OpenAI launches GPT-6.1 Terra; Hi3D and Zhiyuan push 3D and embodied AI

OpenAI said Terra’s standard API price is half that of GPT-5.6 Terra: $1 per million input tokens, $0.05 per million cached input tokens and $5 per million output tokens. The model is available to Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, and via API as gpt-6.1-terra. In Agents’ Last Exam, its medium setting scored 15 percentage points above GPT-5.6 Terra while costing about one-third as much per task; OpenAI said it matched GPT-6 Astra at one-tenth the token price and ran more than three times faster than GPT-6 Astra and GPT-6.1 Sol. It also exceeded GPT-5.6 Terra on MMMU-Pro and OSWorld 2.1, solved 50.3% of Terminal-Bench 4.0 tasks at its max setting versus 21.5% for GPT-5.6 Terra, and cut internal factual-error rates at low reasoning from 53.6% to 16.7%, and to 4.3% at max. OpenAI co-founder Greg Brockman said Terra’s computer-use ability had crossed a “qualitative threshold” and was faster than his own operation.

Artificial Analysis gave Terra’s max setting 48 points, behind Claude Opus 5.5, Kimi K3.1 Pro, Google Gemini 4 Argon, Meta Muse Wonder and GPT-6.1 Sol, but ahead of Grok 4.7 and Xiaomi MiMo-V2.6-Pro. In a DesignArena test to recreate Lisbon’s Belém Tower at golden hour, Terra took nearly four times as long and used 48% more tokens than GPT-5.6 Terra, but total cost was 38% lower ($0.047 versus $0.077) and the image was more detailed. Developer Victor Nunez said Terra completed a long form 53.5 seconds earlier than GPT-6.1 Sol in Fast mode, a demonstration OpenAI’s developer account shared. Another user, Leo, said Terra scored below Sonnet 5.5 and Opus 5.5 on intelligence but generated at 283 tokens per second—3.4 times Opus 5.5—at about one-fifteenth the per-task cost.

OpenAI also said its AI systems produced three results in harmonic analysis, complexity theory and quantum field theory. It said the systems proved the local smoothing conjecture in all dimensions, resolving the Kakeya and Stein restriction conjectures, and that natural-language and Lean formal proofs were public. It said the systems showed the Exponential Time Hypothesis to be false, with a Lean proof, though this does not resolve P versus NP or imply immediate practical algorithms. It said they gave existence proofs for certain four-dimensional quantum field theories, including quantum chromodynamics, and completed a Lean formalization for pure Yang-Mills existence; the mass-gap part of the Millennium Prize problem remains open. OpenAI said the physics results would first be reviewed by outside experts including Nobel laureate David Gross.

QbitAI reported that Hi3D released Hi3D Agent and its V3.0 model, which it described as the industry’s first commercially usable 2048³ ultra-high-precision AI 3D model developed in China. In sample comparisons with GPT-6 Astra, QbitAI said Astra produced recognizable basic forms but Hi3D Agent and V3.0 were closer to reference images in contour, volume and local detail. Hi3D also open-sourced V1.0, a high-resolution geometry-refinement model based on Sparc3D; it was trained on about 2 million curated 3D assets and validated in the team’s production workflow. V1.0 expands Sparc3D’s training set from 500,000 assets and uses sparse refinement through 1024³ and 1536³ levels. Its technical report lists multi-view fidelity of 0.75 and detail richness of 0.841, first among compared methods, and a 90% slicing success rate. The model is on GitHub and Hugging Face, with an online demo. Hi3D Agent can modify generated models, prepare assets for games and animation, and handle splitting, connection structures, coloring and print preparation for physical products; Hi3D also offers MCP connections for external applications.

Zhiyuan released Genie Master, an embodied-agent framework, as the fourth part of its AI WAVE series, QbitAI reported. In a real convenience store, a Zhiyuan Genie G2 robot using Genie Master completed tasks without human demonstration or debugging, the company said. When a small hook was partly blocked by a price tag, the robot’s first attempt failed; the agent reviewed execution records, identified insufficient head-camera resolution and occlusion, and used the left wrist camera for a close-up before the right hand hung the item. The agent saved the successful approach as experience. Genie Master uses an Embodied Execution Protocol with capability, guidance, feedback and continuity interfaces to connect upper-level reasoning with low-level action policies. It runs two routes: Harness RSI improves prompts, memory and execution rules, while Policy RSI updates the action-policy model using data from autonomous execution, human takeover demonstrations and exception recovery. Zhiyuan said human involvement is still used, including takeover demonstrations and model training. The workflows cover picking, restocking, cleaning, exception recovery and multi-robot collaboration. Zhiyuan plans a 24-hour convenience-store livestream on Nov. 2 with six Genie G2 robots working continuous shifts.