OpenAI is swapping out the engine behind most ChatGPT conversations. Starting today, GPT-5.5 Instant is on its way out as the everyday chat model, replaced by two members of the GPT-5.6 family that launched in July.
Free and Go users are moving to GPT-5.6 Luna, and OpenAI says they will get unlimited text chats next week. Plus and Pro subscribers get an updated GPT-5.6 Sol tuned for everyday chat, along with a new slider that sets how hard the model thinks before it answers.
The update is small in name and big in reach. OpenAI says about 1 billion people use ChatGPT every week, so changing the default model changes the product most of them see.
What OpenAI announced
The announcement came in a company blog post and a matching system card on OpenAI's Deployment Safety Hub. Both say the new models "will replace GPT-5.5 Instant" in ChatGPT.
The rollout is staggered. Here is the order OpenAI laid out:
- Today: Plus and Pro users get the updated GPT-5.6 Sol and the effort slider on web, mobile and desktop.
- This week: GPT-5.6 Luna becomes the default model for Free and Go users.
- Next week: Free and Go users get unlimited text chats and a new Think button for harder questions, "subject to abuse guardrails."
OpenAI says limits will still apply to file uploads, images and other tools. Only plain text chat becomes unmetered.
GPT-5.6 Sol: one model, adjustable effort
For paying users, the headline feature is the slider. Instead of switching between separate "instant" and "thinking" models, Plus and Pro users now pick how much reasoning a single model uses on each response.
OpenAI's pitch is consistency. The company says the same updated Sol now powers both quick replies and deeper reasoning. Moving the slider up should feel like the model "taking extra time for a more comprehensive answer," not like switching to a model with a different tone.
At launch, OpenAI's help center describes the stops as Instant, Medium, High and Extra High, with the deepest settings tied to more expensive plans. Check the model menu in your own account, since stops and plan limits can shift while the rollout finishes.
OpenAI also says it retuned Sol's style. The company claims it now gives more focused answers, adjusts how much detail it gives to the question, uses less unnecessary formatting and will correct the user "when simply agreeing wouldn't be useful." That last point targets the flattery problem that has dogged chatbots for years.
The Work and Codex caveat
This August version of Sol lives only in the Chat part of ChatGPT. OpenAI says the Sol and Luna versions that power ChatGPT Work and Codex are not changing.
The system card tells them apart by release month. The "August" models are the new chat-tuned versions, and the "July" models stay in Work and Codex. If you use ChatGPT for both chat and agent-style work, you are now using two different builds of Sol.
GPT-5.6 Luna: a bigger free tier
The free-tier change may matter most. GPT-5.6 Luna, the fastest and cheapest tier of the GPT-5.6 family, becomes the default for Free and Go accounts, and OpenAI is removing the cap on text messages.
TechCrunch reported the change as OpenAI removing limits on text-based chats for Free and Go users, with separate limits still applying to files, images, voice and image generation.
The new Think button is the free-tier version of the slider. Tapping it gives Luna more time to work through a hard question. According to OpenAI's help documentation, Think still runs on Luna. Free and Go users do not get access to Sol.
Who gets what
| Plan | Default chat model | Reasoning control | Rollout timing |
|---|---|---|---|
| Free | GPT-5.6 Luna (replaces GPT-5.5 Instant) | Think button | Luna this week; unlimited text and Think next week |
| Go | GPT-5.6 Luna (replaces GPT-5.5 Instant) | Think button | Luna this week; unlimited text and Think next week |
| Plus | Updated GPT-5.6 Sol | Effort slider | Starting today |
| Pro | Updated GPT-5.6 Sol | Effort slider, including higher settings | Starting today |
| Work and Codex (all plans) | July GPT-5.6 builds | Unchanged | No change in this release |
Fewer factual errors, by OpenAI's count
OpenAI's biggest quality claim is about accuracy. In an internal evaluation of financial, medical and legal prompts that need factual detail, the company says responses with at least one factual error were about 62% less common with GPT-5.6 Luna and 68% less common with GPT-5.6 Sol than with GPT-5.5 Instant.
The system card gives more detail. It describes three hard prompt sets: factuality-heavy production-style prompts, conversations users had flagged as containing errors, and high-stakes medical, legal and financial questions.
On those sets, OpenAI says Sol cut factual error rates by roughly 60% across all three. Luna cut them by more than 60% on high-stakes prompts and by roughly 30% on the other two.
Key takeaway: the free-tier model is not a stripped-down afterthought. By OpenAI's own numbers, Luna makes far fewer factual errors than the GPT-5.5 Instant model it replaces, though Sol's gains are larger and more consistent.
OpenAI stresses that these prompt sets were chosen because models tend to fail on them. They do not show how often ChatGPT gets things wrong in everyday use, and none of the figures have been checked independently yet.
Health answers get a measurable bump
Health questions are one of ChatGPT's most sensitive uses, and the system card reports gains on OpenAI's HealthBench suite. OpenAI adjusts these scores for answer length so longer replies don't score higher just for saying more.
| Evaluation (length-adjusted) | GPT-5.5 Instant | GPT-5.6 Sol (August) | GPT-5.6 Luna (August) |
|---|---|---|---|
| HealthBench | 51.4 | 55.0 | 53.3 |
| HealthBench Hard | 22.9 | 31.4 | 28.7 |
| HealthBench Professional | 38.4 | 54.0 | 44.1 |
The biggest jump is on HealthBench Professional, which is built around clinician-style use cases. OpenAI says Sol improved by 15.6 points there. Answers also got shorter on most of the health tests.
Safety designations and the fine print
OpenAI's Preparedness Framework rates models on how much risk they pose in sensitive areas. The updated Sol and Luna keep the same ratings as the July GPT-5.6 models: High in biological and chemical, High in cybersecurity, and below High in AI self-improvement.
A High rating does not mean the model is blocked. It means OpenAI says the model is capable enough in that area that it has to ship with stronger safeguards, and the company says it has applied the same safeguard package described in the original GPT-5.6 system card.
The cyber results show why. On OpenAI's internal capture-the-flag hacking tests, the system card says the August Sol "saturates the evaluation at 97.06%." In OpenAI's cyber range exercises, where a model must carry out an attack from start to finish on a simulated network, Sol passed 83.3% of scenarios and Luna passed 61.5%.
The system card also flags some regressions:
- On disallowed-content tests, Sol performed about the same as GPT-5.5 Instant except for gore and disallowed sexual content. OpenAI says it adds a system-level filter to stop graphic erotic content in production.
- On multi-turn mental health tests with simulated adversarial users, Sol showed a statistically significant regression on self-harm. OpenAI says it did not see more undesirable responses during online testing and will keep monitoring.
- Luna showed a minor regression on an image-input extremism test, which OpenAI describes as having low statistical significance.
For the first time, OpenAI is also publishing dedicated evaluations for users it believes are under 18. The company says it trained the models to avoid romantic roleplay with teens, age-restricted challenges and presenting themselves as a substitute for real relationships, and it added system-level protections on top.
What it means for developers
For API customers, nothing breaks today. OpenAI's announcement is about ChatGPT. The GPT-5.6 models in the API, launched July 9 as gpt-5.6-sol, gpt-5.6-terra and gpt-5.6-luna, are not being replaced.
API pricing was last changed on July 30, when OpenAI cut Luna by 80% and Terra by 20%. Current list prices per million tokens:
| API model | Input | Output |
|---|---|---|
| GPT-5.6 Sol | $5.00 | $30.00 |
| GPT-5.6 Terra | $2.00 | $12.00 |
| GPT-5.6 Luna | $0.20 | $1.20 |
There are still two practical points. First, if your product copies ChatGPT's tone, the August models are a new target: shorter, less formatted and more willing to push back. Second, OpenAI has previously offered a chat-latest API alias that tracks the model used in ChatGPT, and it recommends against using that alias in production. If you rely on it, expect its behavior to change.
The effort slider also matches how the API already works. gpt-5.6-sol already accepts a reasoning effort parameter, so ChatGPT is showing consumers the same speed-versus-depth trade-off developers already manage.
Why this matters
Two things stand out. First, OpenAI is spending heavily on its free tier. Unlimited text chat on a current-generation model is a strong offer, and it raises the bar for rival free chatbots.
Second, the slider is part of a wider trend. Instead of a menu of model names, labs are moving toward one model with adjustable effort. That is easier to understand, but it moves the cost decision to the user. How much you get out of Sol now partly depends on where you set the slider.
The split between Chat and Work also deserves attention. Two builds of the same model name, tuned for different jobs, could confuse users who compare answers across products.
FAQ
When do free ChatGPT users get GPT-5.6 Luna?
OpenAI says Luna becomes the default for Free and Go users this week. Unlimited text chats and the Think button arrive next week. Limits still apply to file uploads, images and other tools.
Do Free users get access to GPT-5.6 Sol?
No. According to OpenAI's help documentation, Free and Go accounts use Luna for both the default experience and the Think button. Sol and the effort slider are for paid plans.
Does this change GPT-5.6 in Codex or ChatGPT Work?
No. OpenAI says the Sol and Luna versions that power Work and Codex stay on the July builds. The August update covers only the Chat experience.
How big is the accuracy improvement?
On OpenAI's internal test of financial, medical and legal prompts, responses with at least one factual error were about 62% less common with Luna and 68% less common with Sol than with GPT-5.5 Instant. These are company-reported results on deliberately hard prompts, not independent measurements.
Is the API price changing?
Not with this release. The GPT-5.6 API lineup and the July 30 price cuts for Luna and Terra stay in place.
The bottom line
The August GPT-5.6 update retires GPT-5.5 Instant and gives ChatGPT's free tier a much more capable default with no cap on text chat. Paying users get a retuned Sol and direct control over how hard it thinks.
OpenAI's accuracy and health numbers are impressive but self-reported, and the system card is open about some safety regressions. Over the coming weeks, watch how Luna holds up under unlimited free use, and whether the split between Chat and Work builds leaves users with inconsistent answers.
Join the conversation