 | Most capable, hardest end-to-end work |
| GPT-6.1 Sol | $2.00 | $0.10 | $10.00 | Complex coding and agentic workflows |
| GPT-6.1 Terra | $1.00 | $0.05 | $5.00 | Fast, efficient; visual reasoning, computer use, writing |
| GPT-6 Luna | $0.10 | $0.01 | $0.50 | High-volume, focused tasks |
The cost range across the line is 100x on input ($0.10 for Luna to $10 for Astra). Picking the right model per task is the biggest lever on your bill. Verify the latest figures on the official OpenAI API pricing page. Batch is 50% off, and data residency adds 10%.
How Token Pricing Actually Works: Input vs. Output, Cached vs. Fresh
"Per 1M tokens" can feel abstract. On average, 1,000 tokens is roughly 750 English words, so one million tokens is about 750,000 words.
Input Tokens: What You Send
Input includes your system prompt, the user message and any context such as conversation history, retrieved documents, few-shot examples and tool definitions. A simple chatbot request may use a few hundred tokens; a RAG request with several chunks may use thousands. Input is the cheaper half, so sending context is relatively cheap. On Sol, a thousand words of input costs about $0.0027.
Output Tokens: What the Model Generates
Output costs 5x input on every GPT-6 model. If your application writes long answers, output will dominate the bill. A 2,000-word response (about 2,700 tokens) on Sol costs roughly $0.027 in output alone, which is $270 a day at 10,000 such responses. On Luna the same response costs about $0.00135. Set max_output_tokens and design prompts for concise answers.
Cached Tokens: The Discount You Should Be Using
Cached input is billed at a fraction of fresh input: Sol $0.10 vs $2.00, Astra $1.00 vs $10.00, Terra $0.05 vs $1.00 and Luna $0.01 vs $0.10 (per 1M tokens). For example, a 1,500-token system prompt across 100,000 requests a month is 150M tokens: $300 fresh on Sol versus $15 cached, a saving of $285. Caching is prefix-based in practice, so put static content first and dynamic content last.
Reasoning Tokens: The Hidden Cost
Models that think before answering generate internal reasoning tokens that you don't see but that OpenAI's reasoning models have billed as output. As an illustration (assumed numbers, not a measurement): a query with 1,000 input tokens, 500 visible output and 5,000 reasoning tokens costs about $0.057 on Sol and $0.285 on Astra. Check the token breakdown in your usage dashboard before projecting costs.
What Replaced GPT-4, GPT-4o and GPT-4.1?
In April this section compared GPT-4o ($2.50/$10), GPT-4.1 ($2/$8), GPT-4o mini ($0.15/$0.60) and GPT-4.1 nano ($0.10/$0.40). Those models no longer appear on OpenAI's pricing page, so we won't quote prices for them. Check OpenAI's docs for API availability if you still run them. Here is how the new ladder maps to old habits.
If you used GPT-4o or GPT-4.1 as the default: try Sol or Terra
GPT-6.1 Sol ($2/$10) sits where GPT-4o was on price and is OpenAI's pick for complex coding and agents. GPT-6.1 Terra ($1/$5) is half the price and targets interactive use: visual reasoning, computer use and writing.
If you used GPT-4o mini or GPT-4.1 nano: try Luna
GPT-6 Luna ($0.10/$0.50) is the efficient tier for classification, extraction, routing and high-volume Q&A. It is the natural "triage layer" in a multi-model architecture, deciding which requests deserve a bigger model.
If you used GPT-4 Turbo
The original GPT-4 pricing ($30/$60 at launch, per the earlier version of this guide) is long gone. Any app still on it should migrate; every current model is cheaper per token.
Real-world cost examples (list prices)
- Support exchange (600 input + 350 output tokens): about $0.0047 on Sol, $0.0024 on Terra, $0.00024 on Luna.
- Summarizing a 15-page legal document (about 6,000 in + 800 out): about $0.020 on Sol, $0.010 on Terra.
- Classifying 500,000 tickets (about 250 in + 50 out each) on Luna: $25, or $12.50 through Batch.
- A consumer chatbot at 2 million messages a month (400 in + 200 out): $280/month on Luna, $2,800/month on Terra.
For enterprise seat pricing and API-vs-seat decisions, see our ChatGPT API pricing guide.
Beyond Text: Images, Voice and Tools
The older version of this section covered DALL-E 3, Whisper, TTS and embeddings. The current pricing page lists different products, so we've replaced the table. We couldn't confirm embedding prices on the page, so check the docs if you depend on them.
Image Generation
| Model | Role | Price |
|---|---|---|
| GPT-Image-2.6 Flare | Fastest, high-quality everyday images | Image input $8 / cached $2 / output $30 per 1M tokens; text input $5 / cached $1.25 |
| GPT-Image-2.6 Sunburst | Editing precision | Same rates as Flare |
Images are converted to tokens and billed per token, so cost per image depends on size and quality. OpenAI notes that low and medium quality output costs changed with this launch. Both models launched on November 4, 2026.
Voice and Speech
| Model | Price |
|---|---|
| GPT-Live-1 (expressive conversation) | $0.05 per minute |
| GPT-Realtime-2.1 | Audio $32 in / $64 out per 1M tokens; text $4 / $24; image $5 in |
| GPT-Realtime-2.1 mini | Audio $10 in / $20 out per 1M tokens; text $0.60 / $2.40; image $0.80 in |
| GPT-Transcribe (async) | $0.0045 per minute |
| GPT-Live-Transcribe (streaming) | $0.017 per minute |
| GPT-Realtime-Translate | $0.034 per minute |
Transcribing a one-hour meeting with GPT-Transcribe costs about $0.27. Realtime audio is the most expensive endpoint by a wide margin, so route text-only requests through the standard text models.
Tools
The web search tool is $10 per 1,000 calls (search content tokens are free). Containers start at $0.03 for a 1 GB container. Sora is discontinued and no longer available in the API.
Batch API, Rate Limits & Cost Optimization: Cut Your OpenAI Bill in Half
Raw per-token pricing is half the story. OpenAI offers several service tiers and discounts.
Batch API: A Flat 50% Discount
Per the pricing page, Batch saves 50% on inputs and outputs, with tasks run asynchronously over 24 hours.
| Model | Standard Input / Output | Batch Input / Output |
|---|---|---|
| GPT-6 Astra | $10.00 / $50.00 | $5.00 / $25.00 |
| GPT-6.1 Sol | $2.00 / $10.00 | $1.00 / $5.00 |
| GPT-6.1 Terra | $1.00 / $5.00 | $0.50 / $2.50 |
| GPT-6 Luna | $0.10 / $0.50 | $0.05 / $0.25 |
Other Service Tiers
- Flex processing: lower costs in exchange for slower responses and occasional resource unavailability, suited to non-production work.
- Fast mode: high-speed performance, pay-as-you-go.
- Data residency: +10%.
- Scale Tier and Reserved Capacity: via sales, for larger workloads.
Rate Limits and Spend Controls
OpenAI applies rate limits that grow with account usage. We couldn't re-verify the April tier table, so see the rate-limit docs for your account. For spend control, OpenAI says you can set a monthly budget after which requests stop being served (with a possible enforcement delay), set an email alert threshold, and apply per-project billing restrictions. Do this on day one.
Optimization Checklist
- Route by difficulty: Luna first, escalate only when needed.
- Cache static prefixes.
- Batch anything offline.
- Cap output tokens.
- Measure Astra against Sol before committing, since Astra is 5x the price.
Real-World Cost Examples: What Common Workloads Actually Cost
Worked examples at list prices, with assumed token counts. They show the shape of the bill, not your exact number.
| Workload | Assumptions | Luna | Terra | Sol |
|---|---|---|---|---|
| Support chatbot | 50,000 conversations/mo, 800 in + 400 out | $14 | $140 | $280 |
| Content pipeline | 500 articles/mo, 1,500 in + 2,000 out | $1.15 | $5.75 | $11.50 |
| Document analysis | 10,000 docs/mo, 10,000 in + 500 out | $12.50 | $125 | $250 |
With Batch, halve each number. With a cached shared prefix, the cached portion drops further (Sol: $0.10 instead of $2.00 per million).
The Astra Question
Take a research tool running 5,000 queries a month with 2,000 input and 6,000 billed output tokens each (30M output). On Sol that is about $320/month; on Astra, about $1,600/month. Unless Astra's output is measurably better for your use case, you're paying 5x for nothing.
Voice Example
A support line using GPT-Live-1 at $0.05 a minute costs $3 for 60 minutes of conversation. Realtime audio models are billed per token, so test with real calls to estimate.
OpenAI vs. Claude vs. Gemini vs. Other Providers: Cost Comparison
Per 1 million tokens (input / output), as of November 2026, from each vendor's pricing page or documentation. Gemini 4 Argon and Carbon had introductory prices at the time of writing.
| Provider / Model | Input | Output | Note |
|---|---|---|---|
| OpenAI GPT-6.1 Sol | $2.00 | $10.00 | Cached input $0.10 |
| OpenAI GPT-6 Luna | $0.10 | $0.50 | Cached input $0.01 |
| Anthropic Claude Opus 5.5 | $4.00 | $20.00 | 1M context |
| Anthropic Claude Sonnet 5.5 | $2.00 | $10.00 | Cache hit $0.10 |
| Anthropic Claude Haiku 5.5 | $0.10 (up to 100K prompt) | $0.50 | Higher above 100K |
| Google Gemini 4 Hydrogen | $1.50 | $7.50 | Free tier available |
| Google Gemini 4 Argon | $2.00 (intro) | $10.00 (intro) | $4 / $20 afterwards per Google |
| xAI Grok 4.7 | $2.00 | $6.00 | 500K context |
| DeepSeek V4.1-Flash | $0.15 off-peak | $0.60 off-peak | Peak hours are double |
| DeepSeek V4-Pro | $0.66 off-peak | $1.98 off-peak | Peak hours are double |
Mistral also sells hosted models, but we haven't verified its current API prices, so we're not listing numbers.
Takeaways. OpenAI and Anthropic now match at $2/$10 for the workhorse tier and $0.10/$0.50 for the budget tier. Anthropic notes its newer models produce about 30% more tokens for the same text, so a like-for-like test matters. DeepSeek is far cheaper on paper. Google offers a free tier on Hydrogen. See our Claude pricing breakdown for Anthropic's plans.
When to Use Which OpenAI Model: A Decision Framework
Use GPT-6 Luna ($0.10 / $0.50) when:
- You're classifying, extracting, routing or answering simple questions at volume
- Latency and cost matter more than polish
- You want a triage layer that escalates hard cases
Use GPT-6.1 Terra ($1.00 / $5.00) when:
- You need visual reasoning, computer use or good writing at moderate cost
- Luna's quality falls short but Sol is overkill
Use GPT-6.1 Sol ($2.00 / $10.00) when:
- You're building coding tools or agents
- You need the best default balance of capability and price
Use GPT-6 Astra ($10.00 / $50.00) when:
- Sol consistently fails on your hardest tasks
- Accuracy on high-stakes work justifies 5x the price
Avoid legacy models for new projects
GPT-4o, GPT-4.1, GPT-4 Turbo and the o-series are no longer on the pricing page. Don't start new projects on them; if you maintain old code, plan a migration and test quality on your own data.
The Verdict: What OpenAI API Pricing Means for Your Budget
OpenAI's pricing is now a clean four-step ladder: Luna $0.10/$0.50, Terra $1/$5, Sol $2/$10 and Astra $10/$50. The September cuts to Sol and Luna made the middle and bottom of the ladder much cheaper than the April equivalents, while the top moved up.
Practical advice: start on Luna or Terra, move to Sol for coding and agents, treat Astra as an exception, batch everything offline and cache everything shared. Re-benchmark every few months, since the model names have changed completely in six months.
Always confirm current numbers on the official pricing page. For consumer plans see our ChatGPT pricing guide.
Key Takeaways
- 01OpenAI's current API ladder is GPT-6 Luna ($0.10/$0.50), GPT-6.1 Terra ($1/$5), GPT-6.1 Sol ($2/$10) and GPT-6 Astra ($10/$50) per million tokens
- 02GPT-4o, GPT-4.1 and the o-series no longer appear on the pricing page; the o-series is legacy
- 03Output costs 5x input on every GPT-6 model, so cap output length
- 04Batch is 50% off and cached input is far cheaper than fresh (Sol: $0.10 vs $2.00)
- 05Sonnet 5.5 and Haiku 5.5 match Sol and Luna list prices, so compare on quality and real token counts
- 06Astra costs 5x Sol; use it only where Sol demonstrably falls short
Frequently Asked Questions
Related Guides
ChatGPT API Pricing 2026: What 1M Tokens Actually Cost
GPT-6 Astra, Sol, Terra and Luna priced per token, with cost examples, batch math, enterprise notes and a Claude and Gemini comparison.
Claude Pricing 2026: Free, Pro, Max, Ultra and API
Claude plans from Free to Ultra at $200, plus Team, Enterprise and API rates for Fable 5.2, Opus 5.5, Sonnet 5.5 and Haiku 5.5. Which plan you need.
Mentioned Tools
Work on a tool covered here? See how sponsored placements work