OpenAI Lifts Chat Limits as GPT-5.6 Luna Rolls Out to a Billion Weekly Users

A New Baseline for 1 Billion Weekly Users

OpenAI has eliminated text-based chat limits for all ChatGPT users, a move that arrives as the platform surpasses 1 billion weekly active users. The company is rolling out GPT-5.6 Luna as the new default model for free and Go-tier accounts, replacing the earlier GPT-5.5. Free users will retain caps on file uploads, image processing, voice interactions, and image generation, but the core text conversation experience is now unconstrained.

The shift marks a strategic inflection point for OpenAI's monetization model. At DailyTechWire, we've tracked how AI incumbents have navigated the tension between growth and unit economics - Anthropic's Claude recently tightened rate limits for free users, while Google's Gemini has experimented with quota-based access. OpenAI's decision to remove text caps suggests the company believes its subscription funnel is strong enough to absorb the incremental compute cost, or that the marginal expense of serving text inference has fallen sufficiently to make unrestricted access viable as a user-acquisition lever.

Reasoning on Demand

Free and Go users will gain access to a new "Think" button, designed to allocate additional reasoning cycles to complex queries. The feature is OpenAI's first attempt to surface inference-time compute controls to non-paying users, albeit in a binary form. The Think button effectively lets the model spend more tokens on chain-of-thought processing before delivering an answer, a technique that has become central to recent performance gains in large language models.

For Plus and Pro subscribers, the controls are more granular. They receive access to GPT-5.6 Sol, a variant optimized for rapid tasks - web research, decision support, short-form writing, and planning. OpenAI describes Sol as delivering "more compact and robust answers," a framing that suggests the model has been fine-tuned for lower latency and higher consistency at the cost of some reasoning depth. Importantly, this consumer-facing Sol is distinct from the version deployed in Codex and Work products, which remain unchanged.

Plus and Pro users also gain a thinking slider, a continuous control that adjusts how much inference-time compute the model applies to a given prompt. The slider is calibrated to the complexity and step-count of the query, allowing users to trade response time for accuracy. This is the first time OpenAI has exposed such a parameter directly in the consumer interface, and it reflects a broader industry trend toward user-configurable inference profiles. Anthropic's Claude and Google's Gemini have both experimented with similar controls in enterprise contexts, but OpenAI is the first to bring them to a mass-market chatbot.

Error Rates Drop by Two-Thirds

OpenAI shared internal evaluation data showing that GPT-5.6 Luna produces 62% fewer factual errors than GPT-5.5-Instant, while GPT-5.6 Sol reduces errors by 68%. The company did not disclose the benchmark suite or error taxonomy used, but the magnitude of improvement is consistent with gains we've observed in other labs' reasoning-focused models over the past six months.

Factual accuracy has been a persistent friction point for consumer AI products. A study published in March by researchers at Seoul National University found that GPT-5.5 hallucinated in roughly 18% of fact-retrieval queries, a rate that made it unsuitable for high-stakes use cases. If OpenAI's internal numbers hold in the wild, GPT-5.6 Luna would bring that rate below 7%, a threshold that several enterprise customers have cited as the minimum for production deployment.

Still, the error-rate metric is narrow. It measures factual recall but does not capture reasoning coherence, instruction-following, or creative quality - dimensions where model performance is harder to quantify and where user preferences vary widely. The real test will be whether the accuracy gains translate into measurable retention and engagement lifts, particularly among free users who are most sensitive to quality degradation.

Rollout Timeline and Tiering

The upgraded GPT-5.6 Sol is available to Plus and Pro subscribers immediately. Free and Go users will see the other changes - Luna as default, the Think button - roll out over the next week. Unlimited text chats for free accounts will arrive the following week, according to OpenAI.

The staggered rollout is likely a hedge against infrastructure load. OpenAI has historically struggled with capacity management during major launches; GPT-4's release in March 2023 triggered multi-day slowdowns, and the company has since adopted a phased deployment strategy to smooth demand spikes. By giving paying users early access and delaying the most resource-intensive change - unlimited free text - OpenAI buys itself time to monitor cluster utilization and adjust inference scheduling before the full user base hammers the system.

The Economics of Unlimited

Removing chat limits for free users is a gamble that rests on two assumptions: that text inference costs have fallen far enough to make the move sustainable, and that the marginal user acquired through unlimited access is valuable enough to justify the compute spend. OpenAI has not disclosed its cost per token, but industry estimates place GPT-5.5 inference at roughly $0.40 per million output tokens at scale. If GPT-5.6 Luna is comparably efficient, the incremental cost of serving a typical free user - assuming 50 messages per week at an average of 300 output tokens per message - would be around six cents per user per week, or roughly $3 per user per year.

That's a defensible customer-acquisition cost if even a small fraction of free users convert to paid tiers. ChatGPT Plus is priced at $20 per month, and Pro at $200 per month. If 5% of free users upgrade within a year, the economics pencil out. But the calculus becomes more precarious if inference costs rise - due to model complexity, regulatory compliance overhead, or energy prices - or if conversion rates disappoint.

There's also a competitive dimension. Anthropic, Google, and several well-funded Chinese labs are all vying for consumer AI mindshare, and free-tier generosity is one of the few levers that can shift market share at scale. By offering unlimited text, OpenAI makes it harder for rivals to differentiate on access, forcing them to compete on model quality, ecosystem integrations, or vertical-specific features - all of which play to OpenAI's strengths.

What the Slider Reveals

The thinking slider is the most interesting design choice in this release. By externalizing a parameter that was previously hidden inside the model's inference stack, OpenAI is acknowledging that one-size-fits-all latency-accuracy tradeoffs don't work for all users or all queries. A user drafting a quick email wants speed; a user debugging a complex SQL query wants thoroughness. The slider lets them express that preference directly.

But it also shifts cognitive load onto the user. Deciding how much "thought" a model should apply requires some mental model of what the slider actually controls - token budget, search depth, sampling temperature - and most users won't have that model. If the slider is poorly calibrated or its effects are opaque, it risks becoming a source of confusion rather than empowerment.

We've seen similar experiments fail in other contexts. Microsoft's Bing Chat briefly exposed a "creative," "balanced," and "precise" mode selector in 2023, but user research showed that most people either ignored it or felt anxious about choosing the wrong mode. The slider may fare better if OpenAI invests in onboarding and contextual hints, but the early design suggests the company is optimizing for power users rather than the median free-tier account.

Implications for the Broader AI Stack

OpenAI's move will likely accelerate the commoditization of text inference. If the market leader is willing to give away unlimited chat, smaller players will face pressure to match or beat that offer, compressing margins across the industry. That's good news for developers building on top of chat interfaces - lower costs and fewer rate limits make it easier to experiment - but it's challenging for model providers who lack OpenAI's scale or capital cushion.

The error-rate improvements also raise the bar for what counts as a competitive reasoning model. A 60-plus percent reduction in factual errors is a step-function change, and it will reset user expectations for accuracy. Models that were considered state-of-the-art six months ago may now feel unreliable by comparison, forcing other labs to accelerate their own fine-tuning and evaluation cycles.

Finally, the introduction of user-facing reasoning controls hints at a future where inference becomes more configurable and context-aware. We're likely to see more products expose parameters like temperature, top-k sampling, and chain-of-thought depth, and eventually those controls may be automated - models that adaptively allocate compute based on query characteristics, without requiring manual tuning. OpenAI's slider is a small step in that direction, but it's a step that could reshape how we think about model interfaces over the next 18 months.

The Path Ahead

OpenAI's latest release is as much about market positioning as it is about model capabilities. By lifting chat limits and introducing adaptive reasoning controls, the company is betting that it can sustain user growth without sacrificing margin structure, and that the value of platform lock-in outweighs the cost of serving free users at scale. Whether that bet pays off will depend on conversion rates, infrastructure efficiency, and how aggressively competitors respond.

For now, the message is clear: text inference is no longer a scarce resource, and the competitive battleground is shifting to accuracy, latency, and user experience. The labs that can deliver all three - at scale, across tiers - will define the next phase of the consumer AI market.