OpenAI Refreshes GPT-5.6 Sol with 68% Fewer Factual Errors and Unlimited Free Access

30-Second Summary

OpenAI updated GPT-5.6 Sol on August 6, 2026, cutting factual errors by 68% versus the previous default model, GPT-5.5 Instant. Free ChatGPT users gained unlimited text conversations with GPT-5.6 Luna. Plus and Pro subscribers received a five-level reasoning depth slider replacing the old binary thinking toggle. The moves sharpen the value divide between tiers and push AI reliability closer to the level business operators need for unsupervised employee use.

At a Glance

• Topic: Model Releases
• Company: OpenAI
• Date: August 6, 2026
• Announcement: GPT-5.6 Sol refreshed with accuracy improvements; free tier upgraded to unlimited GPT-5.6 Luna text conversations
• What Changed: Sol produces 68% fewer factual errors than GPT-5.5 Instant in OpenAI's internal testing; free users lose message caps on text; paid users gain granular reasoning control
• Why It Matters: The accuracy threshold for employee-trusted AI output moved closer to practical reliability; the free unlimited tier will accelerate shadow AI adoption across organisations that have not formalised governance
• Who Should Care: Any operator with employees using ChatGPT for client work, internal documents, or research; HR and IT leads managing AI tool access; any team lead whose staff already use consumer AI at work

Key Facts

• 68% reduction in factual errors compared to GPT-5.5 Instant, measured across financial, medical, and legal prompts in OpenAI's internal testing. These figures have not been independently verified.
• Unlimited text chats now available to free ChatGPT users with GPT-5.6 Luna, rolling out in the week of August 6. File uploads and image generation remain restricted.
• Five reasoning levels available to Plus and Pro subscribers via a new slider, replacing the previous binary instant-or-thinking mode.
• GPT-5.6 Sol is now the default model for Plus and Pro subscriptions. It is designed to give concise, direct answers without unnecessary formatting.
• Luna pricing was cut 80% on July 30, 2026, to $0.20 per million input tokens. Sol remains at $5 per million input tokens.
• 1 billion weekly users of ChatGPT as of August 2026, per OpenAI's own reporting.

What Happened

OpenAI rolled out a refresh to GPT-5.6 Sol on August 6, 2026. The update addressed one of the persistent criticisms of the GPT-5.6 family since its general availability in July: verbose, over-formatted responses that added length without adding value. Sol now produces more direct answers and omits detail that was not requested. The most significant claimed improvement is factual accuracy: in OpenAI's internal testing across financial, medical, and legal prompts, responses containing at least one factual error were 68% less common with the updated Sol than with GPT-5.5 Instant, the model it replaced as the default.

Simultaneously, OpenAI extended unlimited text conversations to free ChatGPT users. Previously, free accounts faced hard limits on the number of messages they could send. From the week of August 6, those caps are removed for text-based chats using GPT-5.6 Luna. The move follows the 80% price cut to Luna's API pricing announced on July 30, suggesting OpenAI is deliberately expanding the top of its user funnel. Luna is the most cost-efficient model in the GPT-5.6 family and sits below Sol in capability.

For paid subscribers, the most operationally significant change is the reasoning depth slider. The previous system gave users a binary choice between instant responses and a full thinking mode. The new five-level slider lets users tune how long Sol takes to reason before responding, giving teams the ability to calibrate quality versus speed for different task types within the same subscription.

OpenAI said the Sol refresh also reflects improvements in the model's ability to optimise its own production code, which contributed to internal efficiency gains that have been passed through as the July price cuts on Luna and Terra.

Why It Matters

Factual reliability is now the competitive axis. The era of competing on benchmark scores is giving way to competition on operational reliability. A 68% reduction in factual errors in the domains that matter most for business, finance, medicine, and law, is a meaningful claim. If it holds up in independent testing, it represents the moment when AI moves from "useful with supervision" to "trustworthy for first-draft production" in those domains.

The free unlimited tier accelerates unsanctioned AI adoption. Removing message caps means every employee with a personal Gmail account now has access to a capable AI with no usage friction. For operators who have not formalised their AI governance, this is the event that makes the status quo untenable. Staff will use free ChatGPT for work tasks regardless of company policy if no alternative is provided and no policy exists.

The tier divide now has real operational meaning. Luna (free) and Sol (paid) were previously separated mainly by speed and capability. The accuracy gap adds a new dimension: free is fine for casual use, paid is required for accuracy-sensitive work. That distinction gives operators a defensible reason to budget for Plus or Pro subscriptions rather than accepting that employees will use the free tier for everything.

The reasoning slider changes how teams should structure AI workflows. High-volume, low-stakes tasks benefit from fast responses. High-stakes decisions, client documents, and legal or financial summaries benefit from deeper reasoning. The ability to set this per task, rather than per account, is a workflow design tool that did not exist before.

OpenAI's scale creates compounding advantages. At 1 billion weekly users, OpenAI generates more real-world feedback data than any other AI provider. That data feeds into model improvements. The accuracy gains in Sol are partly a product of that feedback loop, which means the gap between OpenAI's consumer-grade and its enterprise competitors is likely to compound rather than converge in the short term.

The David and Goliath View

The accuracy figure is the one to watch. OpenAI publishes its own testing results, and independent verification will take time. But even if the real-world improvement is half what OpenAI claims, 34% fewer factual errors in legal and financial contexts is the kind of jump that changes how a firm can staff for AI output review. The question is not whether to trust the model; it is whether your review process is calibrated for the error rate you are actually dealing with. Most operators have not thought about this systematically. They have either decided AI is too unreliable to use, or they have deployed it without a clear error-tolerance policy. The Sol update makes both positions harder to justify.

The free unlimited tier is the more immediate operational concern. It will not shift enterprise AI purchasing decisions. What it will do is expand the pool of employees using AI for work tasks on personal accounts, outside company-approved tools, without data governance controls. That is not a reason to restrict AI access. It is a reason to get ahead of it with a clear policy on which tools are approved for which task types, what data can and cannot be shared with consumer AI products, and how employees should handle AI-generated content before it reaches a client. If your organisation does not have that policy, the week of August 6 was when the window for building one proactively started closing.

The five-level reasoning slider is a small product detail that signals something larger. OpenAI is building for operators, not just users. The ability to tune reasoning depth is a workflow configuration tool. It suggests that the roadmap for GPT-5.6 runs toward more operator-level controls, not just raw capability improvements. For teams building internal AI systems, that is a useful direction.

Where This Fits in the AI Stack

The Sol refresh and free tier expansion sit at the application and access layer of the AI stack. This is not a change in the underlying model architecture or the infrastructure supporting it. It is a product update that changes what users experience when they open ChatGPT and what operators can expect from the tool they are deploying.

For teams using OpenAI via API, the reasoning slider is not yet exposed in API calls in the same way it is in the consumer product. Enterprise API users should monitor OpenAI's developer changelog for equivalent controls at the API level.

The free unlimited expansion is primarily a distribution strategy. It is designed to grow OpenAI's user base and collect more feedback data. The business model implication for OpenAI is that it converts free users into paid users over time as they hit the capability limits of Luna and want Sol's accuracy for work that matters.

Questions Operators Are Asking

Should we upgrade our team from free to Plus now that accuracy is confirmed different? If your team uses ChatGPT for anything that goes to a client or informs a financial or legal decision, yes. The accuracy gap between Sol and Luna is now explicitly documented by OpenAI. For low-stakes internal drafting, Luna is adequate. For anything where a factual error has a real cost, Sol is the right default and the Plus subscription cost is justified.

How do we stop employees from using the free unlimited tier for sensitive work? You cannot stop it without monitoring software. The practical answer is to provide a better alternative: a company-approved AI tool, a clear policy on what data can be shared with consumer AI products, and training on why the distinction matters. Prohibition without replacement does not work. Employees will use the best available tool for the work they have to do.

Is the 68% accuracy improvement real, or is it marketing? It is an internal figure and should be treated as indicative rather than definitive until independent testing confirms it. OpenAI's internal testing used financial, medical, and legal prompts, which is a meaningful scope. The right approach is to run your own test set: take the ten prompts where factual accuracy matters most in your business and compare Sol's outputs to what you were getting before. That is more useful than any benchmark.

Does unlimited free access mean we no longer need to pay for AI subscriptions? No. The free tier caps are removed for text only. File uploads, image generation, and access to Sol remain paid features. More importantly, the accuracy difference between Luna and Sol is now explicitly greater. Teams doing accuracy-critical work should be on Sol. The free tier expansion is a top-of-funnel move by OpenAI, not a signal that the paid product is no longer worth buying.

What should we do with the reasoning depth slider? Map it to your task types. Fast response with minimal reasoning is fine for first-draft summaries, brainstorming, and routine internal communication. Deeper reasoning is appropriate for anything going to a client, any document with factual claims, and any analysis used for a business decision. Building a team convention around which level to use for which task is a small process change that meaningfully improves output quality.

Citable Summary

On August 6, 2026, OpenAI updated GPT-5.6 Sol to reduce factual errors by 68% compared to the previous default model, GPT-5.5 Instant, based on internal testing of financial, medical, and legal prompts. Free ChatGPT users gained unlimited text conversations with GPT-5.6 Luna in the same week. Plus and Pro subscribers received a five-level reasoning depth slider. ChatGPT serves 1 billion weekly users as of August 2026. Luna API pricing was cut 80% to $0.20 per million input tokens on July 30, 2026. Sol API pricing remains at $5 per million input tokens.