30 Days That Changed How We Write Code: Cursor 3, Codex, Claude 4.7, and GPT-5.5
In the span of roughly 30 days — from April 2 to May 5, 2026 — every major AI coding tool shipped a significant update. Not incremental improvements, but fundamental rethinks of what the tool is actually for. Cursor rebuilt its interface from scratch around parallel agents. OpenAI gave Codex a computer and 90 plugins. Anthropic released Claude Opus 4.7, aimed squarely at the hardest engineering problems. And GPT-5.5 Instant became ChatGPT's new default model, two days ago.
They're not coordinating. But they're all pointing at the same thing: the developer's job is shifting from writing code to directing agents that write it. Here's what actually shipped, and what it means for the web developers building with these tools right now.
Cursor 3: The IDE Built for a World Where Agents Write the Code
On April 2, Anysphere didn't ship an update to Cursor — they threw out the VS Code fork they'd been building on and started over. Cursor 3 is built on a single premise: most professional code will be written by AI agents, and the developer's primary job is to orchestrate them.
The centre of the new interface is the Agents Window — a full-screen workspace that replaces the old Composer pane. You can run as many parallel agents as you need, each working across different files, repositories, or remote environments simultaneously. One agent refactoring a legacy API, another writing tests, a third fixing a CSS layout issue. All at once.
Design Mode is the addition that will resonate most with web developers: in Cursor 3's integrated browser, you can shift-drag to select any UI element and drop it directly into an agent's context with Cmd+L. Instead of describing "the submit button on the checkout page," you point at it. The agent sees what you see. For anyone who has spent time translating visual intent into text prompts, this removes a persistent friction point.
Cursor 3 also ships multi-repo parallel execution, local-to-cloud agent handoff, and a plugin marketplace. The team says they now work almost entirely through agents internally — the strongest possible signal about where they think this is going.
OpenAI Codex: Now with a Cursor and 90 Plugins
On April 16 — the same day Anthropic released Opus 4.7 — Codex received its most substantial update yet: computer use, an in-app browser, image generation, memory, and over 90 plugins.
The headline feature is Background Computer Use. Codex can now operate macOS alongside you, with its own cursor, running multiple agents in parallel without interfering with your foreground work. It's the same conceptual shift as Cursor 3's Agents Window, but applied at the operating system level: agents that exist in your environment and act within it, not just inside a chat pane.
The underlying model powering Codex, GPT-5.3-Codex (released February 2026), carries a notable footnote: it's the first model meaningfully involved in creating itself. OpenAI's team used early versions to debug training runs, manage deployment, and diagnose evaluation results. That feedback loop — AI improving the process of making AI — is either a curiosity or a sign of things to come, depending on how you read the trajectory.
OpenAI also released a research preview of GPT-5.3-Codex-Spark, a smaller, faster sibling delivering over 1,000 tokens per second. Real-time coding suggestions at that speed start to feel less like autocomplete and more like a second brain keeping pace with your thinking.
Claude Opus 4.7: The Hard Problems, Delegated
Also on April 16, Anthropic released Claude Opus 4.7 — the successor to Opus 4.6, which had launched just two months earlier in February. The focus is deliberate and narrow: Anthropic describes it as a "notable improvement in advanced software engineering, with particular gains on the most difficult tasks."
That framing matters. Opus 4.7 isn't trying to be better at everything — it's specifically targeting the work that previously required a senior engineer's direct attention: deeply tangled debugging, cross-repository architectural changes, problems that require holding many things in mind at once. Users report handing off their hardest coding work to it with confidence.
For web developers, the improved vision is worth noting: Opus 4.7 can process images at substantially higher resolution and reasons about visual problems with genuine accuracy. Pass it a design mockup or a screenshot of a broken layout, and it can work with what it sees rather than requiring you to describe it.
Opus 4.7 is available across the Claude API, Amazon Bedrock, Google Cloud's Vertex AI, and Microsoft Foundry. Pricing is unchanged from Opus 4.6: $5 per million input tokens and $25 per million output tokens.
GPT-5.5 Instant: The New Default, As of Two Days Ago
The most recent development in this run of updates: GPT-5.5 Instant became ChatGPT's new default model on May 5 — two days ago. It replaces GPT-5.3 Instant, and the changes are meaningful in ways that matter to developers using ChatGPT as a day-to-day tool.
OpenAI reports that GPT-5.5 Instant produces 52.5% fewer hallucinated claims than its predecessor on high-stakes prompts covering medicine, law, and finance. For web developers using ChatGPT to research APIs, debug stack errors, or draft technical content, that reduction in confident-but-wrong output is practically significant.
It's also considerably more concise — roughly 30% fewer words and lines to make the same point — and OpenAI specifically called out a reduction in "gratuitous emojis" in responses. These feel like small quality-of-life changes, but they reflect a maturation in what we expect from a daily-use AI tool: less performance, more precision.
For Plus and Pro users, GPT-5.5 Instant can now pull from past conversations, uploaded files, and Gmail to give more personalised answers. Mobile support for this feature is coming soon.
What This All Means for Web Developers
Parallel execution is the new normal. Both Cursor 3 and Codex now let you run multiple agents simultaneously across different tasks and repositories. Learning to decompose work into parallel-safe agent tasks is increasingly a core developer skill.
Visual context is becoming part of the conversation. Cursor 3's Design Mode and Opus 4.7's improved vision both point toward a workflow where you hand an agent a screenshot instead of a description. For anyone building web interfaces, this makes AI tools meaningfully more useful on the visual side of the job.
The hardest problems are moving into scope. Codex's self-involvement in its own training and Opus 4.7's focus on "the most difficult tasks" both suggest the ceiling for what AI can handle in software engineering is rising faster than most people assumed.
Accuracy is improving where it matters. GPT-5.5 Instant's hallucination reduction on high-stakes domains is a sign that the reliability gap is closing. Not closed — but the direction is clear.
What Hasn't Changed
The developers getting the most from these tools right now aren't the ones who've stepped back entirely and let agents do everything. They're the ones who've gotten sharper about what to delegate and what to hold. Knowing your codebase, recognising when an agent's output needs rethinking, understanding what "correct" looks like for your specific product — none of that has been automated. Those judgement calls are more valuable in a world where the bottleneck is no longer hours of typing.
Thirty days is a short window to draw sweeping conclusions. But when Cursor, OpenAI, and Anthropic all ship major updates pointing in the same direction, within weeks of each other, it's reasonable to pay attention. The agent-first wave isn't a prediction anymore.