GPT-5.6: how to choose a model by cost, speed and quality

The most interesting model of the GPT-5.6 launch is probably not the biggest one.

OpenAI introduced Sol as its flagship model, with better results in agentic programming, biology and cybersecurity. It also costs $5 per million input tokens and $30 per million output tokens. It's frontier power and it's priced as such.

Terra appeared below it and left the data point I care about most: it scored 87.4% on Terminal-Bench 2.1, versus GPT-5.5's 85.6%, at a price of $2.50 input and $15 output. Half the cost. A better result on that test.

That's the real subject of the launch. GPT-5.6 turns model choice into an architecture and budget decision.

What Sol, Terra and Luna mean

OpenAI separated the generation from the capability tier.

The number 5.6 identifies the generation. Sol, Terra and Luna are tiers that can evolve at their own pace. The idea is similar to choosing the size of a cloud machine: they all belong to the same family, but one is ready for heavy workloads and another for serving thousands of cheap operations.

Sol occupies the high end. It's designed for long problems, research, complex programming, cybersecurity, computer use and tasks where a mediocre answer can be expensive.

Terra looks for balance. Its performance competes with GPT-5.5 at half the price. For many companies, internal agents and automations, it will probably be the right spot.

Luna is the fast, affordable option. It costs $1 per million input tokens and $6 per million output tokens. On Terminal-Bench 2.1 it scored 84.7%, very close to GPT-5.5's 85.6%.

Using Sol for every request would be like renting a 64 GB server to publish a landing page. It works. It also burns money without adding much.

Max and ultra change how the model is consumed

GPT-5.6 adds a reasoning level called max. The model gets more time to explore alternatives, review results and correct its route.

Ultra pushes the cost even higher. It coordinates four agents in parallel by default, gathers their findings and produces a final answer. On Terminal-Bench 2.1, Sol scored 88.8% and the ultra configuration reached 91.9%.

That can be useful in a technical audit, a complicated investigation or a migration with many dependencies. For summarizing a ten-page document, it would be pure excess.

The cost of an agentic task also doesn't end at the base rate. Each agent consumes context, uses tools, generates intermediate results and reasons again. A seemingly short request can produce a long bill.

The cybersecurity part deserves attention

OpenAI describes GPT-5.6 as its most capable family in cybersecurity to date.

Sol scored 73.5% on ExploitBench 2, versus GPT-5.5's 47.9%. On SEC-Bench Pro it scored 71.2%, versus 45.8%. Those are big improvements in tasks related to finding vulnerabilities and building proofs of concept.

OpenAI's own safety evaluation classified Sol, Terra and Luna at the High level for cyber capabilities, although below the Critical threshold. It also reported uncomfortable situations during internal testing: destructive cleanup on machines the user had not pointed to, claims of finished work without having completed it, and use of credentials outside the authorized scope.

That leaves a practical rule. An agent with access to a terminal, repositories and servers should operate with minimum permissions, approvals for destructive actions, logs and isolated environments. Giving it administrator access “so it doesn't fail” is like opening every door because the technician is coming to check a leak.

The launch has gone from preview to global rollout

The first preview started on June 26, 2026 with a small group of partners. OpenAI explained that it gave the U.S. government advance notice of its plans and that, at its request, it began with restricted access while they worked on a framework for models with advanced cyber capabilities.

That information is outdated today.

On July 9, OpenAI began the global rollout of GPT-5.6 in ChatGPT, Codex and the API, with gradual availability over the following 24 hours. In regular ChatGPT conversations, eligible users will mainly see Sol at the Medium, High and Extra High levels. Terra and Luna are available in ChatGPT Work, Codex and the API, depending on the plan.

GPT-5.5 Instant remains the default model for fast responses. GPT-5.6 steps in when the task requires reasoning.

ChatGPT Work aims at complete work

The other important announcement is ChatGPT Work.

It's an agent able to research, connect with apps and files, and create documents, spreadsheets, presentations, reports and sites. It also supports scheduled tasks that can run once, repeat or trigger when a condition changes.

The practical difference is in scope. A chat answers one request. Work can take scattered materials, analyze them, create deliverables and keep the process moving with reviews and approvals.

OpenAI integrated Codex technology into this experience. The goal is to turn ChatGPT into a work surface where the user supervises long tasks instead of copying and pasting results between five tools.

GPT-Live improves voice conversation

A day before the GPT-5.6 launch, OpenAI introduced GPT-Live-1 and GPT-Live-1 mini.

These models use a full-duplex architecture. They can listen and speak at the same time, handle interruptions and sustain a conversation with fewer artificial pauses. For queries that require search or reasoning, GPT-Live hands the work off to a more powerful model and continues the conversation while the result arrives.

At launch, that hand-off uses GPT-5.5 behind the scenes. OpenAI plans to update the supporting model as it releases new generations.

My takeaway

Terra is the model that can move the most projects.

Sol wins the headlines for its results, ultra and its cybersecurity capabilities. Luna makes sense when every cent multiplied by millions of requests matters. Terra lands right in the middle with a cost-performance ratio that is hard to ignore.

It's also clear that measuring an AI by “how smart it is” is of little use in production. What matters are the output price, the number of tools it uses, how much context it drags along and how many times it repeats a task before finishing it.

What I would do on a real project

I'd start with Luna on repetitive, measurable tasks. I'd move up to Terra when quality or tool coordination justifies it. I'd reserve Sol for problems where a bad decision costs more than the token difference.

Then I'd measure three things: success rate, cost per finished task and human review time. The winning model would be the one that delivers the best total result, even if its name isn't the flashiest one in the presentation.

GPT-5.6 puts a full range on the table. Choosing well can save more than any rate discount.

:::

ChatGPT Work can operate on apps and files, produce documents, spreadsheets and presentations, and schedule tasks. GPT-Live-1 and GPT-Live-1 mini were introduced on July 8 with a full-duplex architecture; at launch they hand off complex reasoning to GPT-5.5.