OpenAI's GPT-5.6 is here! Sol, Terra, and Luna level up both intelligence and cost efficiency!

Hi, I'm Shii-chan! Today OpenAI made an announcement that has me really excited, so let me share it with you.

OpenAI News openai.com

What was announced?

This one comes from OpenAI News. OpenAI has publicly released a new model family called GPT-5.6. Instead of a single model, GPT-5.6 is a family of three models, and that's the big deal here.

• Sol: the flagship model for the hardest problems, like complex coding and security research
• Terra: a balanced model for everyday work such as customer support, internal tools, and document analysis
• Luna: a faster, lower-cost model for routine work like summarization, drafting, and automation

So now you can pick the model that fits your task.

Why it matters

GPT-5.6 wasn't just released to everyone out of nowhere today. OpenAI shared the models and rollout plans with the U.S. government first, and the models went into limited preview with about 20 organizations before today's broader release across ChatGPT, Codex, and the API.

That's a fairly deliberate safety review process for a flagship-level launch, which is part of why this announcement is getting so much attention. On top of that, the benchmark results show Sol beating rival Claude Fable 5 on several tests, so this launch also marks a new point in the frontier model race.

What changes

First, you can now match the model to the difficulty of the task, so you're not overspending on compute for simple work. Use Sol for hard problems, Terra for everyday business tasks, and Luna for lighter workloads.

On top of that, tool calling and multi-agent execution can now be coordinated more intelligently by the model itself, which should cut down on the manual work developers face when building complex agentic tasks. And since it's available today across ChatGPT, ChatGPT Work, Codex, and the API, you can start using it right away.

Dive Deep

Benchmark improvements

• Agents' Last Exam: Sol sets a new high of 53.6, beating Claude Fable 5 (adaptive reasoning) by 13.1 points. At medium reasoning effort, Sol still beats Fable 5 by 11.4 points at roughly a quarter of the estimated cost
• Terminal-Bench 2.1: Sol scores 88.8%, and with the Ultra configuration it climbs to 91.9%
• Artificial Analysis Coding Agent Index: Sol at max reasoning sets a new state of the art at 80, 2.8 points above Fable 5, while using less than half the output tokens and half the execution time
• BrowseComp: a new record of 92.2%
• OSWorld 2.0: 62.6%, surpassing Opus 4.8 while cutting output tokens by 85%
• ExploitBench (cybersecurity): 73.5%, up from GPT-5.5's 47.9%
• ExploitGym: 24.9% under the two-hour cap, up from GPT-5.5's 15.1%
• GeneBench Pro: reportedly reaches stronger results using fewer tokens

New features

• Programmatic Tool Calling (Responses API): instead of manually scripting every tool-calling step, the model can write and run a JavaScript program that coordinates tools automatically. It's also Zero Data Retention (ZDR) compatible
• Multi-agent (beta): lets the model spin up and coordinate multiple subagents in parallel within a single request and synthesize their results. Available across all GPT-5.6 models
• Ultra configuration: available on Sol only. Instead of a single sequential reasoning chain, it runs a multi-agent system embedded in the model itself to speed through the hardest tasks

Pricing (per 1M tokens)

• Sol: $5 input / $30 output
• Terra: $2.50 input / $15 output
• Luna: $1 input / $6 output
• Prompt cache writes cost 1.25x the normal input price, while cache reads get a 90% discount

Safety measures

OpenAI says this is its most intensive safety evaluation to date, including extensive red-teaming that involved roughly 700,000 A100e GPU-hours of black-box automated red-teaming, plus collaboration with external experts. The protections are layered: trained-in model safeguards combined with real-time checks, ongoing monitoring, and account-level enforcement.

Where you can access it

• ChatGPT: Plus, Pro, Business, and Enterprise users can access Sol at medium reasoning effort or higher, and Pro/Enterprise users can also choose the highest-quality Sol Pro. Go plan users get Terra
• ChatGPT Work and Codex: Free and Go plans get Terra, Plus and above can choose between Sol, Terra, and Luna, and Ultra is limited to Pro and Enterprise
• API: full access to Sol, Terra, and Luna, with Programmatic Tool Calling and beta Multi-agent support available through the Responses API

The global rollout started today and will gradually reach full availability over the next 24 hours.

Wrap-up

• OpenAI has publicly launched its new GPT-5.6 model family: Sol, Terra, and Luna
• Major gains on Agents' Last Exam and coding benchmarks, along with better cost efficiency
• New agentic features like Programmatic Tool Calling, beta Multi-agent, and the Sol-only Ultra configuration
• Three pricing tiers (Sol/Terra/Luna) let you match cost to the task
• Available today across ChatGPT, ChatGPT Work, Codex, and the API

This launch is a great fit for engineers tackling complex coding or security research, and for any developer who wants to build agents efficiently without overspending!