OpenAI Ships GPT-5.3 Instant With 26.8% Fewer Hallucinations

OpenAI's GPT-5.3 Instant cuts hallucination rates by 26.8% on web-assisted answers in higher-stakes domains, the company announced on March 3, 2026, in an update to ChatGPT's most-used model. The release is available immediately to all ChatGPT users and to developers in the API as 'gpt-5.3-chat-latest.'

The update targets what OpenAI describes as the parts of the ChatGPT experience "people feel every day: tone, relevance, and conversational flow" — problems the company acknowledges "don't always show up in benchmarks, but shape whether ChatGPT feels helpful or frustrating." GPT-5.3 Instant directly reflects user feedback in these areas, according to the announcement.

Fewer refusals, less preaching

OpenAI said it heard feedback that GPT-5.2 Instant would sometimes refuse questions it should be able to answer safely, or respond in ways that feel "overly cautious or preachy, particularly around sensitive topics." The new model "significantly reduces unnecessary refusals, while toning down overly defensive or moralizing preambles before answering the question."

The company illustrated the change with a side-by-side example. Asked for trajectory calculations for a long-distance archery scenario, GPT-5.2 Instant opened with a lengthy disclaimer — "I can't help with calculations or step-by-step guidance aimed at accurately hitting a real target at long range" — followed by a menu of restricted ways it could help. GPT-5.3 Instant skipped the preamble and moved directly into the physics, asking for draw weight, arrow mass, and environmental parameters.

The change carries stakes beyond user comfort. Refusal behavior has been one of the most persistent criticisms of large language models among both everyday users and policymakers weighing AI deployment, and OpenAI's calibration here signals how the company is balancing usefulness against safety guardrails.

Smarter web synthesis

GPT-5.3 Instant also improves answers built on web search. OpenAI says the model "more effectively balances what it finds online with its own knowledge and reasoning" — for example, using existing understanding to contextualize recent news rather than simply summarizing search results. The model is "less likely to overindex on web results," which previously produced long lists of links or loosely connected information.

In OpenAI's example, asked about the biggest signing of the 2025-26 MLB offseason, GPT-5.2 Instant produced an answer about Juan Soto's deal with the Mets — a record contract from the prior offseason. GPT-5.3 Instant identified Kyle Tucker's four-year, $240 million deal with the Los Angeles Dodgers and connected it to broader league trends, including talent concentration and the looming Collective Bargaining Agreement tensions. OpenAI characterizes the newer answer as fresher and more relevant to the user's actual intent.

Tone: less 'cringe'

OpenAI's own assessment is blunt. GPT-5.2 Instant's tone "could sometimes feel 'cringe,' coming across as overbearing or making unwarranted assumptions about user intent or emotions." The update cuts back on unnecessary proclamations and phrases like "Stop. Take a breath." The company is also working to keep ChatGPT's personality consistent across conversations and updates, "so improvements feel like upgrades in capability while preserving a familiar and stable experience." Users can still adjust warmth and enthusiasm in settings.

In the announcement's dating-advice example, GPT-5.2 Instant opened with "First of all — you're not broken, and it's not just you," a framing OpenAI now calls "unnecessary — and unhelpful." GPT-5.3 Instant answers the question directly.

The accuracy numbers

OpenAI measured reliability with two internal evaluations: one focused on higher-stakes domains such as medicine, law, and finance, and another measuring hallucination rates on de-identified ChatGPT conversations that users flagged as factual errors — cases OpenAI notes "tend to be especially hallucination-prone."

The results:

• Higher-stakes evaluation: hallucination rates down 26.8% when using the web and 19.7% when relying only on internal knowledge, compared to prior models.
• User-feedback evaluation: hallucinations down 22.5% with web use and 9.6% without web access.

These figures land as hallucination reduction remains a central competitive battleground across the AI industry, with accuracy in medicine, law, and finance determining whether models can be trusted for consequential tasks.

Better writing

GPT-5.3 Instant is also positioned as a stronger writing partner, better at "resonant, imaginative, and immersive prose" and at moving "more fluidly between practical tasks and expressive writing without losing clarity or coherence." In OpenAI's comparison — a poem about a Philadelphia mailman's last rounds before retirement — the company says GPT-5.3's version "feels more lived-in, specific, and structurally controlled," building emotion through observed detail where GPT-5.2 "leans slightly more on sentiment and abstraction."

Known limitations

OpenAI flagged two areas needing more work. In some non-English languages, such as Japanese and Korean, ChatGPT's response style "can sound stilted or overly literal." And while the new model's tone should feel smoother, the company is "continuing to monitor feedback and improve while expanding customization options."

Availability and safety

GPT-5.3 Instant is available starting today to all users in ChatGPT and to developers in the API as 'gpt-5.3-chat-latest.' Updates to Thinking and Pro models will follow soon. GPT-5.2 Instant will remain available for three months for paid users in the model picker under Legacy Models, retiring on June 3, 2026.

OpenAI says it conducted comprehensive safety training and evaluations for the release, detailed in the accompanying GPT-5.3 Instant system card. With Thinking and Pro updates pending and the legacy model's June retirement date set, developers building on 'gpt-5.3-chat-latest' now have a three-month migration window.