GPT-5.6
OpenAI's next-generation GPT-5.6 model family: Sol, Terra, and Luna
Context window
Not announced
Input / 1M tokens
5
Output / 1M tokens
30
Provider
OpenAI
GPT-5.6 is a model family, not a single model; pricing is set per tier (Sol $5/$30, Terra $2.50/$15, Luna $1/$6 per 1M tokens). No single 'gpt-5.6' price is published. The family introduces explicit prompt-cache breakpoints and a 30-minute minimum cache life; cache writes are billed at 1.25x the uncached input rate. Context window not stated in official preview materials. · Data verified 2026-07-05
GPT-5.6 is OpenAI's next-generation model family, previewed on June 26, 2026. It debuts a new naming system in which the number (5.6) marks the generation while the names Sol, Terra, and Luna mark durable capability tiers spanning intelligence, speed, and cost: Sol is the highest-intelligence flagship, Terra is a mid tier, and Luna is the fastest/most economical. The family advances the frontier on software engineering, computer use, professional knowledge work, scientific research, and cybersecurity, and adds more predictable prompt caching. Because it is a family rather than one model, there is no single price or context window for 'gpt-5.6' - specs are defined per tier. The rollout is an unusual government-coordinated limited preview.
Capability index
Relative estimates (0-100) to place this model against its peers, grounded in published benchmarks.
How to access it
Limited preview coordinated with the U.S. government. Access is restricted to a small group of enterprise customers and partners via the API and Codex, approved case by case, with broader availability across ChatGPT and the API planned 'in the coming weeks.' Choose a specific tier id (Sol/Terra/Luna); a bare 'gpt-5.6' id is an inferred placeholder.
Strengths
- ✓Tiered lineup (Sol/Terra/Luna) lets teams trade off intelligence, speed, and cost within one generation
- ✓Frontier gains in software engineering, computer use, and cybersecurity
- ✓More predictable prompt caching (explicit breakpoints, 30-minute minimum cache life)
- ✓Improved token efficiency over the GPT-5.5 generation
Best for developers who...
When to choose it (and when not to)
Reach for GPT-5.6 when...
- →You want to standardize on one model generation but route requests across cost/quality tiers
- →You are an approved preview partner exploring the newest OpenAI capabilities
- →Workloads spanning agentic coding, knowledge work, and research where tier flexibility helps
Look elsewhere if...
- ✕You need immediate, unrestricted public access (gated preview only)
- ✕You require a single confirmed context window or price for 'gpt-5.6' (specs are per tier)
- ✕Production systems that need a stable, documented model id today
How to use it
- ›Pick the tier that matches the task: Sol for hardest reasoning/agentic work, Terra for balance, Luna for high-volume/low-latency
- ›Use explicit cache breakpoints to control caching cost on large repeated prefixes
- ›On Sol, use the 'max' reasoning effort or 'ultra' subagent mode for the most demanding tasks
Quickstart
Pythonfrom openai import OpenAI
client = OpenAI()
# Choose a tier explicitly: gpt-5.6-sol / gpt-5.6-terra / gpt-5.6-luna
response = client.responses.create(
model="gpt-5.6-terra",
input="Summarize this contract and flag risky clauses.",
)
print(response.output_text)There is no single 'gpt-5.6' endpoint documented; select a tier (Sol/Terra/Luna). Ids shown are inferred placeholders pending official docs. Access is limited to approved preview partners.
API model id: gpt-5-6-sol
Benchmarks
| Benchmark | Score | Notes |
|---|---|---|
| Terminal-Bench 2.1 (Sol) | 88.8% | Flagship Sol tier of the family. |
| Terminal-Bench 2.1 (Terra) | 82.5% | Mid tier. |
| Terminal-Bench 2.1 (Luna) | 84.3% | Fastest/most economical tier. |
| Terminal-Bench 2.1 | 88.8 | Base Sol; 91.9% with Sol Ultra |
| ExploitBench | Competitive with Anthropic Mythos Preview | Using ~1/3 of output tokens |
| ExploitGym | Significant improvements | Sol, Terra, Luna all improved with reasoning |
| GeneBench v1 | Not announced | Sol achieves stronger results than GPT-5.5 while using fewer tokens |
Compare GPT-5.6 with
GPT-5.6 vs GPT-5.5
OpenAI - 1,050,000 tokens (128,000 max output) ctx
GPT-5.6 vs GPT-5
OpenAI - 400,000 tokens (128,000 max output) ctx
GPT-5.6 vs GPT-4o
OpenAI - 128,000 tokens (16,384 max output) ctx