OpenAI has introduced GPT-5.5, marking its first entirely retrained base model since GPT-4.5.

      The model, referred to as “Spud,” is engineered to handle complex multi-step tasks with minimal human guidance. It establishes new standards in agentic coding, computer usage, and knowledge work, while achieving comparable per-token latency to GPT-5.4. API access is currently postponed as further safety measures are being developed.

      For months, it has been an open secret in the AI industry that Anthropic’s Claude has been leading the enterprise market. Since at least December 2025, OpenAI has been in a “Code Red” situation, observing Anthropic’s annual recurring revenue soar from $9 billion to $30 billion, while witnessing a decline in its own B2B positioning.

      On Thursday, OpenAI made a move in response. GPT-5.5, its first fully retrained base model since GPT-4.5, is now being deployed to Plus, Pro, Business, and Enterprise users via ChatGPT and Codex. This model is built to perform tasks with limited human input, functioning across platforms like email, spreadsheets, calendars, and more.

      The main concept behind GPT-5.5 is legibility. While earlier models needed well-structured prompts and extensive supervision, OpenAI asserts that 5.5 can tackle “messy, multi-part tasks” by independently planning, utilizing tools, verifying its output, navigating uncertainties, and continuing until completion.

      The improvements are concentrated in four key areas: agentic coding, computer usage, knowledge work, and preliminary scientific research. OpenAI categorizes these as fields “where advancement hinges on reasoning across contexts and taking action over time.”

      The benchmark results are impressive. GPT-5.5 achieves 82.7% on Terminal-Bench 2.0, which assesses complex command-line workflows that necessitate planning, iteration, and tool coordination. On SWE-Bench Pro, designed to evaluate real-world issue resolution on GitHub across four programming languages, it scores 58.6%, successfully solving more tasks in one attempt compared to prior models.

      On GDPval, which assesses agents in 44 knowledge work professions, it scores 84.9%. For OSWorld-Verified, which checks if a model can autonomously operate real computer environments, it reaches 78.7%. On Tau2-bench Telecom, it scores 98.0% without needing prompt tuning. Overall, OpenAI indicates that GPT-5.5 surpasses the scores of GPT-5.4 while utilizing fewer tokens.

      This efficiency claim holds significant commercial value. Usually, larger, more capable models are slower, creating a cost-quality dilemma for enterprise clients. OpenAI states that GPT-5.5 matches the per-token latency of GPT-5.4 in practical use, suggesting an improvement in intelligence without a delay in response time. Moreover, it requires considerably fewer tokens to execute similar tasks in Codex, directly lowering the cost per task for enterprise implementations. Although GPT-5.5 is priced higher per token than GPT-5.4, OpenAI asserts that the overall effect is superior outcomes at a lower total cost for most workflows.

      The safety approach is notably more cautious than in past launches. OpenAI conducted evaluations of GPT-5.5 within its “entire suite of safety and preparedness frameworks,” collaborated with both internal and external red team members, included specific testing for advanced cybersecurity and biology functionalities, and gathered feedback from nearly 200 trusted early-access partners prior to the launch.

      The focus on cybersecurity reveals heightened caution: OpenAI mentions introducing “stricter classifiers for potential cyber risks that some users may find initially bothersome.” The company recognizes that GPT-5.5 signifies a substantial advancement in cybersecurity capabilities, framing the enhanced protections as a necessary investment for responsible use.

      Notably, the API is missing from this launch. GPT-5.5 is currently available in ChatGPT and Codex for paid subscribers, but OpenAI states that API launches “call for different safeguards, and we are working closely with partners and customers on the safety and security requirements for large-scale implementation.” The company promises that API access will come “very soon,” though no specific date has been provided. This poses a significant delay for enterprise customers who rely on the API instead of the ChatGPT interface. Additionally, GPT-5.5 Pro, a version featuring extended reasoning, is exclusively available to Pro, Business, and Enterprise subscribers.

      The competitive landscape heavily influences every design choice. OpenAI is developing its unified desktop “super-app” around GPT-5.5, intending to integrate ChatGPT, Codex, and the Atlas browser agent within one session. This model is intended to facilitate intent-aware reasoning in that integrated workspace, a product category that has only emerged in the last six months. GPT-5.2 Thinking will continue to be offered for three months as a legacy option before its retirement on June 5, 2026.

      The rapid pace of model releases—GPT-5, 5.1