The New Fontier Model Generation - GPT 5.5 vs Opus 4.7

Ronni Holmvig Strøm · 2026-05-10

Two frontier releases, seven days apart, and the press cycle treated them as a duel. They are not. Anthropic shipped Claude Opus 4.7 on April 16, 2026; OpenAI followed with GPT-5.5 on April 23. Read the release notes side by side and a quieter pattern surfaces. The two labs are no longer chasing the same scoreboard. One is investing in the model's capacity to hold a long task without a human in the loop. The other is investing in the moment between question and answer, making it cleaner, faster, less likely to mislead. Both are frontier. They are answers to different questions.

Anthropic shipped Claude Opus 4.7 on April 16, 2026 and OpenAI followed with GPT-5.5 on April 23. Read the release notes side by side and a quieter pattern surfaces.

The two labs are no longer chasing the same scoreboard. One is investing in the model's capacity to hold a long task without a human in the loop. The other is investing in the moment between question and answer, making it cleaner, faster, less likely to mislead. Both are frontier. They are answers to different questions.

For builders deciding what to put behind a production endpoint, that distinction is the whole story.

Coding and Agentic Depth

Opus 4.7 records the clearest gains on sustained coding workflows. Hex's 93-task coding evaluation shows a 13% resolution lift, with four tasks completed that neither Opus 4.6 nor Sonnet 4.6 could finish. CursorBench rose from 58% to 70%. Rakuten's SWE-Bench reports three times more production tasks resolved, with double-digit improvements in both code quality and test quality.

Testers at Notion and Hex describe the same property in different words: the model holds coherence across hours-long asynchronous tasks, recovers from tool failures, and verifies its own outputs before returning.

GPT-5.5 narrows the gap on precise tool use and file navigation. It leads Terminal-Bench 2.0 at 82.7%, up from 75.1% for GPT-5.4. Early enterprise reports note stronger one-shot system diagnosis. The model identifies why a codebase is failing and predicts downstream effects more reliably than prior versions. On SWE-Bench Verified, however, it trails Opus 4.7 in several independent comparisons. GPT-5.5 favors rapid iteration over extended autonomous execution.

A compact view of the headline coding numbers:

The pattern is not subtle. Opus is built for the run that lasts. GPT-5.5 is built for the turn that lands.

Multimodal and Creative Output

Opus 4.7 ships with 3.75-megapixel vision support. The upgrade improves interpretation of technical diagrams, chemical structures, and dense interfaces. Testers at Solve Intelligence and XBOW report measurable lifts on patent workflows and visual-acuity benchmarks. The model also produces higher-quality professional artifacts (dashboards, slides, documentation) with noticeably better design taste and fewer scaffolded placeholder functions.

GPT-5.5 Instant — the May 5 default-tier release — scored 76 on MMMU-Pro against 69.2 for GPT-5.3 Instant, OpenAI's previous default. Its emphasis is on cleaner, less verbose outputs rather than richer visual fidelity. The Instant variant, now default in ChatGPT, prioritizes immediate readability across text and simple diagrams. For teams that need high-fidelity image understanding inside agent loops, Opus retains the edge. For teams whose users open a chat window and want the right answer in two sentences, GPT-5.5 reads better.