August 9, 2026. OpenAI shipped an updated GPT-5.6 Sol into ChatGPT on August 6, made GPT-5.6 Luna the default for free accounts, and said free users will get unlimited text chats the week after. It was the biggest consumer AI release of the week and it will be quoted at you by clients. Before it is, read the last paragraph of the announcement, because it contains a sentence that most coverage of this release has skipped: the version of GPT-5.6 Sol that powers ChatGPT Work and Codex is not changing as part of this release.
What OpenAI actually shipped on August 6
- An updated GPT-5.6 Sol for Plus and Pro, in Chat only. OpenAI's product post describes more focused answers, tighter formatting, less unnecessary detail, and a model that offers a correction rather than agreeing when agreement would not help.
- One model across Instant and Thinking. For Plus and Pro, the same model now powers both quick responses and deeper reasoning, so moving up the effort scale should read as more time on the same system rather than a handoff to a model with a different voice.
- A reasoning slider. Plus and Pro users get a slider on web, mobile and desktop to set how much thought goes into an answer.
- GPT-5.6 Luna becomes the free default. OpenAI says Luna rolls out as the default for Free and Go users during the week of August 6.
- Unlimited text chats, the week after, with conditions. Free and Go users get unlimited text chats and a new Think button starting the week of August 10, subject to abuse guardrails. Limits still apply to file uploads, images and other tools.
- An accuracy claim with a named baseline. In an internal evaluation of financial, medical and legal prompts requiring factual detail, responses containing at least one factual error were about 62 percent less common with GPT-5.6 Luna and 68 percent less common with GPT-5.6 Sol than with GPT-5.5 Instant.
- A system card covering under-18 users. OpenAI published a system card describing training to avoid romantic roleplay and to apply age-appropriate boundaries for users it believes are under 18.
The release was reported the same day by TechCrunch and others, mostly under the unlimited-free-chats headline.
The sentence that changes who this release is for
OpenAI is explicit about the scope. Because this version of GPT-5.6 Sol is tuned for everyday chats, it is available only in the Chat experience in ChatGPT, and the version of Sol that powers Work and Codex is not changing in this release. The API is not mentioned at all.
That is three separate deployments of a model with the same name. The one that got the accuracy work is the one your prospect uses on their phone. The one your client's team logs into at work is unchanged. The one your product calls through the API is unchanged and was not part of the announcement. A model name is not a deployment, and this release is the cleanest recent proof of it.
The accuracy gain is measured against an older model
The 62 and 68 percent figures are stated plainly and OpenAI names the comparison: GPT-5.5 Instant. That matters in both directions. It is a genuine reduction in factual errors on a hard prompt set, and it is a reduction relative to an older model, not relative to the GPT-5.6 many teams already moved to after the July 30 price-performance release. Anyone telling a client this release means 68 percent fewer errors than what they have today is asserting something the announcement does not say, unless what they have today is GPT-5.5 Instant.
The evaluation is also OpenAI's own, on prompts OpenAI selected, in three domains it names. That is normal for a launch post and it is worth saying out loud when you repeat the number, because the alternative is presenting a vendor's internal benchmark as an independent one.
Three rollout dates means three different systems this month
The post stages the release across three moments: the updated Sol and the slider for Plus and Pro on August 6, Luna as the free default during the week of August 6, and unlimited text chats plus the Think button the week after. A team that runs a comparison today, a colleague who runs the same comparison on Monday, and a client who tries it the following Thursday are not testing the same product. If you publish a bake-off, date-stamp it and name the tier, or it will be wrong within a fortnight and someone will screenshot it.
The free tier deserves one more note. Unlimited applies to text chats. File uploads, images and other tools keep their limits, which is exactly where a free-tier user runs into a wall while trying to do the work that made them upgrade in the first place. If your funnel assumes a free-tier prospect can process a document, that assumption did not change this week.
What it means for operators
This is the week a client forwards a headline and asks why the assistant you built for them did not get better. The honest answer is that the thing that improved is a different deployment of a model with a shared name, and that your build changes when you change it. Three concrete moves.
Pin the model string and say so in writing. If your automation calls a model by a floating alias, a vendor-side swap can change your outputs with no deploy on your side. Pin the version, log which version produced every client-visible output, and put the upgrade decision on your own calendar. This is the single largest reliability difference between a demo and a production AI automation build, and it is the first thing we set up when we put an engineer on a client stack.
Write evals against your surface, not the vendor's. A launch-post benchmark on financial, medical and legal prompts tells you very little about your intake form, your enrichment step or your reply classifier. Twenty to fifty real examples from your own workflow, scored the same way every time, will tell you within an hour whether a model change helps you. Teams shipping a product on top of a model should treat that eval set as part of the codebase, not as a one-off spreadsheet, and it belongs in the plan before you launch anything on top of an API.
Separate the consumer story from the deployment story when you talk to clients. Consumer releases move fast and are genuinely useful as a signal of where capability is going. They are not a schedule for your roadmap. Saying that clearly, in the same message where you acknowledge the news, is what a competent partner sounds like.
The open question worth putting in writing
The announcement says Work and Codex are not changing in this release. It does not say when they will, or whether the Chat tuning is intended to reach them at all, and the same silence covers the API. If your business depends on the Work surface or on API behaviour, that is a fair thing to ask OpenAI in writing rather than to infer from a consumer post. The two things worth asking for specifically: whether the reduced-error tuning is planned for the Work and Codex versions of GPT-5.6 Sol, and whether any API-facing model version changed on or around August 6. Until either is answered, assume nothing on your side moved, and verify with your own eval set rather than with a headline.
Frequently Asked Questions
No. OpenAI's August 6, 2026 announcement states that because this version of GPT-5.6 Sol is optimized for everyday chats, it is available only in the Chat experience in ChatGPT, and that the version of GPT-5.6 Sol powering Work and Codex is not changing as part of this release. The API is not mentioned in the announcement at all.
OpenAI reports that in an internal evaluation of financial, medical and legal prompts requiring factual detail, responses containing at least one factual error were about 62 percent less common with GPT-5.6 Luna and 68 percent less common with GPT-5.6 Sol than with GPT-5.5 Instant. The baseline is GPT-5.5 Instant, not GPT-5.6, and the evaluation is OpenAI's own.
OpenAI says GPT-5.6 Luna becomes the default for Free and Go users during the week of August 6, 2026, and that unlimited text chats plus a new Think button arrive the following week, subject to abuse guardrails. The word text is load-bearing: limits still apply to file uploads, images and other tools.
Not unless your automation calls the ChatGPT Chat experience, which almost no automation does. Production builds call the API or a platform surface, and neither was part of this announcement. If your outputs change without you deploying anything, that usually means you are calling a floating model alias rather than a pinned version, which is worth fixing regardless.
Run your own evaluation first. A vendor benchmark on financial, medical and legal prompts predicts very little about your specific workflow. Twenty to fifty real examples from your own pipeline, scored consistently, will answer the question in about an hour and will keep answering it at the next release.
It is a control for Plus and Pro users, available on web, mobile and desktop, that sets how much thought ChatGPT puts into a given answer. OpenAI pairs it with a change that makes the same model power both Instant and higher-effort responses, so raising the effort should feel like more time on one system rather than a switch to a different model with its own tone.