loading
loading
OpenAI's GPT-6.1 Terra Cuts Agent Costs 50% and Runs 3x Faster OpenAI's new middle-tier model matches GPT-6 Astra on agentic benchmarks at one-tenth the price and triple the speed. - OpenAI launched GPT-6.1 Terra, a mid-tier model with 50% lower API prices than GPT-5.6 Terra. - Matches GPT-6 Astra on Agents' Last Exam at roughly one-tenth the price per token. - Runs over 3x faster than Astra and GPT-6.1 Sol on average agent tasks. - Available now in ChatGPT Work and Codex for Plus, Pro, Enterprise, and Edu users. - Developers can call it via the API as gpt-6.1-terra; Sol and Astra remain for heavy reasoning. - Positioned as OpenAI's answer to Anthropic on price for agentic computer-use workloads. OpenAI has launched GPT-6.1 Terra, a mid-tier model designed for agentic workflows, computer use, and routine coding. OpenAI positions it as the default model for many ChatGPT Work and Codex tasks, with Sol and Astra reserved for research and complex engineering. Standard API prices are 50% lower than those of GPT-5.6 Terra. The previous model, released on July 9, 2026, cost $2 per million input tokens and $12 per million output tokens. Terra 6.1 reduces those rates to $1 and $6, respectively. The middle tier gets a clearer job OpenAI’s current model family spans four tiers, each aimed at a different balance of cost, speed, and reasoning capability. Terra’s role centers on workflows that need stronger reasoning than Luna can provide at lower cost and latency than the top tiers. Typical jobs include navigating applications, completing forms, editing documents, reviewing code, and coordinating multiple tools. Tool use drives the pitch Computer-use models interpret interfaces, decide which actions to take, and operate software through clicks, keyboard input, or tool APIs. Reliability depends on recognizing interface state, recovering from errors, and maintaining the task’s context across many steps. OpenAI reports three main gains for Terra 6.1: - Higher agent scores: Terra substantially outperforms GPT-5.6 Terra on Agents’ Last Exam, a benchmark for operating workplace software. - Lower-cost performance: OpenAI says Terra matches GPT-6 Astra on that benchmark at about one-tenth of the per-token price. - Shorter runtimes: Terra reportedly completes tasks more than three times faster on average than Astra and GPT-6.1 Sol. Agent workflows can make dozens of sequential model and tool calls, so lower latency can cut total completion time substantially. Lower token prices also reduce model spend, although the final bill still depends on prompt length, generated tokens, retries, and third-party tool charges. One agent benchmark provides limited evidence about broader reasoning, coding quality, and consistency. Independent evaluations will need to test the parity claim across different interfaces, long-running tasks, and failure conditions. Average completion speed also leaves tail latency and retry rates unanswered. A dollar in, six dollars out OpenAI says Terra consumes usage allowances more slowly than Sol or Astra inside ChatGPT Work and Codex. The standard API rates above exclude other deployment costs, including external tools, browser infrastructure, storage, and human review. Microsoft left an early trail Public Microsoft repositories showed GPT-6.1 Terra in 18 evaluation configurations that had previously referenced GPT-5.6 Terra. A GitHub commit attributed to Copilot engineer Nate Finch also tested Terra as a replacement for GPT-6.1 Sol in code reviews, re-reviews, integration checks, and continuous-integration reviews using high reasoning effort. The artifacts document prelaunch testing. The scale and production status remain unclear. Agent scores have limits OpenAI continues to recommend Sol or Astra for deep research, difficult coding, mathematical proofs, and long reasoning chains. Luna remains the lower-cost option for simple, high-concurrency workloads. Previous Terra releases occupied an awkward middle position for some teams. Luna offered better economics for basic tasks, while Sol and Astra delivered more dependable results on difficult engineering work. Terra 6.1’s benchmark results suggest stronger computer-use performance, but they leave open questions about code correctness, mathematical reasoning, long-context behavior, and multi-turn consistency. Run these checks before migrating A production evaluation should measure Terra against the current model on representative tasks rather than isolated prompts. Useful tests include: - Task success: Measure completed workflows, partial failures, and cases requiring human repair. - End-to-end latency: Track median and tail completion times across every model and tool call. - Total cost: Include input, output, retries, external APIs, browser sessions, and review time. - Interface resilience: Change layouts, labels, permissions, and error states to test recovery. - Coding quality: Compare defect rates, test pass rates, review accuracy, and regression frequency. - Safety controls: Use least-privilege credentials, sandboxed environments, audit logs, and confirmation steps for destructive actions. Teams migrating from GPT-5.6 Terra can use the same task suite to quantify improvements before increasing traffic. Applications moving from Sol or Astra should test the hardest cases separately, since performance on routine agent tasks may conceal regressions in complex reasoning. OpenAI’s cost bet Lower prices change the deployment economics for agents that make frequent model calls. Forms, application navigation, document work, code review, and light development can generate substantial token volume at scale, making Terra’s pricing relevant even when each task is small. If independent tests confirm OpenAI’s speed and benchmark claims, Terra becomes a strong default candidate for agent development on the OpenAI stack. Its pricing also increases competition with Anthropic’s Haiku models and Google’s mid-tier Gemini variants, which target teams seeking lower inference costs than flagship models provide.
document fetched 2026-10-27 02:55