Semiconductor and AI analyst firm SemiAnalysis just dropped a comprehensive coding assistant benchmark pitting GPT-5.5, Claude Opus 4.7, and DeepSeek V4 against each other. The big takeaway: GPT-5.5 is the first OpenAI coding model in six months to reclaim a spot at the frontier. SemiAnalysis engineers have started switching between Codex and Claude Code — a notable shift, given that the team previously used Claude almost exclusively.
GPT-5.5 is based on a new pretraining run codenamed "Spud," marking OpenAI's first expansion of training scale since GPT-4.5.
The benchmarks revealed an emerging division of labor: Claude handles new project planning and initial scaffolding, while Codex takes over reasoning-heavy bug fixes. Codex is stronger at understanding data structures and logical reasoning, but it struggles to infer vague user intent. In one dashboard task, Claude automatically reproduced the reference page layout but fabricated a lot of the data; Codex skipped the layout but nailed the data.
Here's where it gets interesting. The article exposes a subtle benchmark swap: back in February, OpenAI published a blog post urging the industry to adopt SWE-bench Pro as the new standard for coding benchmarks. But GPT-5.5's announcement used a different benchmark called "Expert-SWE." The reason is buried in the fine print at the bottom of the release: GPT-5.5 was actually outperformed by Opus 4.7 on SWE-bench Pro, and was far below Anthropic's still-unreleased Mythos model (77.8%).
On the Opus 4.7 side, Anthropic released a postmortem a week after launch, admitting that Claude Code had three bugs from March through April that persisted for weeks and affected nearly all users. Several engineers had previously reported performance degradation in version 4.6, but those reports were dismissed as subjective. Additionally, Opus 4.7's new tokenizer can increase token usage by up to 35% — a stealth price hike that Anthropic itself acknowledges.
DeepSeek V4 was described as "close to the frontier but not leading" — the cheapest closed-source alternative. The report also noted that "Claude still beats DeepSeek V4 Pro on difficult Chinese writing tasks," sarcastically adding that "Claude beat a Chinese model in its own language." (For more on DeepSeek V4, check out our coverage here.)
The article introduces a crucial concept: model pricing should be measured by "cost per task" rather than "cost per token." GPT-5.5's unit price is double that of GPT-5.4 ($5 per million input tokens, $30 per million output tokens), but it uses far fewer tokens to complete the same task, so the actual cost may not be higher. SemiAnalysis's preliminary data shows Codex has an input-to-output ratio of 80:1, lower than Claude Code's 100:1. (OpenAI revealed the API pricing details for GPT-5.5 shortly after the announcement.)