OpenAI Unveils GPT-5.5: A More Agentic Model for Coding, Research, and Real Work

OpenAI Unveils GPT-5.5: A More Agentic Model for Coding, Research, and Real Work

OpenAI released GPT-5.5 on April 23, 2026, positioning it as its strongest model yet for complex professional work: coding, research, document analysis, data work, and agent-style use of tools. The headline is not just “better chatbot,” but a model built to carry more tasks through to completion with less hand-holding.

The News in Brief

GPT-5.5 is now available across ChatGPT, Codex, and the API, with OpenAI describing it as a frontier model for “complex professional work.” In its launch materials, the company says GPT-5.5 performs especially well on coding, research, data analysis, information synthesis, and document-heavy tasks, particularly when tools and plugins are involved.

The key benchmark claims are eye-catching. OpenAI says GPT-5.5 scores 84.9% on GDPval, a benchmark for well-specified knowledge work across 44 occupations; 78.7% on OSWorld-Verified, which tests the ability to operate real computer environments; and 98.0% on Tau2-bench Telecom, a customer-service workflow benchmark run without prompt tuning. For coding, OpenAI reports 82.7% on Terminal-Bench 2.0 and 58.6% on SWE-Bench Pro.

The model is also commercially significant: OpenAI is making GPT-5.5 available through Amazon Bedrock, bringing the model into AWS enterprise environments.

What Was Actually Announced

OpenAI announced GPT-5.5 as a model designed for execution-heavy knowledge work, not merely conversation. The company’s framing is clear: GPT-5.5 is supposed to understand goals earlier, use tools more effectively, check its own work, and keep going until a task is done. That is a different emphasis from earlier chatbot launches, where the focus was often on fluency, general reasoning, or multimodal interaction.

What appears to be available now: GPT-5.5 is listed in OpenAI’s API documentation as a frontier model for complex professional work, with text and image input and text output. The API page also lists a 1,050,000-token context window and pricing shown as $5 input / $30 output, though OpenAI’s pricing pages should be checked before quoting this commercially because model pricing can change.

OpenAI also announced GPT-5.5 in relation to Codex, describing it as a stronger agentic coding model. The company says it performs better on command-line workflows, real GitHub issue resolution, and tool-heavy development tasks.

What is less clear is the underlying model architecture. OpenAI did not publicly disclose parameter count, training data composition, full architecture details, or compute budget. That means any claims about “how big” GPT-5.5 is, or exactly how it differs internally from GPT-5.4, should be treated cautiously unless they come from OpenAI or reproducible third-party analysis.

The announcement also sits inside a broader enterprise push. A few days later, OpenAI said GPT-5.5 would be available through Amazon Bedrock, meaning companies can use OpenAI models inside AWS infrastructure and procurement systems.

The Technical Angle

The most important technical shift is not that GPT-5.5 can write more polished paragraphs. It is that OpenAI is presenting it as a better agentic work model: a model that can plan, use tools, operate across software environments, check intermediate results, and complete longer workflows.

The benchmark selection tells us a lot about the intended use case. OSWorld-Verified is about controlling real computer environments. Terminal-Bench 2.0 tests command-line workflows that require planning, iteration, and tool coordination. SWE-Bench Pro tests real-world GitHub issue resolution. These are not just language benchmarks; they are closer to measuring whether a model can do work inside a digital environment.

From a systems perspective, this reflects the industry’s move from “single-turn answer generation” to tool-using AI systems. A model like GPT-5.5 is valuable if it can maintain context over long tasks, choose tools appropriately, recover from errors, and reduce the number of human interventions required. OpenAI’s system card explicitly says the model is designed for writing code, researching online, analysing information, creating documents and spreadsheets, and moving across tools.

The long context window is also technically important. OpenAI’s API documentation lists GPT-5.5 with a 1,050,000-token context window, which matters for codebases, legal documents, large research packets, financial models, and enterprise knowledge work.

However, long context is not the same thing as reliable reasoning. A model can ingest a huge amount of text and still miss details, over-weight irrelevant passages, or fail to maintain a coherent plan. The real technical question is whether GPT-5.5 improves not just context capacity, but context discipline: knowing what matters, checking it, and using it correctly.

On architecture, OpenAI has not disclosed enough to say whether GPT-5.5 is a larger dense model, a mixture-of-experts system, a more heavily post-trained model, or some combination of architectural and training improvements. The cautious interpretation is that GPT-5.5 is best understood by its behavioural profile: better tool use, better long-task persistence, stronger coding, and more efficient task completion.

Why It Matters

GPT-5.5 matters because it shows where the frontier model race is moving: away from chatbots as answer machines and toward AI systems that perform work.

For developers, the obvious impact is coding. If GPT-5.5 genuinely improves command-line workflows, debugging, issue resolution, and multi-step repository work, then the role of a software engineer shifts further toward supervising, reviewing, and orchestrating AI-generated changes.

For enterprises, the significance is workflow automation. GPT-5.5 is being marketed for research, analysis, documents, spreadsheets, coding, and tool use — the messy middle of knowledge work. That is where many white-collar processes live: not fully automatable with old software, but repetitive enough for AI assistance.

For the AI industry, the launch reinforces a competitive direction: models are being judged less by conversation quality alone and more by whether they can complete real tasks in real environments. Benchmarks like OSWorld, SWE-Bench, and Terminal-Bench are imperfect, but they are closer to practical work than traditional question-answering tests.

Is it genuinely new ground? Partly. The idea of tool-using agents is not new. What is new is the degree to which frontier labs are now packaging agentic capability as the main product story.

The Reaction

The early response has been mixed but engaged. The Verge summarised the launch around efficiency, coding, and tool-heavy work, noting that OpenAI is emphasising GPT-5.5’s ability to write and debug code and operate across tools.

There has also been notable attention from the safety and cybersecurity community. The UK AI Security Institute evaluated GPT-5.5’s cyber capabilities and described it as one of the strongest models it has tested on cyber tasks. AISI said GPT-5.5 was the second model to solve one of its multi-step cyber-attack simulations end-to-end.

That is both impressive and uncomfortable. Stronger agentic capabilities help defenders, developers, and analysts — but they may also increase misuse potential. The same ability to reason through complex workflows can apply to benign software engineering or harmful cyber operations.

There was also a lighter but revealing reaction around GPT-5.5 and Codex producing odd “goblin” or mythical-creature references. OpenAI reportedly traced this to training incentives and personality-related behaviour, which became a small viral moment but also highlighted how subtle reward signals can shape model behaviour in unexpected ways.

The Caveats and Open Questions

The biggest caveat is that benchmarks are not reality. A model can score well on Terminal-Bench, SWE-Bench, OSWorld, or GDPval and still fail in messy production settings. Real workplaces involve unclear instructions, broken tools, missing permissions, contradictory documents, legacy systems, and human politics. GPT-5.5 may be better at real work, but benchmark results do not prove it will be reliable in every business process.

There are also major unknowns. OpenAI has not disclosed GPT-5.5’s parameter count, architecture, training data, full post-training process, or compute requirements. Without those details, outside observers cannot fully assess what changed technically.

Safety is another open issue. AISI’s cyber evaluation is important because it suggests GPT-5.5 is powerful enough to matter in offensive and defensive cybersecurity contexts. A model that can complete more complex workflows may require stronger access controls, monitoring, and deployment policies.

There is also the question of reliability. OpenAI says GPT-5.5 uses tools more effectively and checks its work, but businesses should still assume human review is necessary for high-stakes tasks: legal advice, medical decisions, financial modelling, security operations, and production code changes.

Finally, the marketing language around “real work” deserves scrutiny. AI systems can accelerate work, but they do not automatically understand business context, accountability, risk tolerance, or organisational judgement. The practical value will depend on integration, governance, and human oversight.

What Comes Next

The next thing to watch is not just whether GPT-5.5 is smarter, but whether it becomes operationally dependable. The real test will be long-running coding agents, enterprise workflows, cybersecurity defence, research assistance, and document-heavy analysis where mistakes are costly.

OpenAI’s AWS partnership suggests the company wants GPT-5.5 embedded inside enterprise infrastructure rather than used only through ChatGPT.

Expect competitors to respond less with “our chatbot is nicer” and more with claims around agents, tool use, code execution, enterprise integration, safety controls, and cost-per-task. The frontier race is becoming a race to build models that do work — and to prove they can be trusted while doing it.

---

Transformer AI helps SMEs navigate the AI landscape without the jargon. If you would like a frank conversation about where models like GPT-5.5 could have an impact in your business, get in touch.