Sol, Terra, Luna — OpenAI has named its newest models after the sun, the earth and the moon, and the metaphor is doing real work. The GPT-5.6 family, unveiled in a limited preview in late June 2026, is not one model but a three-step ladder of capability and cost: a frontier flagship, a workhorse, and a budget sprinter. For small businesses that have struggled to match the right model to the right job, that ladder may matter more than any benchmark score.
What is in the GPT-5.6 preview
OpenAI describes GPT-5.6 Sol as its most capable model yet, built for frontier reasoning and long-horizon agentic work — the kind of multi-step, tool-using tasks that power modern business automation. GPT-5.6 Terra is the middle tier, delivering performance competitive with GPT-5.5 at roughly half the cost. GPT-5.6 Luna is the lightweight option, optimized for speed and everyday tasks. According to OpenAI, the family advances the state of the art in software engineering, computer use, professional knowledge work, scientific research and cybersecurity.
One important caveat: this is a limited preview. The three models are currently available through the OpenAI API and Codex to a restricted group of partner organizations, and GPT-5.6 is not yet in ChatGPT. OpenAI says general availability is coming in the following weeks.
The new price ladder, in plain numbers
Pricing is where the announcement gets concrete. Sol is listed at $5 per million input tokens and $30 per million output tokens; Terra at $2.50 and $15; Luna at $1 and $6. In other words, the cheapest tier costs a fifth of the flagship — and Terra effectively offers last quarter’s flagship performance at half of last quarter’s price.
That is the same downward cost curve that appeared when Claude Sonnet 5 changed the math on AI agents for small business: each generation quietly halves the price of “good enough,” and the savings compound for any organisation running AI tasks at volume.
Ultra mode and deeper reasoning
Two new controls stand out. A “max reasoning effort” setting lets Sol take more time to think before answering — useful for genuinely hard problems, wasteful for routine ones. And “ultra mode” goes beyond a single agent by spinning up subagents that split complex work and run it in parallel.
Both are dials, not defaults, which reinforces the core lesson: cost discipline in AI is now a configuration decision, and an unmanaged setting can multiply a bill as quickly as a well-chosen tier can shrink it. That is a governance question as much as a technical one — the same reason most AI agent projects stall without clear governance.
What a small business should actually do
The sensible sequence starts with mapping workloads before the general release. Routine drafting, summarising and customer-facing chat almost certainly belong on a Luna-class tier; complex analysis and agent workflows may justify Terra; very few small-business tasks need a frontier flagship at flagship prices. Nothing should be rebuilt on a preview model — stable pricing and terms only arrive at general availability. And model families now turn over quickly: GPT-4.5 was retired with little warning, so any new build should assume its model will eventually be swapped out.
Limitations and what to watch
Everything above describes a preview, and previews change. The published prices, the benchmark claims and even the model line-up could shift before general availability, and OpenAI’s performance comparisons have not yet been widely validated by independent testing. Access is currently limited to a small set of approved partner organisations, so most firms cannot verify the claims first-hand. Worth watching: the actual ChatGPT and API release, independent benchmark results once access widens, and how competitors respond on price.
The takeaway
GPT-5.6 is less a single launch than a pricing philosophy: pick the smallest model that does the job, and reserve expensive reasoning for the problems that earn it. The businesses that benefit most will not be the ones using the biggest model — they will be the ones matching each task to the smallest model that does the job well.
