OpenAI officially launched the GPT-5.6 model family to general availability on July 9, 2026, and it comes with a naming convention unlike anything the company has shipped before. Instead of the usual alphanumeric suffixes, GPT-5.6 introduces three named tiers — Sol, Terra, and Luna — each engineered for a distinct set of workloads, budgets, and performance expectations. If you're a developer or engineering lead trying to figure out which model belongs in your stack, this breakdown is for you.
What the "5.6" Actually Means
The version number and the tier names serve two different purposes. According to MacRumors, the "5.6" represents the model generation, while Sol, Terra, and Luna represent durable capability tiers that can advance independently over time. Think of the number as the platform and the names as product lines built on top of it.
This architecture means OpenAI can upgrade individual tiers — say, improving Luna's speed — without bumping the entire generation number. For teams building on the API, that's a meaningful shift: you're pinning to a capability tier, not just a snapshot version.
The Three Tiers, Broken Down
Each model in the GPT-5.6 family occupies a distinct position on the performance-vs-cost curve. Here's what you need to know about each one before you start routing traffic.
Sol — The Flagship for Complex Work
Sol is the most powerful model in the family, built for tasks that demand deep reasoning and precision. Per OpenAI's own benchmark data, Sol scores 94.1% on GPQA Diamond and shows strong results on ExploitBench², making it a serious contender for cybersecurity and long-horizon coding workflows.
That power comes at a price: Sol is priced at $5.00 per million input tokens and $30.00 per million output tokens, per ZDNet. Reserve Sol for tasks where quality is non-negotiable and latency is a secondary concern.
Best use cases for Sol:
Complex reasoning pipelines: Multi-step logic chains, research synthesis, and tasks requiring sustained context over long sessions.
Long-horizon coding: Architecting systems, refactoring large codebases, and generating production-ready implementations from high-level specs.
Cybersecurity workflows: Vulnerability analysis, exploit research, and security audits where accuracy is critical.
Terra — The Balanced Everyday Model
Terra sits in the middle of the family and is designed for the bulk of day-to-day engineering work. Vellum describes it as a strong performer for scoped implementation and code reviews, delivering solid results at roughly half the cost of Sol — $2.50 per million input tokens and $15.00 per million output tokens.
For most product teams, Terra will be the default workhorse. It offers enough capability to handle real engineering tasks without burning through budget on every API call.
Best use cases for Terra:
Code review and PR summaries: Reviewing diffs, flagging issues, and generating structured feedback at scale.
Scoped feature implementation: Writing and iterating on well-defined components where the problem space is clear.
Documentation generation: Producing technical docs, inline comments, and README files from existing code.
Customer-facing AI features: Powering in-product AI assistants where response quality matters but cost efficiency is also a priority.
Luna — Fast, Cheap, and Surprisingly Capable
Luna is the speed-and-cost champion of the family. Priced at $1.00 per million input tokens and $6.00 per million output tokens, it scores 74.6 on the Artificial Analysis Coding Agent Index and, according to Every, outperforms Claude Opus 4.8 at less than a quarter of the cost.
Luna is the model you reach for when you need high throughput, low latency, or are running tasks where near-perfect accuracy isn't the top priority. It's also an excellent choice for prototyping before you commit to a more expensive tier.
Best use cases for Luna:
High-volume automation: Processing large batches of data, logs, or documents where speed and cost dominate the decision.
Rapid prototyping: Iterating quickly on ideas before graduating to Terra or Sol for production.
Lightweight chat interfaces: Powering conversational UIs where response time is more important than exhaustive reasoning.
Classification and tagging: Structured output tasks like labeling, routing, or categorization at scale.
Pro Tip: Consider a tiered routing strategy in your application — use Luna for initial triage or simple queries, escalate to Terra for implementation tasks, and reserve Sol only for the highest-stakes reasoning steps. This pattern can dramatically reduce API costs without sacrificing output quality where it counts.
The New Prompt Caching System
GPT-5.6 also ships with a reworked prompt caching system that developers should factor into their cost models. According to Vellum, the new system introduces explicit cache breakpoints and a 30-minute minimum cache life. Cache writes cost 1.25x the standard uncached input rate, but cached reads receive a 90% discount on input tokens.
For applications with long, repeated system prompts — think RAG pipelines, agent frameworks, or multi-turn assistants — this caching model can significantly change the economics of running GPT-5.6 at scale. Structuring your prompts to maximize cache hits should be a first-class concern in your integration design.
ChatGPT Work: The Autonomous Agent Built on Top
Alongside the model family, OpenAI launched ChatGPT Work — an autonomous workplace agent that merges the ChatGPT and Codex desktop experiences into a single tool. Per MacRumors, it can execute multi-step tasks across connected apps, local files, and desktop tools without constant user input.
For engineering teams, ChatGPT Work represents a shift from AI-as-assistant to AI-as-collaborator. Tasks that previously required manual orchestration — pulling data from multiple sources, generating a report, and filing it in the right place — can now be delegated end-to-end. It's worth evaluating for internal tooling workflows before building a custom agent from scratch.
Key Takeaways
Three tiers, one generation: Sol, Terra, and Luna are named capability tiers within the GPT-5.6 generation, designed to advance independently over time.
Sol for depth: The flagship tier at $5.00/$30.00 per million tokens is purpose-built for complex reasoning, long-horizon coding, and cybersecurity — use it where quality is non-negotiable.
Terra for balance: At $2.50/$15.00 per million tokens, Terra is the right default for most product and engineering workloads including code review and feature implementation.
Luna for speed and scale: At $1.00/$6.00 per million tokens, Luna punches above its weight class and is ideal for high-volume automation, prototyping, and lightweight interfaces.
Cache your prompts: The new 90% discount on cached reads makes prompt structure a key cost-optimization lever — design for cache hits from day one.
