OpenAI Ships GPT-5.5 with 82.7% Terminal-Bench Score — The First Fully Retrained Agentic Model Reshapes Dev Work

OpenAI released GPT-5.5 on April 23, the first fully retrained base model since GPT-4.5 and the company's strongest agentic system to date. The model is rolling out to Plus, Pro, Business, and Enterprise users in ChatGPT and Codex, where roughly 4 million developers already use Codex weekly — including more than 85% of OpenAI's own staff across software engineering, finance, marketing, and operations. The launch matters less for the version-number bump than for what GPT-5.5 actually does: it executes long-running command-line workflows, debugs its own output, and finishes multi-step tasks with minimal hand-holding.

Terminal-Bench 2.0: A 13-Point Lead and What It Measures

GPT-5.5 scored 82.7% on Terminal-Bench 2.0, the benchmark that measures planning, iteration, and tool use across realistic command-line workflows. Claude Opus 4.7 sits at 69.4% on the same benchmark, putting GPT-5.5 more than 13 points ahead. Terminal-Bench is built around tasks that include reading documentation, running commands, recovering from errors, editing files, and verifying results — the same loop a junior developer runs through dozens of times a day. A double-digit lead on this benchmark is the closest available proxy for the question every engineering manager is asking: how much routine implementation work can an agent finish without a human in the loop?

What This Changes About Entry-Level Coding Roles

The agentic-coding capability gap between top frontier models and the average junior engineer's first six months on the job has narrowed faster than most teams planned for. The work GPT-5.5 demonstrably does well — implementing well-specified tickets, writing tests for existing code, refactoring within a defined module, debugging deterministic failures — is exactly the work most companies historically gave to new hires as ramp-up volume. The work it still struggles with is exactly the work entry-level engineers were not supposed to be doing yet: ambiguous product decisions, cross-system architectural calls, and judgment about what should not be built. The implication for early-career developers is that the apprentice ladder is being rewritten in real time, with the bottom rung shifting from 'execute well-specified work' to 'evaluate and direct AI-executed work.'

GDPval and the Knowledge-Work Generalization

GPT-5.5 also scored 84.9% on GDPval, OpenAI's economic-value benchmark of real-world knowledge-work tasks across 44 occupations. That number is harder to map cleanly to a single workflow than Terminal-Bench, but the direction is the same: the model is finishing more end-to-end tasks per unit of human supervision. For non-engineers, that translates to faster compression of analyst, junior consultant, and entry-level operations work — the same multi-step, document-heavy, tool-coordinating workflows that GPT-5.5 was retrained to handle. Pricing is unchanged versus GPT-5.4's per-token rate while using fewer tokens per task, which means actual cost-to-finish has dropped in a way that will accelerate enterprise rollout decisions through Q3.

What Professionals Should Do This Quarter

If your job description still describes the work you actually do day to day, your job description is probably stale. The professionals who will compound advantage out of GPT-5.5's release are not the ones who memorize its features — they are the ones who systematically map their own workflow against what the model can now finish unsupervised, and aggressively reroute their daily time toward the parts AI cannot do. For software engineers specifically, that means more code review of agent output, more time on system design and product judgment, and faster comfort with running multi-agent workflows in tools like Codex, Cursor 3, and Claude Code. For analysts, consultants, and operations staff, it means rebuilding your portfolio around AI-augmented deliverables rather than your prior individual-contributor output.