GPT-6.1 Terra
OpenAI’s release of GPT‑6.1 Terra, a mid-tier large language model positioned between Luna and Sol, is being evaluated mainly on its price–performance tradeoffs and fit for real-world workloads like technical writing, data tasks, and computer use. Commenters note that benchmarks and cost-per-task charts place Terra close to more expensive frontier models in some areas, but weaker on coding-heavy tests, leading to questions about whether developers should still favor Sol, Astra, or rival models such as Anthropic’s Opus and DeepSeek. The launch also reignites frustration over OpenAI’s fast-changing model lineup, naming scheme, and safety constraints, which many see as making it harder to choose and reliably standardize on a single model over time.
Overall positioning and performance
- Terra 6.1 is framed as a mid-range model between Luna and Sol on both price and capability.
- Some readers see it as close to Sol/Astra at much lower cost; others point to benchmarks (e.g., Terminal Bench 4.0, AA Intelligence Index) where it clearly underperforms Sol and Opus 5.5.
- Several note that benchmark curves are saturated and may be cherry‑picked; trust in vendor benchmarks is low.
- External eval aggregators cited in the thread often show Sol 6.1 x‑high both cheaper and smarter than Terra max on some suites, but Terra fills a cost/perf gap on others.
- PacBench, Runebench, and a data‑analytics benchmark show Terra as a strong, fast “cheap model” that lands on or near the Pareto frontier.
Use cases and target audience
- Many think Terra is not aimed at SWE-heavy workflows but at “computer work”: technical writing, office automation, vision/computer-use tasks, dashboards, and data analysis.
- Multiple commenters say for serious coding they prefer Sol, Astra, Claude Opus/Fable, or DeepSeek; Terra looks more appealing for help‑desk answers, question‑to‑SQL, PR reviews, and cheaper agents.
- For some coding harnesses, model‑switching costs (cache misses) make sticking with Sol more economical even when Terra might be “good enough.”
Cost, plans, and “cost per task”
- Terra’s API pricing (half Sol, 10× cheaper than Astra) is attractive for production agents where marginal cost matters.
- Subscription users are divided: those on flat‑rate plans feel price hikes + cheaper models reduce plan value; others argue cost-per-task has genuinely improved.
- There is skepticism about the “cost per task” metric, seen by some as marketing spin after subscription cuts, and about comparisons vs Anthropic where assumptions about token usage matter.
Naming, model matrix, and confusion
- Strong recurring theme: too many axes to reason about—model family (Luna/Terra/Sol/Astra), version (5.6/6/6.1), reasoning effort (none–max), plus “pro/instant” modes.
- Staggered releases (Sol 6.1 before Terra 6.1, no Astra 6.1 yet) make cross‑generation comparisons (e.g., Sol 6.1 vs Astra 6, Terra 6.1 vs Sol 6) confusing.
- Some want simpler size names (small/medium/large), or auto‑selection modes (“Auto medium”) that pick a Pareto‑optimal model+effort for them, or an LTS‑style stable line.
- An OpenAI employee explains the intent: Luna < Terra < Sol < Astra is primarily size/price; version numbers track generation; effort is separate “thinking budget.” But they acknowledge multidimensional tradeoffs and welcome naming suggestions.
Behavioral notes and limitations
- Terra is described as very fast but more verbose than Sol at higher efforts.
- Some users see models (including Terra/Luna) stopping early despite prompts, and refusing certain cybersecurity or GPU‑unlocking tasks, reinforcing demand for local/unrestricted models.
- Overall sentiment: technically impressive and potentially very useful for API/enterprise workloads, but underwhelming for devs hoping for an OpenAI answer to Opus 5.5 and adding to model/choice fatigue.