New AI models this week: OpenAI’s GPT-6.1 Terra and Nanbeige4.3-33B

OpenAI’s GPT-6.1 Terra, priced at $1.00 per 1M input tokens, was the one major commercial model to reach the price tracker between October 25 and November 1. The rest of the week’s notable releases were open weights on Hugging Face: a 33B agentic model from Nanbeige, a 1.7B edge model from Liquid AI, two OCR models and a speech translation model from Tencent. Prices below come from our API pricing tracker, which reads OpenRouter’s catalogue.

GPT-6.1 Terra: OpenAI’s mid-tier model at $1 per 1M input tokens

OpenAI released GPT-6.1 Terra on October 26, along with a Pro variant that OpenRouter lists at identical prices and context. Both cost $1.00 per 1M input tokens and $5.00 per 1M output tokens, with a context window of 1,050,000 tokens. The OpenRouter listing shows support for tool calling, structured output, and adjustable reasoning effort and verbosity. Press coverage of the launch reports that OpenAI describes the model as roughly half the price of its earlier Terra and GPT-6.1 Sol and more than three times faster than GPT-6 Astra. We could not load OpenAI’s own announcement, so those figures are second-hand and unverified. They are also OpenAI’s own tests, not independent results. See the GPT-6.1 Terra listing and the Terra Pro listing.

Nanbeige4.3-33B: An open 33B teacher model for agentic work

Nanbeige4.3-33B is a 33B-parameter text model released under the Apache-2.0 license. Its card calls it the flagship teacher model of the Nanbeige4.3 series, meant to pass agentic skills down to smaller models such as a planned 3B version. It supports English and Chinese with a context of up to 512K tokens. The publisher’s self-run results include 80.9 on Terminal-Bench 2.1, 60.3 on SWE-Bench Pro and 95.2 on AIME 2026. It trails Qwen3.8-27B on SWE-Bench Pro and Gemma4-31B on GPQA Diamond in the same table. Details are on the model card.

LFM3-1.7B: Liquid AI’s small model for local agents

LFM3-1.7B is a text-only model from Liquid AI with 1.73B parameters and a new tri-hybrid architecture. It offers a 256K-token context, 16 languages and a 4-bit footprint of about 1 GB. It ships under the LFM Open License v1.0. The card targets local agents, tool calling and structured output, and advises against agentic coding or complex multi-step reasoning. Liquid AI reports an average benchmark score of 49.37 against 47.24 for Qwen3.5-9B, though the larger Qwen model scores higher on GPQA Diamond (80.40 versus 63.37). All figures are vendor-reported. See the model card.

MiniCPM-OCR 1.0 and Unlimited-OCR-2: Two open models for document parsing

OpenBMB’s MiniCPM-OCR 1.0 is the first OCR model in its MiniCPM-V family, at about 1.3B parameters under Apache 2.0. It pairs a SigLIP vision encoder with a Qwen3.5 language model and covers text, tables, formulas, charts and seals, mainly in Chinese and English. OpenBMB reports 93.68 on Real5-OmniDocBench and about 1.2 pages per second on an RTX 5090.

Baidu’s Unlimited-OCR-2 is a 3B model under the MIT license built for very long documents. Its Sparse Reference Sliding Window Attention keeps per-step decoding cost nearly constant as page count grows, according to the card. Baidu also reports leading results on MPDocBench and MarathonDocBench, a benchmark it releases with the model. Both sets of results come from the makers.

Hy-SimulMT-30B-A3B: Tencent’s live Chinese-English speech translator

Tencent Hunyuan’s Hy-SimulMT is a 30B-parameter real-time simultaneous translation model for Chinese-English speech in meetings, speeches and interviews. The FP8 checkpoint outputs text only, so speech output needs a separate text-to-speech service. It is licensed under Apache-2.0. Tencent reports a StreamLAAL of about 3 seconds on the RealSI long-form benchmark, the lowest among the systems it compared. In a small study of 18 recordings with six human interpreters, the authors say quality was on par with experts, and note the result applies only to those samples. The model card has the details.

Also seen: Claude Duet, a work agent app from Anthropic

Anthropic’s Claude Duet is an iPhone and iPad app, not a new model. It is described as a proactive work agent that summarizes email and calendar items, drafts replies and carries out handed-off tasks. It asks for approval before it sends, pays or signs anything. Access is limited to a waitlist, and the app requires iOS or iPadOS 26.0 or later.

The week in numbers

• Models on the price tracker: 2, both from OpenAI (GPT-6.1 Terra and Terra Pro).
• Cheapest per 1M input tokens: $1.00 (GPT-6.1 Terra and Terra Pro).
• Most expensive per 1M input tokens: $1.00 (the same two models, so no spread this week).
• Open-weight releases covered: 5 (Nanbeige4.3-33B, LFM3-1.7B, MiniCPM-OCR 1.0, Unlimited-OCR-2, Hy-SimulMT).
• Largest context window: 1,050,000 tokens (GPT-6.1 Terra), against 512K for Nanbeige4.3-33B.