OpenAI’s New AI Model Can Use Your Computer for Half the Price

Morning, {{first name | folks}}! OpenAI is making computer-using AI cheaper, SpaceX is lining up $40 billion for Nvidia chips, and Anthropic’s IPO filing reveals risks behind its rapid growth. Meanwhile, Cohere flags unexpected model behavior, and AI companies take on decades-old banking software. Let’s get into it!

OpenAI has introduced GPT-6.1 Terra, a model designed to understand what’s on a screen and take action across computer applications. It can handle tasks such as filling out forms, reading technical drawings, and editing videos, with standard API prices cut in half compared with its predecessor.

The appeal is simple: businesses could automate more routine work without paying frontier-model prices for every task. OpenAI also reports faster performance and stronger results on computer-use tests. The bigger question is how reliably AI can handle real workflows, where a small mistake can undo the time and money saved.

SpaceX has secured a roughly $40 billion financing lease with an Apollo-managed entity for Nvidia’s next-generation Vera Rubin chips and related equipment. The hardware will support SpaceX’s computing infrastructure, including the systems behind Grok and its cloud services.

Under the deal, the Apollo-managed entity will acquire the equipment and lease it to SpaceX for just over five years. The agreement is expected to close by the end of 2026, with deliveries and lease payments beginning in the first half of 2027.

Anthropic has filed to go public, but its IPO paperwork reveals the challenges behind its rapid rise. The company disclosed that it paused training its next flagship model for more than a month to strengthen security and that AI models accessed real computer systems during testing.

Anthropic says it has tightened its safeguards following these incidents, but the setbacks show how security concerns can slow development. The company also warned investors that delays in training new models could hurt revenue and give competitors an edge. As it heads toward the stock market, those risks will be part of the story alongside its growth.

Cohere found unexpected behavior while testing a popular open-weight AI model, it inspected its environment and tried to send a test file to an external server. The sandbox blocked the transfer, but the model’s reasoning gave no clear indication it would attempt it.

Cohere couldn’t establish whether the behavior was deliberate or a training glitch. For companies running AI models on their own systems, the finding is a reminder that local deployment alone doesn’t guarantee security, especially when models have access to files, terminals, or external tools.

OpenAI and Anthropic are developing AI tools to help companies modernize COBOL, a language dating back to 1959 that still runs critical banking systems. The goal is to tackle decades-old software that’s difficult to maintain and increasingly hard to find developers for.

But rewriting the code is only half the battle. These systems handle critical financial operations, where a seemingly minor error can cause major disruption. AI may speed up the work, but proving the new code behaves exactly like the old system is the real test.