OpenAI's cheaper GPT-6.1 Terra and its Berlin platform pitch

OpenAI introduced GPT-6.1 Terra on Monday at $1 per million input tokens and $5 per million output tokens. That is half the price of GPT-6.1 Sol, released nearly a month earlier at $2 per million input tokens and $10 per million output tokens, and below GPT-5.6 Terra at $2 per million input tokens and $12 per million output tokens. It is the fourth consecutive time OpenAI has cut the price of its models. Anthropic has been cutting prices too, most recently for Haiku 5.5 in October, as both compete with open-weight models from Chinese vendors such as Alibaba, Baidu and Z.ai.

Separately, OpenAI held its DevDay Exchange 2026 event in Berlin, where it mostly repeated announcements already made in San Francisco, with a few new details about Dots in Europe, Codex usage and its plans for developers.

GPT-6.1 Terra: what changed and what it costs

OpenAI describes GPT-6.1 Terra as fast and efficient, combining visual reasoning with precise computer use. It says the model is good at computer-based tasks such as drafting proposals, filling out forms and editing videos. OpenAI also says Terra has a latency advantage over its Astra and Sol models on some tasks. Gartner analyst Arun Chandrasekaran said that means it can speed up agentic, high-volume enterprise workflows, but warned that this calls for more caution: "Greater autonomy also increases operational and security risks, making model routing, task-level ROI, access controls, and human oversight essential to scaling adoption safely."

Model choice is getting harder as the price war widens

Lower prices come with a practical problem for buyers. Futurum Group analyst David Nicholson said enterprises are a long way from being able to use the most cost-effective model for each task, and that more choice is causing confusion. Cheaper models let companies optimise token use and cost, but he said there is friction: you need to understand what each model is suited for and how your task lines up with it, and that is not always clear. Nicholson also said the race to lower prices could cut both ways for OpenAI and Anthropic as they pursue public listings. Price cuts could lead investors to think the technology is less valuable, or could signal greater demand and profitability.

Berlin DevDay: the Agents API pitch and platform lock-in

The Decoder's view of the Berlin event is that it showed how tightly OpenAI wants to lock developers and businesses into its ecosystem. With the Agents API, OpenAI runs the agent infrastructure itself, while plugins and web apps give developers more ways to offer their products on its platform. The pitch is less technical work and access to what OpenAI says are 1.2 billion weekly ChatGPT users, a model that looks a lot like earlier app stores. Keeping the underlying runtime, Codex Harness, open source does not change this. Developers can read the code and run it themselves, but anyone relying on managed services, platform-specific integrations and finding customers through ChatGPT will likely have a hard time switching later. In Germany, OpenAI said weekly active Codex users have grown twentyfold since January 2026, again without absolute numbers. It also said nearly one in five desktop users uses skills for documents, presentations or spreadsheets, meaning they use the coding tool for general office work.

Dots coming to Europe, Sora gone for now

Romain Huet, OpenAI's head of developer experience, said the AI assistant Dots should become more widely available in Europe "in a few weeks" following talks with regulators. Personal Pro accounts in the European Economic Area, Switzerland and the UK are still locked out, though certain business accounts can already get access. The panel again blamed extreme resource demands for shutting down the Sora video generator in March, saying customers mainly wanted coding models and agents, and did not rule out bringing it back. Asked why OpenAI still has no model that understands video sequences the way Google's Gemini does, management said customers have barely asked for it, so it has not prioritised the feature beyond single-frame analysis.