OpenAI released GPT-5.6 on July 9 as three models rather than one: Sol, the flagship; Terra, for balanced everyday work; and Luna, the cheapest. The launch post made them available in ChatGPT, Codex and the API, with reasoning effort as the main lever. Within six weeks the prices moved twice, so the figures quoted at launch no longer describe what the API charges.
Launch price and price today
| Model | Launch (July 9), input / output per 1M tokens | Current, input / output | Change |
|---|---|---|---|
| Sol | $5 / $30 | $4 / $20 | August 21, promotional |
| Terra | $2.50 / $15 | $2 / $12 | July 30, −20% |
| Luna | $1 / $6 | $0.20 / $1.20 | July 30, −80% |
On July 30 OpenAI cut Luna by 80% and Terra by 20% and left Sol unchanged. On August 21 it lowered Sol to $4 per million input tokens and $20 per million output, which the launch post describes as a drop of over 20% and the API changelog as promotional pricing available at least through November 21, 2026; batch processing costs half.
The mechanics that move the bill
Three rules matter more than the headline rates. Above 272,000 input tokens, the whole request is billed at twice the input rate and 1.5 times the output rate. Cache writes cost 1.25 times the uncached input price while cache reads keep a 90% discount, so caching a prompt that is rarely reused can cost more than it saves. And on July 30 OpenAI renamed Priority Processing to Fast mode: up to 2.5 times faster for gpt-5.6-sol, at twice the price, sharing Standard’s rate limits.
The throughput limits show what each model is for. Sol and Terra go from 500 requests and 500,000 tokens per minute at Tier 1 to 15,000 requests and 40 million tokens at Tier 5; Luna, built for volume, reaches 30,000 requests and 180 million tokens per minute at Tier 5.
Sol and Terra share a context window of 1,050,000 tokens with 128,000 of output and a knowledge cutoff of February 16, 2026. OpenAI’s pages did not show Luna’s window in legible form, and the family runs on the Chat Completions, Responses and Batch endpoints only.
Two tools for agents
The family shipped with two features for agentic work in the Responses API. Programmatic Tool Calling lets the model write programs that coordinate tools and process intermediate results in memory, which OpenAI says makes it compatible with Zero Data Retention. A multi-agent mode in beta runs concurrent subagents inside one request, with all three models.
Whose numbers
Every comparison with other models in the launch material, from Agents’ Last Exam to cost per task against Claude Fable 5, is OpenAI’s own measurement, and OpenAI footnotes its cost and latency comparisons as offline simulations whose real-world results may differ substantially. The system card places GPT-5.6 below OpenAI’s Critical threshold in cybersecurity and biology, while calling it the company’s strongest cybersecurity model at the time.
Since September 3 GPT-5.6 is no longer OpenAI’s flagship: GPT-6 Astra took that place at $10 and $50. OpenAI’s documentation still recommends Terra to balance intelligence and cost and Luna for cost-sensitive, high-volume work, and no GPT-5.6 model is on its deprecation list.
Related reading
- Advancing the price-performance frontier with GPT-5.6 — OpenAI, July 30
- GPT-5.6 system card — OpenAI
- GPT-6 Astra ships as a product — IA al Día