7 stories on Models.
Carbon-X is trained on DNA sequences and ordinary text, and its creators report a 66.7% average across five DNA benchmarks.
Why it matters: An open model for DNA analysis could lower the cost of genomic work for life-science teams, though the performance and cost claims come from its creators and need independent checks.
Users found a Mythos 5.5 mention in the HTML of Anthropic's site, but a commenter says the line is a tagline that covers all Mythos models.
Why it matters: Speculation about upcoming Mythos models matters for buyers planning cybersecurity and biology work, but the evidence is currently too thin to plan around.
Commentators say GPT-6.2 Astra is visible in Codex usage analytics and may release very soon, but the claim is unconfirmed.
Why it matters: A near-term model release could change model choices for teams building on OpenAI, so procurement plans may need a flexible review point.
OpenAI says GPT-6.1 Terra adds cheaper visual reasoning and computer use across Work, Codex and the API, at half the per-token price of GPT-5.6 Terra.
Why it matters: Teams paying per token for vision or computer-use workloads may see lower costs if they move from GPT-5.6 Terra to this model.
Mistral says Large 4 is available to download, reports better coding after more reinforcement learning, and cut API prices to $0.70 and $2.10 per million tokens.
Why it matters: Companies can self-host Large 4 for coding work, while hosted-model providers face a licensing requirement before offering it.
NVIDIA's open-weight Nemotron 3.5 Transcribe accepts an hour of audio, takes transcript instructions, and supports 40 languages, speaker labels and timestamps.
Why it matters: Companies needing multilingual transcription can self-host an open-weight option, which may reduce reliance on hosted speech APIs.
Liquid AI's LFM3 puts six open-weight models for text, vision, audio and decisions on one backbone, and its 1.7B text model fits in about 1 GB at four-bit precision.
Why it matters: Small, fine-tunable models that run on a single GPU or in limited memory lower the cost of deploying specialized AI on edge hardware.