OpenAI Cuts GPT-6.1 Terra Price in Half

OpenAI quietly but effectively refreshed its mid-tier enterprise models. The company released GPT-6.1 Terra, replacing GPT-5.6 Terra. The new model centers on multimodal Computer Use and visual reasoning; its API pricing is 50% lower than the previous version.

The timing is notable. While server costs and latency for flagship models are being debated, OpenAI is pushing not its most powerful model, but a cheaper and faster one built for everyday workflows. All performance figures the company shared are based on its own tests; independent measurements are not yet available.

It manages the screen and desktop software

GPT-6.1 Terra is not positioned as just a chat model that answers questions. It is presented as a work partner that can directly use the user’s screen and desktop applications. The three use cases OpenAI highlights are:

• Reading technical drawings: It can examine architectural plans, CAD drawings, and engineering schematics at the pixel level, convert them into 3D models, or produce structural error reports.
• Video review: It scans raw footage and animation frames one by one, producing timecode-stamped scene logs and consistency checks.
• Form and invoice processing: It can read scanned PDF invoices and enter them into the relevant fields in enterprise ERP and accounting software.

These capabilities are not new; however, OpenAI’s emphasis is on doing the same tasks at a lower cost. The Computer Use feature means the model understands what it sees on the screen and carries out clicking and typing steps itself. Because agents like this can produce hard-to-reverse results when they make mistakes, approval mechanisms become critical in enterprise use.

Three-times speed claim on the Agents’ Last Exam test

According to the technical data OpenAI shared, the model’s most obvious advantage is speed. On the Agents’ Last Exam test, which measures real-world workflows and multi-step tasks, GPT-6.1 Terra completed tasks more than three times faster than the company’s upper-tier models GPT-6 Astra and GPT-6.1 Sol.

The speed difference directly translates into cost and wait time in live customer support, code completion tools, and security operations that filter large data streams simultaneously. However, the comparison here was made against more expensive and slower models; it was not explained where Terra stands against competitors’ models.

Sharp drop in hallucination rate

The biggest obstacle to enterprise use is the model producing incorrect information in a confident tone. OpenAI reports a clear improvement in its internal stress tests using deliberately misleading and contradictory prompts. According to the company, the factual error rate at low reasoning levels fell from 53.6% in the old GPT-5.6 Terra to 16.7% in the new model.

OpenAI also says the model achieves a more natural style in emails, proposal text, and everyday business correspondence. These claims are also based on the company’s own evaluation; results from independent testing organizations are still awaited.

What happened to API prices?

The pricing side is the most concrete part of the update. Prices per million tokens are as follows:

• Standard input: $1.00 (50% discount compared to the previous generation)
• Cached input: $0.05 (95% discount)
• Generated output: $5.00 (50% discount)

The 95% drop in cached input is the biggest savings item for applications that repeatedly work on the same document or codebase. Agents working with long system prompts and fixed reference texts can recover a large portion of the cost from this item.

GPT-6.1 Terra fills the gap between the upper tier and the budget tier in OpenAI’s model family. The company’s recent focus on price cuts is directly related to cost pressure from enterprise customers. The model’s real-world performance will emerge through tests developers run in their own workflows.

Source: WebTuring