GPT-6.1 Terra surfaces in Microsoft evals

// 1h agoNEWS

GPT-6.1 Terra surfaces in Microsoft evals

A merged EF Core pull request replaces GPT-5.6 Terra with GPT-6.1 Terra as the judge model across 18 coding-evaluation configs. The change also moves task execution from GPT-5.6 Sol to GPT-6.1 Sol, suggesting pre-release testing rather than a public launch.

// ANALYSIS

This looks like credible pre-release evaluation plumbing, but the config is not proof that GPT-6.1 Terra has shipped.

–Seventeen existing configs migrated, while a new query-pipeline evaluation brings the total to 18
–Terra judges coding results while Sol performs the coding tasks, revealing a deliberate model-role split
–Medium reasoning effort and a ten-minute timeout are evaluation settings, not published model specifications
–OpenAI’s public model catalog still lists GPT-6.1 Sol, not GPT-6.1 Terra
–Developers should treat this as a strong availability signal, not an API or pricing announcement

// TAGS

gpt-6.1-terrallmevaluationbenchmarkai-codingcoding-agent

DISCOVERED

1h ago

2026-10-20

PUBLISHED

1h ago

2026-10-20

RELEVANCE

9/ 10

AUTHOR

mark_k

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

VIDEO50m ago

Orca Brings Agent Fleets Into One Workspace

Orca is an open-source agent development environment for running Claude Code, Codex, OpenCode, Pi, and other CLI agents in parallel Git worktrees across desktop, remote, and mobile environments. Its control-plane approach centralizes monitoring, diff review, and agent steering without bundling proprietary models.

MODEL1h ago

Xing4.0-29B-Intent Ships Unified JSON Extraction

China Telecom’s XingChen AGI Lab released a 29B MoE model with 4B active parameters that jointly recognizes intents and extracts structured arguments into JSON. The Apache-2.0 release includes a 2.75-million-sample dataset and inference SDK.

INFRA1h ago

Qwen Coder Gets AMD NPU Wrapper

Modelers.cn has listed AMD’s Qwen2.5-Coder-7B-Instruct variant, quantized for Ryzen AI NPU deployment with Token Fusion and a 16K context window. It is a hardware-targeted repackaging of an existing open model, not a new frontier release, and publishes no benchmark scores.