<h2>What launched</h2>
GPT-5.5 is now live in ChatGPT and Codex for paid subscribers, and available in the API as both <code>gpt-5.5</code> and <code>gpt-5.5-pro</code>. It’s also rolling out across GitHub Copilot (Pro+, Business, and Enterprise), Cursor, Perplexity Max, Devin, and a growing list of third-party tools.
The model is designed for long-horizon agentic work — the kind of task where previous models would quietly lose the plot halfway through. Key benchmark numbers:
<ul>
<li>82.7% on Terminal-Bench 2.0 (multi-step command-line workflows)</li>
<li>78.7% on OSWorld-Verified (operating real computer environments)</li>
<li>73.1% on Expert-SWE (tasks with a 20-hour median completion time)</li>
<li>35.4% / 39.6% on FrontierMath Tier 4 (standard and Pro)</li>
<li>64.66% on Databricks’ OfficeQA benchmark, up ~13% from GPT-5.4</li>
</ul>
Context window is 1M tokens via API, 400K inside Codex. OpenAI claims it matches GPT-5.4 latency while using fewer tokens per task — which matters because API pricing went up. Whether token efficiency actually offsets the cost increase is something people are actively testing.
Early feedback is mostly positive: less defensive boilerplate code, faster responses even at high reasoning settings, and better at staying on task without needing constant re-prompting.
<hr>
<h2>Codex got a major update alongside it</h2>
This isn’t just a model swap. Codex 3.0 is now a substantially different product:
<ul>
<li>Computer use — can operate apps on your Mac in parallel without interfering with your own work</li>
<li>In-app browser — interacts with web apps, clicks through pages, captures screenshots, and iterates until the task is done</li>
<li>Image generation and editing</li>
<li>Docs, PDFs, Sheets, and Slides support</li>
<li>OS-wide dictation</li>
<li>Auto-review mode — described as much safer than the previous “yolo” approach</li>
<li>Memory — learns from previous actions and remembers how you like to work</li>
<li>Thread automations for recurring tasks</li>
<li>Multi-terminal tabs per thread</li>
<li>SSH into devboxes</li>
<li>90+ new plugins including Atlassian, GitLab, Microsoft, CircleCI, Render, CodeRabbit, and Remotion</li>
</ul>
The computer use feature in particular seems to be landing well. Sam Altman called it “even more useful than I expected.” The Sky by Software acquisition, which brought real computer use to the platform, is looking like a good bet in hindsight.
Codex has hit 4 million weekly active users, adding over 1 million in less than two weeks.
<hr>
<h2>Enterprise rollouts</h2>
NVIDIA deployed Codex with GPT-5.5 to over 10,000 employees across all departments. The infrastructure runs on GB200 NVL72 rack-scale systems, with dedicated cloud VMs per employee and a zero-data retention policy. OpenAI has committed to deploying over 10 gigawatts of NVIDIA systems going forward.
OpenAI also launched a formal Codex partner program targeting enterprises that need help implementing and governing AI coding tools. Named partners include Cognizant, CGI, Accenture, Capgemini, PwC, TCS, and Infosys (via its Topaz AI platform). Use cases extend beyond code generation to legacy modernization, vulnerability detection, and agentic DevOps workflows.
For context: OpenAI’s enterprise segment accounts for over 40% of revenue and is on track to match consumer revenue by end of 2026.
<hr>
<h2>What people are building with it</h2>
Early users are reporting things like:
<ul>
<li>One-shotting F-Zero game clones with working AI opponents</li>
<li>Converting After Effects plugins to Remotion components in a single prompt</li>
<li>Building six products (iOS app, web app, investor deck, launch video, landing page, X automation) in under two hours</li>
<li>Full-stack QA testing at Ramp, where engineers use Codex to test end-to-end changes and flag what needs fixing</li>
</ul>
CodeRabbit, which had early access, reported meaningful improvements in code review quality. Cursor lists GPT-5.5 as the top model on CursorBench at 72.8%.
<hr>
<h2>The honest tradeoffs</h2>
API costs went up. That’s the main complaint. The argument from OpenAI is that fewer retries and better task completion mean you use fewer tokens overall — but that’s a claim worth verifying on your own workloads before committing.
The model is also genuinely more capable in ways that feel qualitatively different from previous increments. The jump from “agentic demo” to “actually finishes the task” is real for a lot of users, even if it’s hard to quantify.

Collections

GPT-5.5 is now live in ChatGPT, Codex, and the API (as gpt-5.5 and gpt-5.5-pro), rolling out across GitHub Copilot, Cursor, Devin, and other tools. The model targets long-horizon agentic tasks, posting strong benchmark scores including 82.7% on Terminal-Bench 2.0 and 73.1% on Expert-SWE. Codex 3.0 ships alongside it with major new capabilities: computer use on Mac, an in-app browser, image generation, document support, memory, auto-review mode, SSH into devboxes, and 90+ new plugins. NVIDIA deployed Codex to 10,000+ employees on GB200 infrastructure. OpenAI launched a formal enterprise partner program with Cognizant, Accenture, Capgemini, and others. API pricing increased, though OpenAI argues improved task completion reduces total token usage. Early user reports highlight less boilerplate code, faster responses, and better task persistence without re-prompting.

Codex adds computer use, browser, image generation, and 90+ plugins

Sweet you knew this was coming since Claude dropped opus 4.7 today. Looks like lots of great new stuff!!