Tech • AI • Robotics • Game
OpenAI's New GPT-6.1 Terra Is Way Better Than It Should Be (10x Cheaper, 3x Faster)
9.4/10TL;DR
OpenAI has launched GPT 6.1 Terra, a cheaper and faster mid-tier model that narrows the gap with flagship systems as competition from Anthropic and new open models intensifies.
KEY POINTS
Terra moves from leak to launch
OpenAI officially released GPT 6.1 Terra on October 26 after developers spotted references to it in public code tied to Microsoft and GitHub tools. In one notable clue, internal test setups switched the model used to judge AI outputs from GPT 5.6 Terra to GPT 6.1 Terra, suggesting strong confidence in its reasoning. GitHub Copilot code references also indicated testing for code review and pre-shipping checks.
Lower price, higher speed
GPT 6.1 Terra is priced at $1 per million input tokens and $5 per million output tokens, half the token price of GPT 6.1 Soul. It is available through the OpenAI API, ChatGPT Work, and Codex. On professional agent tasks, OpenAI says Terra runs more than three times faster than both Soul and Astra.
Flagship-level results on some tasks
On Agents Last Exam, which measures AI performance on real professional tasks, Terra performs on par with GPT 6 Astra at roughly one-tenth the price per token. On Terminal Bench Science, focused on code-driven scientific problem solving, Terra reportedly beats the more expensive GPT 6.1 Soul. Its coding score on Terminal Bench 4.0 rose from 21.5 to 50.3, more than doubling from the previous version.
A reshaped product ladder
The GPT 6 lineup is increasingly defined by use case and cost rather than a simple intelligence ranking. Luna is the low-cost fast model, Terra the mid-range value option, Soul the premium workhorse, and Astra the flagship. While Soul and Astra remain stronger on the hardest coding and reasoning tasks, Terra’s combination of speed, price and broad capability makes it more attractive for visual work, computer use and everyday professional tasks.
Pressure from Anthropic is mounting
The release comes as Anthropic’s Claude family gains momentum among developers. Data from OpenRouter showed weekly token usage for Anthropic models rising by 3.6 times after Opus 5.5 launched. A common workflow now pairs cheaper Haiku 5.5 for routine tasks with Opus 5.5 for harder reasoning, putting pressure on rivals at both the low and high ends of the market.
OpenAI’s pricing looks defensive and aggressive
OpenAI appears to be responding by pushing stronger performance into a lower-cost tier. The company says Terra completes certain professional tasks at roughly nine times lower cost than Claude Opus 5.5. If those results hold outside internal testing, the move could force competitors to cut prices or improve throughput rather than rely only on top-end benchmarks.
A Chinese recruiting firm enters the model race
Kanzhun, the company behind the Chinese hiring platform Boss Zhipin, released an open model called NanBei 4.33 33B aimed at coding agents and office work. The model is free for commercial use on Hugging Face, supports a 512,000-token context window, and is intended both as a directly usable model and as a teacher for distilled smaller versions.
NanBei posts strong but partly unverified claims
NanBei says it beats Alibaba’s Qwen 3.8 27B on 18 of 20 tests, though a closer reading suggests a fairer summary is 12 wins, six ties and two losses. Its largest leads were on long multi-step agent tasks such as bug fixing, command-line work, office automation and planning. But the comparisons are self-reported, there is no independent intelligence score yet, and it remains unclear whether Qwen was run with its own best settings.
Claude audits a year of physics work in 10 hours
In a separate sign of rapid capability gains, physicist Emre Onur Kaya of Istanbul Technical University gave Claude Opus 5.5 his PhD-era quantum gravity calculations, a project that originally took about a year. Roughly 10 hours later, the system returned a 183-page derivation, code for repeated numerical checks, and a list of errors including a sign mistake, omitted terms and incorrect constants. After corrections, the main scientific conclusion remained unchanged.
CONCLUSION
The release of GPT 6.1 Terra highlights a new phase in the AI race, where cost, speed and task completion matter as much as raw flagship prestige. With Anthropic, OpenAI and new open entrants all pushing harder, the market is shifting toward cheaper systems that can handle serious professional work.
Ask a question