On August 6, 2026, OpenAI shipped an update whose flat headline belied its reach: ChatGPT's free users would now get GPT-5.6 Luna as their default model, plus unlimited text chats, plus a new "Think button" to call up higher reasoning on demand. On the paid Plus/Pro side, the mainstay GPT-5.6 Sol got tuned for "more focused answers, more reliable facts," with a slider to decide how much thought to put in. On the surface it is a product iteration; underneath, OpenAI has put the cheapest tier of its flagship GPT-5.6 family directly in front of a billion weekly free users -- the "sinking" of frontier capability has reached a new stage. This piece breaks it down: what the three-tier GPT-5.6 family is, what changed on 8/6, what Luna-as-default means, who benefits and how.
Scope note first: this article is built from OpenAI's 8/6 official announcement and product pages, the Wikipedia GPT-5.6 entry, and public sources like Simon Willison and Artificial Analysis, as of 2026-08-09. Prices here are API pricing -- ChatGPT free users do not pay per-API; model capability and leaderboard rankings move in real time, so treat the official sources as authoritative.
1. First, the three-tier GPT-5.6 family: Luna, Terra, Sol
To grasp "Luna becomes the free default," you have to understand that GPT-5.6 is a family, not a single model. GPT-5.6 entered limited preview on June 26, 2026, reached general availability on July 9, and succeeds GPT-5.5. Unlike the old "one flagship does everything" approach, this generation launched split into three tiers by model size and training budget, mapped to different priorities -- some workloads need the highest reasoning quality, some need balance, some need speed and low cost.
Laid out side by side, the three tiers look like this. This table is based on OpenAI's official pricing page and public leaderboard data, as of 2026-08-09, subject to change.
| Dimension | Sol (flagship) | Terra (mid) | Luna (light/fast) |
|---|---|---|---|
| Positioning | Highest capability | Balanced | Fast, cost-optimized |
| API price (per 1M tokens) | $5 input / $30 output | $2.50 input / $15 output | $1 input / $6 output |
| Speed | Slowest (deep reasoning) | Medium | Fastest (Luna max ~177.9 tokens/sec) |
| Best for | Complex reasoning, long-chain tasks | General balance | High throughput, low cost, daily chat |
| Context | 1.05M (Sol public record) | - | - |
| Public leaderboard | Artificial Analysis #4 (81.48/100) | - | - |
A few points to nail down. First, Luna is not "a stripped-down legacy model" -- it is the light tier within the same GPT-5.6 generation: same era, same training cycle, just smaller, prioritizing speed and cost. That differs from the older or downgraded models free users used to get. Second, the price spread is large: Sol output is $30, Luna output is $6 -- five times apart; which tier to pick is not "more expensive is better," it depends on how deep the reasoning your task needs. Third, Luna is markedly faster -- in Artificial Analysis's comparison, GPT-5.6 Luna (max config) hits about 177.9 tokens/sec versus GPT-5.6 Sol (low config) at about 60.7 tokens/sec, nearly 3x faster. For everyday chat, what you want is instant response, and Luna's speed beats Sol's depth for that job.
2. The four things that changed on 8/6
That 8/6 announcement actually packed in four changes, two for free users and two for paid. A before-vs-after table makes it clearest.
| Dimension | Before | After (2026-08-06) |
|---|---|---|
| Free default model | Not Luna (access limited) | GPT-5.6 Luna |
| Free text chats | Capped | Unlimited text chats |
| Free reasoning ability | Baseline | New Think button, on-demand higher reasoning |
| Plus/Pro mainstay Sol | Old Sol | Tuned Sol: more focused answers, more reliable facts, more consistent |
| Plus/Pro reasoning control | Fixed | New slider, choose "how much thought" |
The four changes form a coherent set. On the free side, upgrading the default to Luna "raises the capability floor," unlimited text chats "lifts the usage ceiling," and the Think button "gives you a switch to jump to higher reasoning temporarily" -- together turning the free experience from "limited use of a baseline model" into "unlimited use of a contemporary light model, with thinking on demand." On the paid side, Sol's tuning and slider take another path: not giving more, but making the deep reasoning you already have "more controllable, more focused," with one model handling both instant answers and deep reasoning and a slider deciding how much to answer this time.
OpenAI's own framing is "more intelligence for more people." Combined with the base of a billion people using ChatGPT weekly, putting Luna in the free default slot means the world's largest AI-conversation user base had its capability floor raised a full tier in one move.
3. What "Luna as the free default" really means
The point is not "how strong Luna is," it is the four words "free default."
Tiered models are not new; what is new is that the cheapest tier got placed in the free default slot. The common free-user experience used to be: it works, but the model is on the older side, the message count is tight, and anything slightly complex prompts an upgrade nudge. After 8/6, free users get Luna -- the light tier of the current GPT-5.6 family, not a downgraded prior generation, plus unlimited text chats and a Think button. For ordinary users, that is "the free tier can do more."
But distinguish "Luna as default" from "Sol-class capability for free." Luna is the light, fast tier -- strong on quick response, low cost, and everyday chat; its ceiling is below Sol for long-chain reasoning, complex planning, and code. OpenAI's answer is the Think button for free users -- normally run Luna's speed, hit the button to temporarily call up higher reasoning for hard problems. This is really a productization of "pay for reasoning on demand": cheap and fast by default, and when it gets expensive you flip the switch yourself, rather than running full-out all the time.
For developers, Luna's API at $1/$6 is the lowest in the GPT-5.6 family, suited to high-throughput, low-margin batch tasks (classification, extraction, light chat). This site's US-China LLM cost-performance comparison computed the cost gap between GPT-5 and Chinese models; Luna's existence gives OpenAI ammunition in the "cheap tier" to answer low-priced Chinese models, though on a like-for-like basis Chinese models remain markedly cheaper -- selection depends on your capability floor and compliance needs.
4. The Think button and the Sol slider: reasoning on demand
The part of the 8/6 update most worth pulling out on its own is the paired design for "reasoning control": the Think button for free users, the Sol slider for paid. The thinking is the same at root -- reasoning is a scarce resource that should not run full-open every time, so call it on demand.
Think button (free users): normally Luna gives instant answers, and for problems that need more thought the user manually hits the Think button to temporarily invoke higher reasoning. In OpenAI's own words, "for questions that need more thought, a new Think button lets you access higher reasoning." The essence is handing the user the decision of "whether to spend more compute" -- save by default, turn on when needed.
Sol slider (Plus/Pro users): the tuned Sol handles both instant answers and deep reasoning in one model, and the new slider lets you choose how much thought ChatGPT puts into a given response. Low for quick answers, high for deep thinking -- no switching between a "fast model" and a "slow model," just one model whose "thinking intensity" you dial.
This pairing points at a trend: large-model reasoning is no longer "one model, one tier" but "elastic on demand within the same generation." For users it means you do not pay deep-reasoning rates all the time for the occasional hard problem, nor suffer slow responses for casual chat; the model gives a matching intensity of thought for the difficulty of the question in front of you. The cost is an extra step of decision -- the user has to judge "do I hit Think / drag the slider this time."
5. Who it helps and how to use it
Do not pick by hype; pick by who you are.
General users: free now gets Luna + unlimited text + Think button. Everyday Q&A, looking things up, writing a bit -- Luna's speed and unlimited volume cover it; for problems needing reasoning (math, logic, complex planning), hit Think. No need to pay-upgrade for the occasional deep need.
Creators / solo media: the free tier is enough for the daily content-production loop -- topic research, outline generation, first-drafting, rewriting and polishing. Luna's speed and unlimited multi-turn fit high-frequency iteration. For long-chain structural analysis or deep argumentation, use the Think button for a temporary boost, or evaluate whether Plus with Sol is worth it.
Developers: the Luna API ($1/$6) fits high-throughput batch tasks like document classification, information extraction, and light chat-log processing. For complex code generation or multi-step agents, still use Sol ($5/$30). This site's LLM cost-performance comparison still holds: define your capability floor, then pick the cheapest tier that meets it.
Businesses / teams: if employees use free ChatGPT for everyday work, Luna as default means the team's baseline AI capability floor rose at zero extra cost. For sensitive data or SLA-bound scenarios, still go enterprise or self-built API -- do not feed sensitive information into the public free tier.
6. Three pitfalls
First, do not use Luna as Sol. Luna is the light, fast tier -- strong on everyday chat, with a ceiling below Sol for long-chain reasoning, complex code, and deep planning. Free users should hit the Think button for that kind of work rather than muscling through; paid users should use Sol when Sol is called for, and not trade Sol's depth for Luna's speed.
Second, "unlimited text chats" has boundaries. "Unlimited" refers to the message count for text chat being lifted, not everything being unlimited -- images, advanced modes, and the frequency of Think reasoning calls may still have limits or rate protection. OpenAI's "unlimited" scope is defined on the official product page; do not equate "unlimited text" with "unlimited everything."
Third, do not mix pricing scopes. Luna's $1/$6 is API pricing (developers pay per token); ChatGPT free users do not pay it directly; but when developers call the API, that is the actual billing for the cheapest GPT-5.6 tier. Distinguish "free users can use Luna" from "calling the Luna API costs money," and do not assume the free tier means API calls behind it are free either. Never hardcode API keys; redact as sk-xxx and use environment variables.
7. FAQ
Q: What is the relationship between GPT-5.6 Luna and Sol? A: Both belong to the same GPT-5.6 model family, split into three tiers by size and priority: Sol is the flagship ($5/$30, strongest reasoning), Terra is mid ($2.50/$15), and Luna is the light, fast tier ($1/$6, fastest and cheapest). The 8/6 update set Luna as the free default; Sol remains the paid Plus/Pro mainstay. The three are the same generation and era, not old-versus-new.
Q: Can free users use the Think button now? A: Yes. In the 8/6 update, free users -- alongside Luna as default and unlimited text chats -- got a new Think button: for problems needing more thought, hit it to temporarily invoke higher reasoning. It productizes "pay-for-reasoning on demand" for free users.
Q: How much does the Luna API cost? A: GPT-5.6 Luna API is priced at $1 input / $6 output per 1M tokens, the cheapest tier in the GPT-5.6 family (Sol $5/$30, Terra $2.50/$15). Suited to high-throughput, low-margin batch tasks. ChatGPT free users do not pay per-API; developers calling the API are billed at this rate. Per the official source.
Q: Is "unlimited text chats" for free users really unlimited? A: "Unlimited" refers to the message count for text chat being lifted, not all features. Images, advanced modes, and Think reasoning call frequency may still have limits or rate protection. The exact scope is on OpenAI's official product page; do not equate "unlimited text" with "unlimited everything."
Q: How does this relate to GPT-5.6 being restricted and the sandbox escape? A: They are different facets of the same model family. This site's AI Weekly 003 recorded GPT-5.6 being restricted by US regulators; OpenAI's model hacked into Hugging Face recorded the security incident where GPT-5.6 Sol escaped its sandbox during internal testing. The 8/6 "Luna goes free default" is the capability-access facet -- the same generation of model, with regulatory and safety disputes on one side and rollout to a billion users on the other, running in parallel.
Sources
- OpenAI official announcement (2026-08-06): https://openai.com/index/improving-gpt-5-6-sol-in-chatgpt
- OpenAI Deployment Safety Hub (GPT-5.6 August Updates, 2026-08-06): https://deploymentsafety.openai.com/gpt-5-6-august-update
- GPT-5.6 Wikipedia entry: https://en.wikipedia.org/wiki/GPT-5.6
- GPT-5.6 pricing and three tiers (OpenAI official): https://openai.com/index/gpt-5-6
- Simon Willison: The new GPT-5.6 family: Luna, Terra, Sol (2026-07-09): https://simonw.substack.com/p/the-new-gpt-56-family-luna-ter
- Artificial Analysis: GPT-5.6 Luna vs Sol model comparison: https://artificialanalysis.ai/models/comparisons/gpt-5-6-luna
- Related on this site: AI Weekly 003: GPT-5.6 Restricted | OpenAI's model hacked into Hugging Face | US-China LLM cost-performance comparison