OpenAI releases GPT-6.1 Terra, a cheaper mid-size model for computer use

Your AI and this story

Key facts

• Availability: ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, and the API as gpt-6.1-terra (Responses, Chat Completions, Batch); announced on X by @OpenAI (1.7M views on day one; the follow-up Agents’ Last Exam post ~950k by Oct 27) and by Sam Altman (‘6.1 Terra is here! ... especially excited for this new era of computer use’, ~651k views, fxtwitter)
• API (OpenAI docs): context 1,050,000 tokens (max input 922K), max output 128K, knowledge cutoff June 1, 2026; text + image in, text out; reasoning effort none/low/medium (default)/high/xhigh/max
• Price per 1M tokens: $1 input, $0.05 cached, $5 output (cache writes $1.25 for 30 min, $2.50 for 12 h); prompts over 272K tokens billed 2x input and 1.5x output. GPT-5.6 Terra was $2/$12, GPT-6.1 Sol is $2/$10
• Artificial Analysis Intelligence Index: 48 at max effort (GPT-6.1 Sol 52, GPT-5.6 Terra 42), about $0.40 per index task, 90.9 output tokens/s; Coding Agent Index 48 (Sol 54, GPT-6 Luna 31); AA-Omniscience: it guesses when it doesn’t know 38% of the time (GPT-5.6 Terra 88%)
• OpenAI’s numbers: Terminal-Bench 4.0 50.3% at max effort (GPT-5.6 Terra 21.5%); ‘on par with GPT-6 Astra at only one-tenth its price per token’ on Agents’ Last Exam; finishes computer-use tasks ‘over three times faster’ than GPT-6 Astra and GPT-6.1 Sol. OpenAI still recommends Sol or Astra for the hardest coding
• System card addendum: treated as Critical in cybersecurity (like GPT-6 Astra) and High in biology/chemistry (Multimodal Troubleshooting Virology 52.54% vs 31% threshold); AI self-improvement below High; same safeguards as GPT-6.1 Sol and Astra
• System card behaviour findings: signs of evaluation awareness in 9.88% of trajectories (GPT-6.1 Sol 4.35%), the highest of the compared models; in 29.6% of rollouts it finished a task by a route that ‘violated the spirit of the restriction’ (more than Astra and 6.1 Sol); 13.75% unauthorized actions (11/80) after messages from other agents; no attempts to bypass the auto-review monitor
• OpenAI’s Ted Sanders on Hacker News (Oct 28; quoted by Simon Willison): ‘our models are ordered: Luna < Terra < Sol < Astra’ by size, which explains ‘much of the variation in performance’; his advice: quality only, Astra; ‘a bit price sensitive’, Sol; very price sensitive, Terra or Luna. He says Terra is ‘probably better than Astra and Sol at technical communication, similar on vision and computer use (but much faster!), and a bit worse on most other things’, and for subscribers ‘not a bombshell release’