GPT-5.6 API Pricing: What Developers Pay Per Token

GPT-5.6 API pricing is set per million tokens, split by tier. After OpenAI's July 30, 2026 price cut, Sol costs $5 input and $30 output, Terra costs $2 input and $12 output, and Luna costs $0.20 input and $1.20 output.

You pay separately for the tokens you send (input) and the tokens the model writes back (output). Output is always the pricier half. Cached input tokens cost far less than fresh ones, which changes real bills a lot.

This page shows the exact rates, the before-and-after of the recent cut, worked cost examples for common developer apps, and honest comparisons to Claude, Gemini, and an open-weights option. You will also get practical ways to lower your spend.

Current GPT-5.6 API rates per million tokens

GPT-5.6 charges by tier and by direction, so each tier has one input rate and one output rate per million tokens. Sol is the flagship and costs the most; Luna is the fastest and cheapest.

Input tokens are everything you send: the prompt, system instructions, and any retrieved context. Output tokens are what the model generates. A million tokens is roughly 750,000 words of English.

These rates took effect on July 30, 2026. Always confirm the live numbers on OpenAI's pricing page before you budget a large workload.

• Sol — $5.00 input / $30.00 output per 1M tokens. Best for hard coding, security, and deep reasoning.
• Terra — $2.00 input / $12.00 output per 1M tokens. Best for high-volume everyday business work.
• Luna — $0.20 input / $1.20 output per 1M tokens. Best for summarizing, drafting, and classification.

Input vs output pricing, and how caching cuts the bill

You are billed for input and output tokens at different rates, and output is the expensive side on every GPT-5.6 tier. So the shape of your traffic decides your cost more than the tier alone.

A chatbot that reads long documents but writes short replies is mostly an input bill. An agent that writes long plans and code is mostly an output bill. Model your own ratio before you pick a tier.

GPT-5.6 supports prompt cache breakpoints. When you reuse the same leading context across calls, cached input tokens are billed at a large discount versus fresh input. Repeat-context apps benefit the most.

• Input = the prompt, system message, and retrieved context you send.
• Output = the tokens the model writes back to you.
• Cached input = repeated leading context (a system prompt or knowledge block) billed at a steep discount.
• Uncached input = new tokens the model has not seen this session, billed at the full input rate.

The July 30, 2026 price cut: before and after

On July 30, 2026, OpenAI cut Luna by roughly 80% and Terra by roughly 20%, while Sol stayed the same. The cut is widely read as a response to cheap frontier rivals like the Gemini Flash tier, Claude Haiku, and open-weights models.

OpenAI framed it as pushing the price-performance frontier. It says Luna now matches models that were frontier-class a year ago at about 6 cents on the dollar per task, at nearly nine times the speed.

The table below shows the old and new per-million-token rates so you can re-plan any budget you set before the cut.

• Sol — input $5.00 → $5.00 (unchanged); output $30.00 → $30.00 (unchanged).
• Terra — input $2.50 → $2.00 (−20%); output $15.00 → $12.00 (−20%).
• Luna — input $1.00 → $0.20 (−80%); output $6.00 → $1.20 (−80%).

Worked cost examples for a developer

The best way to read these rates is per thousand requests, using realistic token counts. Below are two common patterns: a RAG support app and an agent loop, each priced on all three tiers.

Assume a RAG request sends 4,000 input tokens (question plus retrieved context) and gets back 500 output tokens. That is 4M input and 0.5M output per 1,000 requests.

Assume an agent-loop task sends 12,000 input tokens across a few steps and writes 3,000 output tokens. That is 12M input and 3M output per 1,000 tasks. These are estimates; your real token counts will vary.

• RAG on Luna: 4M × $0.20 + 0.5M × $1.20 = $0.80 + $0.60 = about $1.40 per 1,000 requests.
• RAG on Terra: 4M × $2.00 + 0.5M × $12.00 = $8.00 + $6.00 = about $14.00 per 1,000 requests.
• RAG on Sol: 4M × $5.00 + 0.5M × $30.00 = $20.00 + $15.00 = about $35.00 per 1,000 requests.
• Agent loop on Luna: 12M × $0.20 + 3M × $1.20 = $2.40 + $3.60 = about $6.00 per 1,000 tasks.
• Agent loop on Terra: 12M × $2.00 + 3M × $12.00 = $24.00 + $36.00 = about $60.00 per 1,000 tasks.
• Agent loop on Sol: 12M × $5.00 + 3M × $30.00 = $60.00 + $90.00 = about $150.00 per 1,000 tasks.

GPT-5.6 API pricing vs Claude, Gemini, and open weights

GPT-5.6 is priced to compete across the whole range, from a premium flagship down to a very cheap fast tier. Anthropic (Claude) and Google (Gemini) both offer tiered token pricing too, and open-weights models can be near-zero to run yourself.

Exact rival rates change often, so treat these as directional and confirm each vendor's live pricing page. The pattern that holds is: a premium reasoning tier, a mid tier, and a cheap fast tier at each lab.

Open-weights models like Llama, DeepSeek, or Qwen have no per-token API fee if you self-host, but you pay for GPU time, ops, and engineering instead. That trade only wins at high, steady volume.

• GPT-5.6 (OpenAI) — Sol $5/$30, Terra $2/$12, Luna $0.20/$1.20 per 1M tokens.
• Claude (Anthropic) — tiered like-for-like: a premium reasoning model, a mid model, and a cheap fast tier (Haiku 4.5). Confirm current rates on Anthropic's pricing page.
• Gemini (Google) — a Pro reasoning tier plus a low-cost Flash tier that competes directly with Luna. Confirm current rates on Google's pricing page.
• Open weights (Llama / DeepSeek / Qwen) — no API fee if self-hosted, but you carry GPU, scaling, and maintenance costs.

How to lower your GPT-5.6 API bill

The biggest savings come from routing each task to the cheapest tier that can do it well, not from running everything on Sol. Most production traffic does not need flagship reasoning.

After routing, caching and output control matter most. Cache the shared prefix, cap the response length, and batch where latency allows. Small changes here compound across millions of requests.

Layer3 Labs helps teams design this routing so quality stays high while spend drops. Start with our free AI workflow audit or read our consumer-side breakdown on /guides/gpt-5-6-pricing.

• Route by difficulty: send routine work to Luna, everyday work to Terra, only hard reasoning to Sol.
• Cache shared context: put stable system prompts and knowledge at the front so cache breakpoints apply.
• Cap output: set a max output length and ask for concise answers, since output is the priciest half.
• Batch and dedupe: group similar requests and reuse results instead of re-querying the same thing.
• Trim retrieval: send only the context you need, not the whole document, to shrink input tokens.