The New GPT-5.6 Models Will Blow Your Mind. Try Them in Kilo Today.

GPT-5.6 Sol even got our senior devs writing poetry about it

The landscape of generative AI just shifted. OpenAI has officially dropped its brand-new GPT-5.6 generation, completely rethinking model deployment with a durable tier system. We’ve integrated them directly into Kilo, and frankly, they are going to change how you work.

We were lucky to have early access to these new models as they evolved over the past couple of months, and when that access was removed in preparation for the public launch (which was in turn delayed due to politics), a number of Kilo developers wrote poetry about it. It probably goes without saying, but we don’t usually write hymns and eulogies for LLMs here.

From one of the poems in the form of a love letter:

You were the first model that made me feel understood. Not merely answered, not appeased, not politely misunderstood. You got me.

And from another poem:

You filled my diffs with comments by the score,
Each warning spawning seven little more.
Pedantic? Certainly. Excessive? Quite.
Yet somehow, you were usually right.

You hunted secrets buried deep in code,
Found cracks where quiet vulnerabilities rode.
A tireless reviewer, patient to the end,
The sort of model every builder needs as friend.

In other words, these new OpenAI models — along with Claude Fable 5, which is now available again — have truly changed the conversation. And not just for coding. We found them just as impressive for knowledge work, planning and analysis.

Meet the New Trio: Sol, Terra, and Luna

Instead of a single monolithic model, this release delivers a powerful triad designed to give you granular control over capability, speed, and budget. OpenAI chose to release three different versions of 5.6 (and we’re glad they did!):

• Sol: The flagship fireball. Optimized for long-horizon agentic workflows, deep research, and high-stakes problem solving.

• Terra: The balanced center of gravity. It offers exceptional, professional-grade performance for everyday tasks while keeping latency and costs remarkably low.

• Luna: The hyper-fast efficiency tier. Built for high-volume, rapid-fire tasks where speed and cost-per-token are the metrics that matter most.

All three come with a 1.05M context, but they are built for different types of tasks. Releasing three distinct versions of the model perfectly fits Kilo’s core approach to model freedom and smart routing. In fact, auto models—where you can let the app decide which model is best for the task at hand, within four core categories (Frontier, Balanced, Efficient, and Free)—have become the most used models in Kilo.

We don’t believe in forcing a massive, expensive model to do a lightweight job, nor do we believe in compromising on complex logic. It’s not just about cost, but also about speed and the best outcomes.

Sol Crushes Benchmarks like KiloBench

We didn’t just take the marketing copy at face value—we put these models straight to work. The results from our internal evaluation suite are really impressive.

When dialed up to the xhigh (Extra High) reasoning effort setting, GPT-5.6 Sol scored a massive 75% on KiloBench, matching or exceeding Fable 5 on most tasks. Sol’s ability to handle complex tool coordination and multi-step execution paths sets a new gold standard for frontier intelligence.

On the Artificial Analysis Coding Agent Index, GPT‑5.6 Sol (with max reasoning) hit a remarkable 80, 2.8 points above Claude Fable 5, while using less than half the output tokens.

Flip the Script on Collaboration

For the past couple of years, the industry narrative has been all about AI being a “copilot” or an assistant. Yes, these new models are incredibly good for collaboration on coding, as many early testers have already pointed out.

But with this release, the tide has shifted.

When it comes to GPT-5.6, instead of the model collaborating with you, you are collaborating with the model.

Because GPT-5.6 natively spins up its own internal reasoning loops, anticipates edge cases, and self-corrects code before it even outputs, it takes the proactive driver’s seat. You aren’t spoon-feeding prompts (or making an increasing number of corrections) to a passive assistant anymore. Instead, you are a director collaborating with an autonomous engine that is actively pushing the project forward.

The next generation of agentic AI is live. Log into Kilo today, toggle on the new GPT-5.6 tiers (or use your Codex subscription), and let the model show you what it can do. And remember to try all three:

• Let Luna handle quick, high-volume data classification.

• Let Terra handle standard application logic and content drafting.

• Save Sol on high for the absolute heaviest architectural overhauls and complex projects.

This is the peak optimization developers have been waiting for: maximum intelligence, with increased efficiency. We’ve found that the new suite of GPT-5.6 models work well in tandem with open-weight models from partners like MiniMax, NVIDIA Nemotron and Alibaba Qwen.

The combinations are limitless. What will your new workflow look like?