OpenAI has released GPT-5.6 after two weeks of preview access limited to government-approved organizations. The model comes in three variants: Luna (smallest), Terra (medium), and Sol (largest), priced from $1 to $5 per million input tokens. All three feature a one-million-token context window and a maximum of 128,000 output tokens.

Alongside the models, OpenAI is launching ChatGPT Work: an agent that moves between apps (Google Drive, Slack, Salesforce) and files, stays on a project for hours if needed, and completes complex tasks without requiring you to orchestrate each step. Sol demonstrates advanced capabilities: it can perform autonomous post-training on Luna and, in Ultra mode, produced a mathematical proof of the Cycle Double Cover Conjecture.

Why this matters to you. If you use language models in daily work, your choices just got more complex: three sizes of the same model, five reasoning levels (Light through xhigh, plus Max and Ultra), and now an agent meant to handle entire projects. The practical question becomes: when Luna suffices, when Sol is necessary, and whether ChatGPT Work can truly sustain a full workday or stalls midway. OpenAI publishes benchmarks showing Sol ahead of Claude Fable 5 on extended agentic tasks, but on the same day they flag roughly 30% of SWE-Bench Pro—a coding test where Fable had dominated Sol—as broken. Benchmark numbers need this context.

Where to look. OpenAI’s model guidance page lists new API capabilities (programmatic tool calling, multi-agent, cache breakpoints). An independent comparison compiled costs for 18 variations of the same prompt: from 0.71 cents (Luna, no reasoning) to 48.55 cents (Sol, maximum effort).

In detail

The context: the limited preview

Two weeks ago, OpenAI released GPT-5.6 only to government-approved organizations in a limited preview that sparked debate. On July 9, the Trump administration cleared it for public release. Sam Altman called it “the best model we’ve ever built.”

Three variants and pricing

Luna, Terra, and Sol cost $1/$6, $2.50/$15, and $5/$30 per million tokens (input/output) respectively. For comparison, Claude Opus is $5/$25 and Fable 5 is $10/$50. But price per million tokens tells only part the story: reasoning-enabled models consume very different internal token quantities for the same task.

All three variants have a knowledge cutoff of February 16, 2026, a one-million-token context window, and 128,000-token maximum output. Five reasoning levels (none, low, medium, high, xhigh) let you choose how much internal work the model performs. Above xhigh are Max and Ultra, which activate parallel sub-agents.

Benchmarks: OpenAI leads on agents, trails on code

OpenAI emphasizes extended agentic tasks: on Agents’ Last Exam (55 professional fields), GPT-5.6 Sol reaches 53.6 points, beating Fable 5 by 13.1 points. Even at medium reasoning, Sol outperforms Fable by 11.4 points at a quarter of the estimated cost. Luna and Terra surpass Fable at a sixteenth of the cost.

On SWE-Bench Pro (a coding test), Fable 5 achieved 80% versus Sol’s 64.6%. The day before launch, OpenAI published an article estimating roughly 30% of SWE-Bench Pro tasks are “broken” and recommending caution with that benchmark. The timing did not go unnoticed.

ChatGPT Work: the agent meant to complete projects end-to-end

ChatGPT Work is described as an agent that takes an objective and completes it, moving across apps and files, working on a project for hours if needed. It’s powered by Codex (OpenAI’s programming tool, now integrated into ChatGPT) and GPT-5.6. It can access Google Drive, Slack, Salesforce. The promise: give it a task, return when it’s done.

The catch is that autonomous agents tend to get stuck or diverge when tasks stretch longer. The ability to “stay on a project for hours” is the hardest part to verify without hands-on use.

Sol Ultra and advanced mathematics

GPT-5.6 Sol, in Ultra mode, produced a proof of the Cycle Double Cover Conjecture, an open problem in graph theory. The proof PDF is public. According to OpenAI, Sol can also perform autonomous post-training on Luna—essentially self-improving by training smaller versions of itself. These are bold claims beyond everyday use, but they signal direction.

New API capabilities

The model guidance lists features that change how you use these models:

• Programmatic Tool Calling: the model writes and executes JavaScript to orchestrate tool calls instead of returning a sequence for external execution.
• Multi-agent: the model can activate focused sub-agents for parallel work directly from the API.
• Prompt cache breakpoints: like Claude, you can specify cache points rather than relying on automatic detection.
• detail: original for images: the model doesn’t resize the image before processing.

Programmatic tool calling resembles the dynamic filtering mechanism Anthropic added to web search: instead of raw results, the model executes code on those results in a single turn.

What we don’t know

Early-access reviewers describe Sol as “definitely quite competent,” but not better than Fable on the complex coding tasks they tried. That’s anecdotal, not a benchmark, but it’s the kind of observation that matters: from someone using these models daily.

We don’t yet know how ChatGPT Work holds up on real, extended projects, or how reliable Sol’s autonomous post-training on Luna is. And the fact that OpenAI flagged roughly a third of SWE-Bench Pro as broken the day before launch raises the question: how robust are the other benchmarks they cite?

Where this leads

Models are now families, not singles. Choosing the right model for a task becomes part of the work: when Luna suffices, when Terra is needed, when Sol justifies the cost. And if agents like ChatGPT Work deliver as promised, the work changes: you move from step-by-step instructions to stating objectives and verifying they’re met. But we’re still in the phase where promises must be tested in the field, not taken on faith from announcements.