Tool of the Week: GPT-6.1 Terra, and what it costs per finished task
On October 26, OpenAI released GPT-6.1 Terra, a mid-tier model aimed at computer use: reading a screen, filling in forms, moving between apps. It is available in the API, and in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu plans. InfoWorld reports API pricing of $1 per million input tokens and $5 per million output, with cached input at $0.05. OpenAI says that is half the price of the model Terra replaces.
The number worth reading is cost per task, not cost per token. On OSWorld 2.1, a desktop-agent benchmark, OpenAI reports Terra at 74.2% for about $1.19 a task, against 57.1% at $4.76 for the previous Terra. InfoWorld notes that Claude Sonnet 5.5 still scores higher on the same test (83.9%) at about $5.74 a task. Those are OpenAI's published figures, and the coverage I found reports no independent testing yet.
Two analysts quoted by InfoWorld make the point that matters for anyone building with this. Pareekh Jain says the model suits high-volume routine work like invoice processing and document analysis, and recommends testing on your own workload, measuring completion rate, accuracy, execution time, cost per workflow, human intervention and compliance. Sanchit Vir Gogia argues cheaper inference is only one piece of agent economics: in his illustrative example, a few extra seconds of human review per item cancels out the savings.
That is the billable part. The model is the cheap component. The work is the harness around it: the queue that feeds it, the validation that checks its output, the review step for what fails validation, and the log that tells you what each completed item actually cost. A model that is right 74% of the time on a benchmark is a different proposition when the other 26% lands on a person's desk.
One more caveat from the same coverage: OpenAI says Terra reaches the "Critical" cybersecurity threshold under its Preparedness Framework. If the agent will touch systems of record, that belongs in your risk review before the pilot, not after.
Quick Hits
dbt v2 and dbt State are generally available. At its September 16 summit, dbt Labs said dbt v2, a Rust rewrite of the engine, is generally available. The company says it parses a 10,000-model project up to 10x faster than v1 and flags errors and lineage problems before anything runs. dbt State uses warehouse metadata to decide which models to build, skip, clone or defer, to cut unnecessary compute. Both v1 (Python) and v2 (Rust) stay Apache 2.0. If you run scheduled transformations, the compute-saving piece is the one to benchmark on your own project. Speed figures are dbt Labs' own.
The AI agent part of the dbt platform is still in preview. Same announcement: dbt Wizard, an agent that understands project lineage, tests and contracts and validates changes before they ship, is in public preview, with a CLI in public beta. Fivetran's Context Layer, which pairs structured data with docs and chat threads for agents, is in private beta. Preview and beta mean you can evaluate them, but they are not something to promise a client this quarter.
Prompt of the Week: Cost Per Completed Item
I run a recurring document workflow and want to know what it
really costs per completed item with an AI model in the loop.
Workflow: [e.g. vendor invoices arriving by email, about 800/month]
Model and prices: [model name, $ per 1M input tokens, $ per 1M output tokens]
Average tokens per item: [input] in, [output] out. If I do not know,
tell me how to measure it from a sample of 20 items.
Share of items that fail validation and need a person: [percent, or UNKNOWN]
Minutes a person spends on a failed item: [minutes, or UNKNOWN]
Loaded hourly cost of that person: [dollars]
Calculate:
1. Model cost per item and per month.
2. Human review cost per month from the failure rate.
3. Total cost per COMPLETED item (correct and accepted), not per
attempt.
4. Break-even: how much would the failure rate have to rise for
this to cost more than doing the work by hand?
5. The three inputs that move the answer most, ranked.
Show the arithmetic. Do not guess any value I marked UNKNOWN. List
what I need to measure first.Start with a sample of 20 real items. The measured failure rate is the finding.
One last thing
Writing this is one side of the work. The other is building it: data pipelines, file and reporting workflows, integrations between systems that were never meant to talk, and the automation that takes the manual steps out.
If a process in your week still depends on someone remembering to run it, book a free 15-minute audit and I will find your top 3 time-wasters. If it is not worth automating, I will tell you that.
Scott