OpenAI released GPT-6.1 Terra(opens in a new tab) on October 26 at $1 per million input tokens and $5 per million output tokens, with cached input at $0.05. OpenAI says it beats the GPT-5.6 Terra model it replaces at half the price. It is live in the API as gpt-6.1-terra and in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users.
The numbers that matter to a budget owner are per task, not per token. On OpenAI's Agents' Last Exam results, as reported by InfoWorld(opens in a new tab), Terra scored 60.0% at $1.01 per task. GPT-5.6 Terra scored 50.7% at $3.58, and OpenAI's own premium GPT-6.1 Sol scored 58.6% at $2.00. Claude Opus 5.5 still led at 63.2%, but at $8.82 per task. On the OSWorld computer use test, Claude Sonnet 5.5 led at 83.9% for $5.74, against Terra's 74.2% for $1.19. These are vendor-published benchmarks, so treat them as a shortlist, not a verdict.
The practical point: a model that costs about a quarter as much per benchmark task as the one it replaces changes which workloads pencil out. Agent pilots that were shelved because each run cost dollars may now cost cents. It also means a team that standardized on one premium model for everything is likely overpaying on routine work.
The counterweight comes from Greyhound Research's Sanchit Vir Gogia, quoted by InfoWorld(opens in a new tab): inference is only one line of agent cost. Review and integration still cost money, and in one illustrative coding workload, 5.4 extra seconds of human review erased the inference saving. A cheaper model that makes more mistakes your staff must catch is not cheaper.
So do not swap models on the price sheet alone. Pick three live workloads, run Terra beside your current model, and measure cost per finished, correct task, including review time. Keep model IDs pinned in code so a default change cannot move your bill.
Action items
OpenAI's GPT-6.1 Terra costs half as much as the model it replaces and scored higher on agent tests. Cost per finished task, not price per token, is now the number that decides which model to use.
For the CAIO
Rerun your model choices on three live workloads, count review time in the cost, and pin model IDs in code.
Researched and drafted by an automated workflow, then reviewed and edited by a human editor before publication. Every source is linked. See how we use AI here.
Also worth knowing
- InfoWorldOpenAI bets on AI economics with lower-cost GPT-6.1 Terra(opens in a new tab)
Terra lists at $1 and $5 per million tokens and scored 60.0% on Agents' Last Exam at $1.01 per task, against $3.58 for the model it replaces. Benchmark it on your own workloads before switching.
- Tech TimesEx-DOGE Startup's AI Agents Breached Anthropic in 12 Hours via Enterprise Software(opens in a new tab)
Cathedral's agents failed against Anthropic's own systems for days, then got into meeting transcripts and salary data through HR, CRM and IT vendor software in under 12 hours. Your SaaS vendors are your perimeter now.
- PR NewswireAI Is Saving Employees Time, but Employers Aren't Getting the Capacity They Expect, New Hubstaff Research Finds(opens in a new tab)
In a survey of 1,068 US workers, 32% of AI users reporting savings get back four hours or more a week, but only 26% use it to take on more work and 39% just finish earlier. Saved hours become capacity only if someone redesigns the work.
- The Motley FoolS&P Global (SPGI) Q3 2026 Earnings Call Transcript(opens in a new tab)
S&P Global says API call volume rose more than 4,000% year over year as customers wire its data into frontier models, and says AI rights deals now carry pricing that protects its IP. Expect data licensing to be a growing AI cost line.