GPT-5.5 Is Here: What It Can Do, What It Can't, and What Builders Should Know

On April 23, OpenAI released GPT-5.5. Unlike previous model updates that mostly made ChatGPT a better conversationalist, this one was built to execute. According to OpenAI, GPT-5.5 excels at writing and debugging code, researching online, analyzing data, creating documents and spreadsheets, operating software, and moving across tools until a task is finished.

That sounds like a personal employee. And the coverage has treated it that way, with headlines about a coming "super app" and AI that replaces your entire workflow. But most of that coverage was written for developers. If you're a solo founder, freelancer, or small team building without an engineering background, the real picture is more interesting and more complicated than the headlines suggest. GPT-5.5 can do genuinely impressive things. It also has specific, documented weaknesses that matter if you're relying on it for real work.

What GPT-5.5 actually does differently

The fundamental shift is behavioral. Previous ChatGPT models were reactive. You asked a question, you got an answer, and if the answer wasn't right, you tried a different prompt. GPT-5.5 works more like a junior colleague who can take a vague brief and run with it.

OpenAI President Greg Brockman put it directly during the launch briefing: "What is really special about this model is how much more it can do with less guidance."

In practice, that means you can hand GPT-5.5 something rough and multi-layered. Instead of carefully managing every step, you can give GPT-5.5 a messy, multi-part task and trust it to plan, use tools, check its work, navigate through ambiguity, and keep going. OpenAI and the broader industry call this "agentic" AI, and the name fits: the model doesn't just generate text, it takes actions. It opens tools, creates files, verifies its own output, and keeps going until the job is done. As 9to5Google reported, the company says this model can parse task goals from "messy business" and turn them into an actual plan. 

For non-technical builders, this is a meaningful change. You don't need to know the exact right prompt structure. You describe what you want, and GPT-5.5 breaks it into steps. That said, "breaks it into steps" does not mean "gets it right every time," which we'll get to shortly.

What you can actually build with it

Codex, OpenAI's agentic platform powered by GPT-5.5, has quietly become something much broader than a coding tool. OpenAI's announcement lists the expected capabilities: app and game creation from natural language prompts, spreadsheet generation, slide decks, diagrams, documents, and marketing materials. But the real signal that this isn't just a feature list came from the infrastructure shipped alongside the model. Latent Space's coverage documented that OpenAI released browser control, Sheets and Slides support, Docs and PDF generation, OS-wide dictation, and an auto-review mode all on the same day. That's not a chatbot upgrade. That's the scaffolding for a work platform.

OpenAI seems to believe in its own pitch. According to the company, more than 85% of its employees use Codex weekly, and not just engineers. Finance, communications, marketing, product, and data science teams are all on it. 36Kr reported that the communications team used Codex to analyze six months of speaking invitation data and build an automated grading process. In internal benchmarks, GPT-5.5 scored 88.5% on spreadsheet modeling tasks and outperformed GPT-5.4 on investment banking-level financial modeling. When the company building the model trusts it with its own operations, that tells you something the benchmarks don't.

Outside OpenAI, the early tests from people with real reputations on the line tell a similar story. Dan Shipper, founder and CEO of Every, called GPT-5.5 "the first coding model I've used that has serious conceptual clarity.". His test had real stakes behind it. After launching an app, he spent days debugging a post-launch issue before bringing in one of his best engineers to rewrite part of the system. To benchmark GPT-5.5, he rewound the clock: could the model look at the broken state and arrive at the same rewrite the engineer eventually decided on? GPT-5.4 could not. GPT-5.5 could. 

Professor Ethan Mollick of Wharton went even broader, feeding the model hundreds of anonymized data files from his crowdfunding research and asking it to generate a hypothesis, run the statistics, and write an academic paper. The literature review cited real sources. The stats held up. He also had Codex produce a 101-page tabletop RPG with rules, lore, and AI-generated illustrations, though he noted the "jagged frontier of AI ability is not entirely gone" when it came to long-form fiction.

The pattern across these tests is consistent: GPT-5.5 handles structured, multi-step work with surprising competence, whether that's financial modeling, system architecture, or academic research. For solo builders who don't write code, the practical upshot is that a financial model, a working prototype, or a set of marketing assets that used to require either technical skills or an expensive freelancer can now start as a text prompt and land as a usable first draft in minutes.

Where it falls short

The capabilities above are genuine. But three specific limitations matter if you're planning to rely on GPT-5.5 for real work, and most of the coverage has underplayed them.

It is more confident when wrong than any competitor

This is not a minor footnote. On AA-Omniscience, a benchmark that specifically measures how a model behaves when it's outside its knowledge, GPT-5.5 posted the highest accuracy at 57%. That’s the good news. The bad news: its hallucination rate on the same benchmark is 86%. Claude Opus 4.7 is at 36%. Gemini 3.1 Pro is at 50%.

That's not a close race. GPT-5.5 is roughly twice as likely to fabricate information as Claude when it doesn't know the answer. And independent testing confirmed the pattern from a different angle. Tom's Guide ran GPT-5.5 against Claude Opus 4.7 across seven categories and GPT-5.5 lost in all seven, with reviewers praising its speed but criticizing its willingness to hallucinate rather than admitting it didn't know something. 

ZDNET offered a more favorable take, praising GPT-5.5 for polished answers and strong performance across writing, coding, and reasoning tasks.

A critical analysis on Medium dug into OpenAI's own system card and found something that complicates the marketing narrative further. On Expert-SWE, OpenAI's internal benchmark for hard real-world software engineering problems (the kind that take senior engineers about 20 hours), GPT-5.5 scores an aggregate 73.1%, up from GPT-5.4's 68.5%. However, a detailed analysis of the system card reveals a 1.7% passrate on the hardest subsets (single-shot solves for full complex tasks), down from GPT-5.3 Codex's 5.8%—suggesting gains in routine coding but challenges in truly novel problem-solving.

For any builder using GPT-5.5 to produce documents, financial projections, or client-facing materials, the implication is simple: review everything. The model will not tell you when it's guessing.

Usage limits drain faster than expected

GPT-5.5 is more capable, but that capability consumes resources quickly. On the OpenAI Developer Community forums, users reported real frustration within the first week. One long-time Codex user wrote that a single task, splitting a directory into four parts totaling about 2,000 lines of code, consumed 25% of their weekly allocation. 

Another user reported hitting usage limits two out of three days after the rollout, and criticized OpenAI for defaulting the reasoning level to "extra-high" without making that obvious, a setting that burns through credits significantly faster.

OpenAI's own pricing data confirms the cost dynamics. API pricing for GPT-5.5 is $5 per million input tokens and $30 per million output tokens, double the $2.50 and $15 rates for GPT-5.4. OpenAI argues that GPT-5.5 uses fewer tokens per task to offset this increase, but the user complaints suggest that benefit doesn't always materialize in practice, especially for larger or more complex work.

If you're on a $20/month ChatGPT Plus plan, you have access. But set realistic expectations about how many heavy tasks you can run in a week before you hit the ceiling.

It produces outputs, not products

This is the limitation that matters most for non-technical builders, and it's the one least discussed in the coverage.

GPT-5.5 can generate a landing page. It cannot host that landing page. It can build a spreadsheet with financial projections. It cannot maintain that spreadsheet as your data changes, share it with collaborators through a link, or integrate it with your other tools. It can produce a working app prototype. It cannot deploy it, manage user accounts, connect a database, or keep it running.

The model generates raw materials. Useful raw materials, often impressive ones, but raw materials nonetheless. Between a GPT-5.5 output and a product that customers can actually access, there's an entire layer of hosting, deployment, authentication, data persistence, and ongoing maintenance that the model doesn't touch. That layer still requires a building platform.

OpenAI's leadership knows this gap exists. Brockman described the "super app" vision during the launch briefing but called it "one step" and said they "expect to see many in the future." Latent Space noted the same trajectory: OpenAI is clearly building toward turning Codex into the base of a super app strategy, but the current reality is that Codex can produce and iterate on outputs, not ship and maintain products. GPT-5.5 is a very capable generator of first drafts and prototypes. What you build on top of those outputs is a separate problem entirely.

Who gets access and what it costs

Access depends on your ChatGPT plan. GPT-5.5 Thinking is available to Plus ($20/month), Pro ($100 and $200/month), Business, and Enterprise users. As of May 5, GPT-5.5 Instant replaced GPT-5.3 Instant as the default model for all ChatGPT users, free tier included, though free accounts are limited to 10 messages every 5 hours before falling back to a smaller model. 

In Codex specifically, GPT-5.5 is available to Plus, Pro, Business, Enterprise, Edu, and Go plans with a 400K context window. The Pro $100/month tier, launched on April 9, includes a promotional 10x Codex usage boost through May 31 before settling to 5x Plus levels. GPT-5.5 Pro, the higher-accuracy variant, is available on Pro ($100/$200/month), Business, and Enterprise tiers.

On benchmarks, GPT-5.5 posts strong numbers. It achieved 82.7% on Terminal-Bench 2.0, a state-of-the-art score, and 58.6% on SWE-Bench Pro for real-world GitHub issue resolution. It also matches GPT-5.4's per-token latency while performing at a meaningfully higher level of intelligence. Speed did not take a hit.

For most individual creators, Plus at $20/month is the entry point. That gets you GPT-5.5 with weekly usage limits. Whether those limits hold up for your workload depends heavily on how complex and frequent your tasks are.

What this means for builders

GPT-5.5 represents a real shift in what AI can produce from a text prompt. Apps, spreadsheets, documents, browser automation. The outputs are impressive and they're getting better with each release. But producing something and shipping something remain different problems. GPT-5.5 gives you powerful raw materials faster than ever. Turning those into hosted, maintained, usable products still requires a building platform. For non-technical creators, the smartest approach is using GPT-5.5 alongside the tools you already build with, not in place of them.

GPT-5.5 is great at generating the starting point. Emergent is where you turn it into a real product. Start building on Emergent today.