Code Red in San Francisco: OpenAI returns fire (and takes on Gemini) with GPT-5.2

If you thought you’d have a quiet December without updating your APIs, think again. After Google’s Gemini 3 stole the spotlight (and the lead on LMArena) in recent weeks, the alarm bells started ringing in San Francisco.

Literally. According to The Information, Sam Altman had declared an internal “code red” after seeing ChatGPT traffic decline in favor of Mountain View. The response arrived this Thursday: welcome GPT-5.2.

Not one, but three “flavors”

OpenAI has realized that “one size fits all” no longer works. GPT-5.2 is available to paying users and via API in three distinct versions:

1. Instant:

   The speedster. Optimized for routine tasks, translations, and quick answers. It costs little and moves fast.

2. Thinking:

   The brain. Here, the challenge is directly aimed at Google’s Deep Think mode. It’s designed for complex coding, long-document analysis, and math.

3. Pro:

   The heavyweight. Maximum accuracy, no compromises, for problems where getting it wrong isn’t an option.

The Battle of the Benchmarks (and APIs)

Fidji Simo (OpenAI’s CPO) was clear: “We designed 5.2 to unlock real economic value”. Translated from corporate speak: enough talk, we need tools that get the job done.

OpenAI says the Thinking version beats Gemini 3 and even Claude Opus 4.5 (the former coding king) on almost every reasoning benchmark, from complex math to software engineering. The figure we like? 38% fewer errors than its predecessor on complex responses. For anyone debugging all day, that’s music to their ears.

What’s missing? (Spoiler: the Banana)

Curiously, this launch has no new image generator. While Google is having fun with its Nano Banana (yes, the name still makes us laugh) and hyper-realistic photos, OpenAI has focused only on text and code. Rumors say the visual response will arrive in January.

Graffico’s Take 🔴

Why is this release crucial for a Tech Partner (and for you)?

GPT-5.2 isn’t just “smarter.” It’s designed for Agentic Workflows. That means we can build software where AI doesn’t just answer a question, but carries out multi-step tasks (e.g., “analyze this Excel file, look for trends online, write the report, and send it to the CRM”) with a level of reliability that was previously a pipe dream.

We like the AI race because, in the end, we’re the ones who win when we use these tools to build solid things.

Status: Available via API starting today (and we’re already updating the keys).