GPT-5.2: The Model That Had to Answer Google

On December 11, 2025, OpenAI officially launched the GPT-5.2 model family — and the timing was no coincidence. The previous month, Google had introduced Gemini 3, which for the first time in history knocked OpenAI off the top spot in several key benchmarks. According to The Information, Sam Altman even declared an internal "Code Red" — a state of emergency aimed at accelerating development so OpenAI could regain its lead.

GPT-5.2 arrived in three versions: Instant (a fast model for everyday tasks), Thinking (deeper reasoning for complex projects), and Pro (the most accurate — and most expensive — variant). The model offered a context window of 400,000 tokens, output of up to 128,000 tokens, and a knowledge cutoff of August 31, 2025.

The key benchmarks spoke clearly: on the GDPval test (44 professions, simulating knowledge work), the Thinking version excelled — according to human evaluators, it beat or matched professionals at 70.9% of tasks. In coding on SWE-bench Pro, it reached 55.6% (new SOTA), and on ARC-AGI-1, the Pro version became the first model to surpass 90%. And what matters most to the average user: hallucinations dropped by 38% compared to GPT-5.1, as confirmed by Max Schwarzer of OpenAI's post-training team.

How Much Does It Cost?

API prices reflect the massive computational demands. GPT-5.2 Thinking costs $1.75 per million input tokens and $14 per million output tokens. The Pro version jumps to $21 per input and $168 per output — that's 40% more than the previous generation. For comparison: Gemini 3 Pro charges $2/$12, Anthropic's Claude Sonnet 4.5 charges $3/$15. For Czech developers and businesses, this means model selection is increasingly an economic decision — not just a technical one.

Since the release of GPT-5.2, OpenAI has shipped further generations — GPT-5.5, GPT-5.5 Instant, and now in the summer of 2026 even GPT-5.6 as the preferred model in Microsoft 365 Copilot. The December launch of GPT-5.2 thus represented a pivotal moment when OpenAI proved that despite growing competition, it can still deliver top-tier performance.

Disney + OpenAI: A Billion Dollars for 200 Characters in Sora

On the same day, OpenAI and The Walt Disney Company announced a groundbreaking deal. Disney invested $1 billion in OpenAI and became the first major licensing partner of Sora, OpenAI's generative video platform. Under the three-year agreement, Sora could create short videos featuring more than 200 characters from the Disney, Marvel, Pixar, and Star Wars portfolios.

Among the available characters were Mickey Mouse, characters from Frozen, Toy Story, the Avengers — from Iron Man to Thor, from Darth Vader to the Mandalorian. An important detail: the contract explicitly excluded the likenesses and voices of real actors.

"Disney is the global gold standard of storytelling," said Sam Altman. "This deal shows how AI companies and creative leaders can collaborate responsibly." Robert A. Iger, Disney CEO, emphasized that it was about "thoughtfully and responsibly expanding the reach of our stories through generative AI."

The deal also had a practical dimension: Disney became a major enterprise customer of OpenAI, using ChatGPT for employees and the API for new products including Disney+. Fans could watch a curated selection of Sora videos directly on the platform.

What does this mean for Europe? The Disney–OpenAI deal set a precedent for licensing content for AI training and generation. The EU AI Act requires transparency around training data — this deal demonstrates one possible model for how large corporations can legally and responsibly enter the world of generative AI without violating copyright.

DeepSeek in Trouble: Banned Blackwell Chips and a Diplomatic Skirmish

In December 2025, Chinese startup DeepSeek found itself at the center of controversy. The Information reported that DeepSeek had acquired banned Nvidia Blackwell chips for training its next models — chips the U.S. government had explicitly prohibited from being exported to China. Nvidia, however, quickly denied the report, as covered by CNBC.

Interesting context: at practically the same time, the Trump administration approved the sale of H200 chips to China — a partial easing of previous sanctions, albeit with conditions. This raised questions about whether Washington can truly control the flow of advanced AI chips effectively.

Since then, DeepSeek has managed to launch V4, a model that by its own metrics surpasses Google Gemini, ChatGPT, and other U.S. systems — at a fraction of the cost. DeepSeek V3.2 runs at $0.28 per million input tokens and $0.42 per output, which is in a completely different price league compared to GPT-5.2.

Where AI Has Moved in Seven Months

A lot has happened since December 2025. OpenAI has since released GPT-5.5 and GPT-5.6. Anthropic introduced Claude Opus 4.7, then 4.8, and Claude Mythos/Fable 5 — a model that found over 10,000 software bugs in its first month. Google launched Gemini 3.5 Pro and Flash and lowered subscription prices. Meta shifted its strategy toward agentic AI.

Czech companies, meanwhile, are gaining access to these models via API as well as through cloud services like Microsoft Azure or Google Cloud. GPT-5.6 has been the preferred model in Microsoft 365 Copilot since July 2026, and it supports Czech. Claude and Gemini both have Czech localization within their language portfolios.

One thing is certain: December 2025 showed that the AI industry is entering a phase where it's no longer just about who has the best benchmarks. It's about money, content rights, chip geopolitics, and user trust. And in this race, the battle is being fought on all fronts simultaneously.