GPT-5.5 Returns to Cutting Edge in Coding, But OpenAI Switches Benchmarks After Losing to Opus 4.7

Gate News message, April 27 — SemiAnalysis, a semiconductor and AI analysis firm, released a comparative benchmark of coding assistants including GPT-5.5, Claude Opus 4.7, and DeepSeek V4. The key finding: GPT-5.5 marks OpenAI’s first return to the cutting edge in coding models in six months, with SemiAnalysis engineers now alternating between Codex and Claude Code after previously relying almost exclusively on Claude. GPT-5.5 is based on a new pretraining approach codenamed “Spud” and represents OpenAI’s first expansion of pretraining scale since GPT-4.5.

In practical testing, a clear division of labor emerged. Claude handles new project planning and initial setup, while Codex excels at reasoning-intensive bug fixes. Codex demonstrates stronger data structure comprehension and logical reasoning but struggles with inferring ambiguous user intent. On a single dashboard task, Claude automatically replicated the reference page layout but fabricated large amounts of data, while Codex skipped the layout but delivered significantly more accurate data.

The analysis reveals a benchmark manipulation detail: OpenAI’s February blog post urged the industry to adopt SWE-bench Pro as the new standard for coding benchmarks. However, GPT-5.5’s announcement switched to a new benchmark called “Expert-SWE.” The reason, buried in fine print, is that GPT-5.5 was surpassed by Opus 4.7 on SWE-bench Pro and fell significantly short of Anthropic’s unreleased Mythos (77.8%).

Regarding Opus 4.7, Anthropic published a postmortem analysis one week after release, acknowledging three bugs in Claude Code that persisted for several weeks from March to April, affecting nearly all users. Multiple engineers had previously reported performance degradation in version 4.6 but were dismissed as subjective observations. Additionally, Opus 4.7’s new tokenizer increases token usage by up to 35%, which Anthropic openly admitted—effectively constituting a hidden price increase.

DeepSeek V4 was assessed as “keeping pace with the frontier but not leading,” positioning itself as the lowest-cost alternative among closed-source models. The analysis also noted that “Claude continues to outperform DeepSeek V4 Pro on high-difficulty Chinese writing tasks,” commenting that “Claude won against the Chinese model in its own language.”

The article introduces a key concept: model pricing should be evaluated by “cost per task” rather than “cost per token.” GPT-5.5’s pricing is double that of GPT-5.4 (input $5, output $30 per million tokens), but it completes the same tasks using fewer tokens, making the actual cost not necessarily higher. Initial SemiAnalysis data shows Codex’s input-to-output ratio at 80:1, lower than Claude Code’s 100:1.