SAN FRANCISCO, 5 MAY 2026 — OpenAI has replaced GPT-5.3 Instant with GPT-5.5 Instant as the default model behind ChatGPT, less than two weeks after GPT-5.5 reached paid subscribers on 23 April. The change reaches every tier, including the free version; OpenAI's chief executive said in October 2025 that more than 800 million people use ChatGPT every week.
Key Takeaways
- GPT-5.5 Instant is now the model that answers when any ChatGPT user — free or paid — sends a message without selecting an alternative.
- The model is also available in the API under the alias
chat-latest, which dynamically points to whatever model OpenAI currently considers the production default. - OpenAI says GPT-5.5 Instant reduces hallucination in sensitive areas such as law, medicine and finance, and reports gains on the AIME 2025 and MMMU-Pro benchmarks.
- Paid users can still choose GPT-5.3 Instant, but only for three months.
- Plus and Pro users on the web gain personalisation drawn from past chats, files and connected Gmail; OpenAI says Free, Go, Business and Enterprise users will follow in the coming weeks.
The Facts
The 5 May change follows the 23 April 2026 launch, when GPT-5.5 first arrived for ChatGPT Plus, Pro, Business, and Enterprise users, alongside the heavier-weight GPT-5.5 Pro variant aimed at long-reasoning tasks. The 5 May milestone is narrower but reaches more people: GPT-5.5 Instant becomes the default model — meaning the model a user gets without making any selection — for the entire ChatGPT user base, including the free tier.
OpenAI's announcement post accompanies a launch in which OpenAI said the model reduces hallucination in sensitive areas such as law, medicine and finance while keeping the low latency of its predecessor. The company published a small set of benchmarks. On the AIME 2025 math test, GPT-5.5 Instant scored 81.2, against 65.4 for GPT-5.3 Instant; on the MMMU-Pro multimodal reasoning benchmark, it scored 76 against 69.2.
The model is the new operational default in two places. First, in the consumer ChatGPT app — web, iOS, and Android — any user message arriving without an explicit model selection is now served by 5.5 Instant. Second, in the API, GPT-5.5 Instant is available under the alias chat-latest.
OpenAI also expanded the model's personalisation features. As TechCrunch reported, Plus and Pro subscribers on the web now see ChatGPT drawing on past conversation history, uploaded files, and — where enabled — content from a connected Gmail account, to ground responses in user-specific context. OpenAI said it plans to extend the feature to Free, Go, Business and Enterprise users in the coming weeks.
The free tier change matters at scale. OpenAI last put ChatGPT's weekly users at more than 800 million, in October 2025. Until 5 May, a free-tier user who did not pick a model received responses from GPT-5.3 Instant. The new default sets a higher floor on what an unauthenticated, unpaid user gets — meaningful for adoption metrics and meaningful for the threat model when ChatGPT is used in adversarial workflows.
Technical Deep-Dive
GPT-5.5 Instant is built on the unified GPT-5 architecture, which merged what were previously two separate model families: the "GPT" line optimised for fluency and the "o-series" line optimised for explicit reasoning. The unified architecture launched with GPT-5 in mid-2025 and has been iterated through several point releases since. The "Instant" suffix denotes the variant optimised for low-latency single-turn and multi-turn chat without extended thinking; "Pro" denotes the variant that runs longer internal reasoning chains before responding.
OpenAI has claimed improvements in areas such as coding and knowledge work, but in the material we reviewed it did not publish detailed figures on tool-call accuracy or long-context behaviour. Teams that depend on either should measure them on their own workloads.
For developers, the most consequential change is the chat-latest alias. Pinning code to an explicit model ID gives stability — your code keeps using the model you chose until you change it. Pinning to chat-latest gives you whatever OpenAI currently serves as the ChatGPT default. This is OpenAI's clearest move toward managing the rollout pace itself, similar to how cloud platforms manage their LTS versus current channels. Production teams should think carefully about which alias matches their risk appetite.
ASEAN Perspective
Southeast Asia is one of the regions where the free-tier default change matters most. ChatGPT's user base in the region skews heavily toward free-tier users — paid subscriptions are likely to be less common than in North America or Western Europe, given differences in purchasing power. When OpenAI improves what the free user gets, the lift hits ASEAN proportionally harder than it hits the US.
In the Philippines, the BPO sector has absorbed generative AI tools into routine workflows — from drafting English-language correspondence for back-office accounting clients to summarising long compliance documents for offshore legal teams. The shift to GPT-5.5 Instant gives Filipino contact-centre and BPO workers who use ChatGPT a stronger default tool for English-second-language writing and reading tasks.
In Indonesia, OpenAI has not published Bahasa Indonesia benchmarks for 5.5 Instant. Jakarta-based startups building consumer chatbots on the OpenAI API should re-test their evaluation suites against 5.5 Instant rather than rely on 5.3 Instant numbers.
In Singapore, the Monetary Authority of Singapore (MAS) and the Cyber Security Agency have both issued AI governance and risk guidance over the past 18 months — including the CSA Advisory on Risks associated with Frontier AI Models in April 2026. Singapore enterprise buyers will want to validate that the new model's behaviour aligns with their existing model-risk-management documentation. In particular, once the personalisation feature drawing on connected Gmail data reaches business tiers, it will need careful review under PDPA cross-border transfer rules.
Vietnam and Thailand have smaller but growing developer communities building OpenAI-API-backed products. For these teams, the chat-latest alias creates a hidden risk: a model swap that arrives without notice can shift quality, latency, and cost characteristics. For production traffic, pinning to an explicit model ID is the more defensive posture.
What Organisations Should Do
A model swap of this scale at the default tier is the kind of change that should trigger a re-evaluation cycle, not a sigh of relief. Five actions to take in the next 30 days:
-
Re-run your internal LLM evaluation suite against GPT-5.5 Instant. If your suite includes a set of representative prompts from your production workflow, run them through 5.5 Instant and 5.3 Instant side-by-side. Document any regressions. Note that "regressions" includes things like response style or formatting changes that downstream parsers depend on.
-
Decide your alias policy. Pick deliberately between an explicit model ID (stability) and
chat-latest(always-current). For production traffic, almost always pick the explicit ID. For internal tools where you want to track OpenAI's quality improvements automatically, the alias is fine. -
Audit the personalisation surface. If your organisation has connected Gmail or file uploads as part of an enterprise ChatGPT deployment, understand what personalisation data is being drawn into prompts. For some regulated industries, this changes the data classification of every prompt and may require a compliance review.
-
Re-test prompt injection defences. Each new model generation can change how well known prompt-injection patterns evade or trigger guardrails. Run your existing red-team prompt corpus against 5.5 Instant.
-
Communicate the change to internal users. If your organisation has ChatGPT in any sanctioned form, send a one-paragraph note to users explaining that the underlying model changed, that response style may shift, and that prompt patterns they had memorised may need adjustment. Surface change is high.
RECATOOLS Verdict
We think the model swap matters more than the model itself. GPT-5.5 Instant is a modest, real improvement over 5.3 Instant — not a leap. But the operational decision to move 800 million weekly users to a new default in a single rollout is a milestone for OpenAI: the company is now treating its consumer model channel like cloud providers treat their compute fleets, with rolling defaults, named aliases, and behind-the-scenes capacity management. That maturity is what enterprises have been asking for.
The risk is that "always upgrading" silently rewrites the behaviour of every product built on the API. We believe the responsibility for managing that risk rests with the developer using the API, not with OpenAI. Pin your model IDs in production. Run your evals on every release. Treat chat-latest as a development convenience, not a production target. The pattern is the same one cloud-native teams already follow for managed databases and Kubernetes versions.
For ASEAN developers building consumer-facing AI products on top of the OpenAI API, our view is: this is the moment to look hard at the cost equation. GPT-5.5 Instant is a strong English writer, but lower-priced models from other providers deserve a benchmark against your own workloads before you commit volume to any one of them.
Frequently Asked Questions
Will my existing prompts still work with GPT-5.5 Instant? Probably, but test rather than assume. A new default model can change response length, formatting and tone, which matters if downstream code parses its output. Production teams should re-test rather than assume zero-diff.
How do I lock my API calls to GPT-5.5 Instant explicitly?
Specify an explicit model ID in the model parameter of your API request. Avoid chat-latest for production workloads — that alias follows whatever OpenAI currently serves as the ChatGPT default, and changes when that default does. Plan any rollback window with care: TechCrunch reports that GPT-5.3 will remain available to paid users for only three months.
Correction, 18 October 2026: An earlier version included benchmark, pricing, Codex, Azure availability, tool-use and model-retirement details that we could not confirm, and described feedback from developer teams that we had not interviewed; these have been removed, and the benchmark figures now match those OpenAI published. It also referred to a Google model released after this article's date and misstated which ChatGPT users will receive the new personalisation features.
View author profile → · Editorial policy
