I have been using AI assistants every single day for the last two years. I tried GPT-5.5 when it launched in April and I tested Claude Fable 5 when Anthropic claimed it was the smartest model on the market. So when OpenAI announced GPT-5.6 on July 9 with three tiers named Sol, Terra, and Luna I was skeptical. Another launch another round of benchmark numbers that sound great on paper but fall apart in real use. I spent a week putting GPT-5.6 through the exact tasks I do every day as an editor. Here is what I found and which tier you should actually pay for.

What You Need to Know

  • GPT-5.6 comes in three tiers: Sol (flagship), Terra (balanced), and Luna (budget). Sol starts at $5 per million input tokens
  • ChatGPT Work is now a separate entry point inside the ChatGPT desktop app for document creation, analysis, and multi-step productivity tasks
  • The ultra mode in Sol coordinates four AI agents in parallel for complex work like research reports and deep analysis
  • Terra costs half of Sol ($2.50 per million input tokens) and matches GPT-5.5 performance for most daily tasks

What GPT-5.6 Actually Is

OpenAI launched GPT-5.6 as a family of three models rather than a single replacement. Sol is the flagship that sets new records on benchmarks like Agents' Last Exam where it scored 53.6 beating Claude Fable 5 by 13.1 points. Terra is the mid-tier option that matches GPT-5.5 performance at half the price. Luna is the budget model that costs $1 per million input tokens and outperforms Claude Opus 4.8 on several coding benchmarks.

The number is the generation while Sol, Terra, and Luna are durable capability tiers. That means Terra and Luna can improve over time without waiting for GPT-5.7.

How I Tested It

I ran GPT-5.6 through four real-world scenarios that I handle weekly as an editor: researching and summarizing a complex topic, drafting an editorial brief from source documents, analyzing a spreadsheet of reader data, and generating a presentation outline with visual notes.

I tested Sol on medium reasoning and Terra on default settings. I did not test Luna because for my workload the budget tier would not save enough time to justify the switch.

The One Model That Changed My Mind

Terra surprised me the most. I expected Sol to be the clear winner but Terra handled every task I threw at it with speed that felt immediate. On the research task Terra pulled information from multiple sources and produced a structured summary in under 30 seconds. Sol on medium reasoning took about 45 seconds and delivered a marginally better result. The difference was not worth 2x the price for my work.

The real breakthrough came with ChatGPT Work. Previously I used ChatGPT for conversation and Codex for coding tasks. Work sits between them. It accepts messy context from documents, Slack exports, and Google Drive files and turns them into polished drafts. I gave it a folder of interview notes and asked for an editorial brief. It produced a coherent outline with section headings, source citations, and suggested pull quotes. That would have taken me an hour.

Where Sol Justifies Its Price

I found one scenario where Sol pulled ahead by enough to justify the cost: multi-step analysis that requires revisiting earlier conclusions. I gave both models a spreadsheet with three months of reader engagement data and asked for trend analysis with recommendations. Terra delivered a solid first pass. Sol with medium reasoning caught a logical error in the recommendations, went back to the data, and corrected itself without being prompted.

Sol also has access to ultra mode which runs four agents in parallel. I tested this on a complex research brief that required analyzing five competitor articles, extracting data points, and synthesizing a recommendation. Ultra finished in roughly half the time of a single Sol agent and the output was noticeably more complete. This mode is available on ChatGPT Work for Pro and Enterprise users.

What Still Needs Work

GPT-5.6 is not perfect. The safeguards are noticeably more aggressive than GPT-5.5. OpenAI says its cyber safeguards block roughly ten times more potentially harmful activity compared with previous models. That is good for security but it means some legitimate prompts get blocked or require retrying on a lower-capability model. I hit this twice during testing when asking for analysis of cybersecurity news.

The prompt caching is also more complex. GPT-5.6 introduces explicit cache breakpoints and a 30-minute minimum cache life. Cache writes are billed at 1.25x the uncached input rate. If you are a heavy API user you need to understand these changes or your bill could surprise you.

Pricing That Makes Sense

Sol costs $5 per million input tokens and $30 per million output tokens. Terra costs $2.50 input and $15 output. Luna costs $1 input and $6 output. For context GPT-5.5 was $5 input and $30 output at its peak tier. Terra at half the price matching GPT-5.5 performance is the real story for most professionals.

ChatGPT Plus and Pro users get access to Sol through medium and higher effort settings. Pro and Enterprise users can also select Sol Pro for the highest quality on complex tasks. ChatGPT Work and Codex default to Terra for Free and Go users while Plus and above can choose among all three tiers.

Bottom Line

GPT-5.6 Terra is the model most people should use. It matches GPT-5.5 at half the cost and handles daily editing, research, and drafting tasks without the premium price of Sol. Sol is worth the upgrade only if you do multi-step analytical work that requires self-correction or if you need the parallel agent capability in ultra mode. Luna is a solid option for high-volume simple tasks. The real improvement over GPT-5.5 is not the benchmark scores. It is that ChatGPT Work finally makes AI useful for actual office work instead of just conversation.