GPT-6.1 Terra

Canonical version: GPT-6.1 Terra.

GPT-6.1 Terra is OpenAI's model between GPT-6 Luna and GPT-6.1 Sol. It came out in late October 2026, about a month after the GPT-6 family launched without a Terra at all. When GPT-6 Sol and Luna replaced the GPT-5.6 models on 22 September, GPT-5.6 Terra got no successor, and the common reading on Hacker News was that Terra had never found its niche. So OpenAI changed its mind.

OpenAI's docs describe it as "a fast, efficient model for visual reasoning, precise computer use, and clear, natural writing". That's a different pitch from Sol (agentic coding) and Luna (cheap, high-volume steps). Terra is sold on what it does with a screen and a keyboard, and on how its prose reads.

Specifications and pricing

Batch and Flex cost half of Standard, Fast mode costs twice as much, and prompts over 272K input tokens cost more. Terra works on Chat Completions, Responses and Batch, but tool calling needs the Responses API, same as 6.1 Sol. The built-in tools are the same set as Sol's: web search, file search, image generation, code interpreter, hosted shell, apply patch, skills, computer use, MCP and tool search.

Two things stand out:

• It's exactly half of 6.1 Sol, on every token type. Against GPT-5.6 Terra, input is halved and output drops by 58%
• It keeps none. 6.1 Sol dropped the none reasoning effort, so on the 6.1 generation Terra is now the cheapest way to call something smarter than Luna without paying for reasoning tokens. If your pipeline still depends on none (or on temperature, which only works with it), that alone is a reason to look at Terra

Where it fits

The lineup now has four tiers again: Luna ($0.10 / $0.50), Terra ($1 / $5), Sol ($2 / $10) and GPT-6 Astra ($10 / $50). The gaps aren't even: Luna to Terra is a 10× jump, Terra to Sol 2×, Sol to Astra 5×.

Ted Sanders, who works at OpenAI, summed up the logic in the HN thread: "If you care about quality only, pick Astra. If you're a bit price sensitive, pick Sol." He described Terra's niche as "uniquely fast computer use", and said OpenAI might drop it and "go back to 3" tiers if adoption is poor. So Terra is on probation, and OpenAI says so openly.

I couldn't read OpenAI's launch post (it returned a 403), so this note has no first-party benchmark numbers. The few numbers below come from commenters.

Reception on Hacker News

The launch thread reached 633 points and 426 comments (around 27 October 2026).

• "A weird middle child." That was minimaxir's top comment, and many agreed. frisbee6152: "If Luna is not good enough for a task, then the marginal improvement of Terra tends not to be worthwhile." cesarvarela saw it the other way: "6.1 Terra gets you 90% of what 6.1 Sol gets you at half the price." h14h and eknkc said Terra-class models matter for cost-sensitive API work and for staying under subscription limits
• Benchmarks pointed both ways. chriddyp said Terra joins the Pareto frontier between Luna and Sol on Plotly's data-analytics benchmark, and ranks among the fastest of the 33 models they tested. XCSme (aibenchy) found Terra at max worse than Sol at twice the cost, but a good fit at medium. ralusek, citing Artificial Analysis, said "Sol 6.1 x-high is both cheaper and smarter than Terra 6.1 max". user43928 noted that Terra needs xhigh on Terminal-Bench 4.0 to get close to Sol
• Eval noise. When shepherdjerred asked why medium beat high on one chart, Sanders answered that error bars are "commonly like ±5%" and that "we don't massage any of our eval numbers"
• Naming took over the thread. Someone1234 asked: "Did Microsoft's chief branding officer move over to OpenAI?" Proposals included size suffixes, date-based names and an automatic picker. mediaman and mrighele defended the scheme as four price tiers, like Intel's i3 to i9. margalabargala's objection was the staggered releases: with models shipping one at a time, nobody can tell whether Terra 6.1 beats Sol 6. throw20261022 said Terra "should've been released together with 6.1 Sol and Luna"
• Mixed hands-on reports. desireco42: "it kind of does what I would ask Sol or Astra, but dont have to count tokens as much". agluszak said it "will stop at an earliest excuse", the same complaint they had about Luna. wellthatsinsa said it "drained my usage SO FAST". deno runs it as a cheaper tier in an orchestrator, with DeepSeek for subagents
• Refusals. notenchange said Sol 6.1 kept refusing GPU-unlocking work on mining cards over "cyber concerns", while Luna complied with weaker results. altcognito reported the same kind of refusal from Claude Opus 5.5 on a read-only security investigation
• The usual side debates. Whether to compare cost per task or cost per token (mapontosevenths called OpenAI's cost-per-task framing "weaselly"; ComputerGuru and wj said it's the metric that matters in production), Anthropic's ban on third-party harnesses for subscriptions, and whether Pangram can detect AI writing at all

Simon Willison posted his pelicans across reasoning levels, and commenters argued (again) about whether the pelican test means anything.

My take

Terra is the model I'd test for two kinds of work: computer-use agents, where latency per step adds up fast, and pipelines that need a bit more than Luna but still want none and Chat Completions. For agentic coding, the thread's numbers (and Sanders' own advice) point to Sol at medium or high, which can beat Terra at max for less money.

But I wouldn't build anything long-lived on it yet. OpenAI said itself that Terra might go away if adoption is poor, and Terra already disappeared once in the GPT-6 generation. Keep the model ID in config, run your own evals on Terra at medium against Sol at medium, and compare cost per task rather than price per token. On paper, Terra costs half as much. In practice, that only holds if it doesn't need twice the effort to get the job done.

Caveats

• OpenAI's launch article (https://openai.com/index/introducing-gpt-6-1-terra/) returned a 403 when I fetched it. The specifications and prices above come from OpenAI's model docs; the exact release date, OpenAI's benchmark charts and availability in ChatGPT and Codex are unverified
• One commenter (legulere) said Terra and Sol are both available in GitHub Copilot Pro; I couldn't confirm that from GitHub
• All benchmark claims in this note come from Hacker News commenters, not from OpenAI or Artificial Analysis directly
• The release date is inferred from the thread's age ("2 days ago" on 29 October 2026)

References

• OpenAI, "Introducing GPT-6.1 Terra" (not readable at the time of writing): https://openai.com/index/introducing-gpt-6-1-terra/
• OpenAI API model docs, GPT-6.1 Terra: https://developers.openai.com/api/docs/models/gpt-6.1-terra
• Hacker News, "GPT-6.1 Terra": https://news.ycombinator.com/item?id=50248124

Related

• GPT-6.1 Sol
• GPT-6 Sol
• GPT-6 Luna
• GPT-6 Astra
• GPT-5.6
• OpenAI
• OpenAI Codex
• GitHub Copilot
• Claude Opus 5.5
• Claude Sonnet 5.5
• Claude Haiku 5.5
• Artificial Analysis
• Model routing
• Large Language Models (LLMs)
• Context Window
• Model Context Protocol (MCP)
• Simon Willison
• Hacker News
• Claude Computer use

About Sébastien

Ready to get to the next level?