OpenAI Suspected of Secretly Testing GPT-5.6 as ChatGPT Quality and Reasoning Improve Sharply

OpenAI rolled out the GPT-5.6 family on July 9, 2026, positioning Sol as its flagship reasoning model and Luna as its fastest, most cost-efficient tier. The update matters because ChatGPT serves 1 billion people each week, giving even small changes in accuracy, consistency and access broad consumer impact. Extending a current-generation model to the $0 Free plan also sharpens competition among AI providers over both performance and the cost of everyday use.

On Aug. 6, OpenAI said an updated GPT-5.6 Sol would immediately power both Instant replies and deeper reasoning for Plus and Pro subscribers, alongside a slider controlling reasoning effort. In an internal test of finance, medical and legal prompts, answers with at least one factual error were 68% less common for Sol and 62% less common for Luna than for GPT-5.5 Instant. Free and Go users switch to Luna by default this week and receive unlimited text chats and a Think button next week, while uploads, images and other tools remain capped.

OpenAI has divided GPT-5.6 into three capability tiers that can evolve independently: flagship model Sol; Terra, designed to balance everyday performance and cost; and Luna, focused on speed and affordability. The structure lets businesses allocate spending according to task complexity. Sol strengthens long-horizon agentic work, software development and cybersecurity research, and carries tighter safeguards because of potential misuse risks.

At the U.S. government’s request, OpenAI gave about 20 approved partners a limited preview on June 26, 2026, before releasing the models on ChatGPT, Codex and the API on July 9. Input/output prices per million tokens are $5/$30 for Sol, $2.50/$15 for Terra and $1/$6 for Luna. Sol scored 92.2% on BrowseComp and 62.6% on OSWorld 2.0, both record highs.

OpenAI launched GPT-5.5 on April 23, 2026, and expanded Codex's role beyond software development to research, data analysis, document and spreadsheet creation, and software operation. The shift means AI agents can do more than assist with coding: they can complete longer, general-purpose computer workflows across multiple tools. The model became available through the API on April 24.

The UK AI Security Institute published an evaluation on April 30, 2026, finding that GPT-5.5 performed comparably to Anthropic's Claude Mythos Preview across 95 cybersecurity tasks spanning four difficulty levels. It also became the second model to complete an end-to-end simulated enterprise-network attack, a test estimated to take a human 20 hours. API pricing is $5 per million input tokens and $30 per million output tokens.

OpenAI continues to position its Instant series at the heart of everyday conversations on ChatGPT, emphasizing speed, concise answers and low latency. Generative AI can fabricate information in high-risk areas such as healthcare, law and finance, making fewer hallucinations and better contextual understanding critical to strengthening trust and user retention.

As of July 19, 2026, OpenAI had rolled out GPT-5.5 Instant to all users, replacing GPT-5.3 as ChatGPT’s default model. Company data showed that the new model reduced hallucinations in high-risk scenarios by 52.5%. It also improved personalized memory and contextual understanding, producing more concise responses that better match users’ needs.

GPT-5.3 Instant is OpenAI’s fast model for everyday ChatGPT conversations, succeeding GPT-5.2 Instant. Rather than extending reasoning, it focuses on improvements users encounter most directly: tone, relevance and conversational flow. The update also addresses excessive refusals, lengthy warnings and insufficient context in web-search responses, all of which affect reliability across ChatGPT’s most common use cases.

OpenAI rolled out GPT-5.3 Instant to all ChatGPT users on March 3, 2026, under the API name gpt-5.3-chat-latest. It did not disclose a separate price for the update. Internal evaluations showed hallucination rates in high-stakes domains fell 26.8% with web access and 19.7% without it, while errors reported by users declined 22.5% and 9.6%, respectively.