tensorfolk

OpenAI ships GPT-6.1 Terra, a cheaper mid-tier model built for computer use

Big news

OpenAI released GPT-6.1 Terra on October 26 to replace GPT-5.6 Terra, pitching it at everyday screen-driven work like form filling and hopping between apps. API pricing drops to $1 input and $5 output per million tokens, with a 1.05M-token context and up to 128K output. It is available in ChatGPT paid plans, Codex and the API as gpt-6.1-terra.

Why it matters

Computer-use agents get much cheaper to run; Artificial Analysis puts cost per task at under a third of the previous Terra.

How to try it

GPT-6.1 Terra is OpenAI's lower-cost model for visual reasoning, computer use and writing, with a 1.05M-token context window and 128K max output.

  1. In ChatGPT on a paid plan, or in Codex, choose GPT-6.1 Terra from the model picker.
  2. In the API, set the model name to gpt-6.1-terra in your existing Responses or Chat Completions calls.
  3. Keep prompts under 272K input tokens where you can: longer requests are billed at a higher rate.
  4. Use Batch or Flex for non-urgent jobs, which cost half the standard price.

Open the official docs

Source: Read the full story on OpenAI

openaigpt-6.1computer-useapipricing

More in Foundations & Models

Foundations & Models / Model Releases & Landscape

Moonshot launches Kimi K3.1 Pro, with open weights promised by Nov 4

Moonshot shipped Kimi K3.1 Pro, a further-trained Kimi K3, in its apps and API, plus a Highspeed variant said to be 2-3x faster. Moonshot reports 65.7% on Terminal-Bench 4.0, near the top closed models, and a big drop in claim-level hallucination. It says full weights will be published by November 4.

Why it matters: A near-frontier coding and agent model that may soon be self-hostable; scores are vendor-reported, so test on your own tasks.

Big newsRead on The Next Web
Foundations & Models / Open-weight Models

Nanbeige4.3-33B ships under Apache 2.0 with a 512K context, aimed at agentic coding

Nanbeige LLM Lab put a dense 33B reasoning model on Hugging Face under Apache 2.0, with a context of about 524K tokens and XML/JSON tool calling. Its model card claims it beats Qwen3.8-27B on most agentic coding and automation tests, but trails on SWE-Bench Pro and GPQA Diamond. No one has independently reproduced the numbers yet.

Why it matters: A permissively licensed ~30B dense model pitched for coding agents gives self-hosters a fresh option to benchmark against Qwen and Gemma.

How to try it

Open BF16 weights on Hugging Face. The model card documents serving with vLLM, using Qwen3-style reasoning and tool-call parsers, which gives you an OpenAI-compatible endpoint.

  1. Download Nanbeige/Nanbeige4.3-33B from Hugging Face to a local path
  2. Set MODEL_PATH to that directory
  3. Run the vllm serve command below (it uses tensor parallel across 4 GPUs)
  4. Send OpenAI-compatible chat requests to port 8000. The card recommends temperature 1.0 and top_p 0.95 for agentic work
bash
MODEL_PATH=/path/to/your/Nanbeige4.3-33B
vllm serve ${MODEL_PATH} \
  --host 0.0.0.0 \
  --port 8000 \
  --tensor-parallel-size 4 \
  --gpu-memory-utilization 0.8 \
  --enable-auto-tool-choice \
  --reasoning-parser qwen3 \
  --tool-call-parser qwen3_coder
Open the docs
Read on Hugging Face (Nanbeige LLM Lab)
Foundations & Models / Small & On-device Models

NVIDIA releases Nemotron 3.5 Bolt E4B for local agents in about 4GB of VRAM

NVIDIA took Gemma 4 E4B and post-trained it with reinforcement learning alone, giving a 4B-effective model with 64K context tuned for tool calls and short conversational turns. Quantized builds fit in roughly 4GB, aimed at RTX cards and Jetson, and it ships under OpenMDW-1.1 on Hugging Face and Ollama. Qwen3.5-4B still beats it on some agent benchmarks.

Why it matters: An openly licensed tool-using model that runs on consumer GPUs and edge boards without cloud calls.

Read on NVIDIA Research