Beam activates 23B of its 501B parameters per token, has 1M-token context, and Reflection reports 80.9 on SWE-bench Verified and 80.1 on Terminal Bench 2.1, roughly level with GLM 5.2 at 3–4x less inference compute but behind Kimi K3 and GLM 5.3. It's waitlist-only for now, with weights, a technical report, and fine-tuning tooling promised later in October.
Instead of writing out an answer, the Apache 2.0 model returns probabilities for answers you define (like safe and unsafe ) for a prompt, tool call, code change, or alert, so your own code picks a threshold to allow, block, or escalate; Superagent reports catching 599 of 600 BIPIA prompt injections. Prompt injection screening is the best tested use so far, it flagged 12% of tricky benign NotInject inputs, and the numbers are Superagent's own, so tune it on your own traffic and keep real permission checks outside the model…
The first drop of OpenAI's 28 day improvement or reset run is an inference speedup of roughly 50% for GPT 6 Astra and GPT 6.1 Sol when used through a ChatGPT subscription, including partner tools that use Sign in with ChatGPT like OpenCode, Pi, Amp, and Devin, with nothing to change on your end. API speeds aren't part of it, and one early forum benchmark saw 6.1 Sol get slower mid rollout, so check your own runs…
As of Oct 5, secret scanning flags lovable api key , Supabase OAuth and scoped personal access tokens, and Pydantic Logfire and AI Gateway keys in your repos, and Lovable joined the partner program, so a Lovable key pushed to a public repo gets sent to Lovable to revoke or rotate. If your agent ever committed a .env , check your repo's security alerts…
OpenAI's Thibault Sottiaux posted that for the next 28 days the Codex and ChatGPT Work team will each day either ship one improvement relevant to most users or hand out a full usage reset. It's an either/or rather than 28 guaranteed resets, and he hasn't said how a reset would work or which plans it covers…
AgentField's Apache 2.0 CodeAF ships as one 53 MB binary ( curl fsSL https://agentfield.ai/get/codeaf | bash ) that turns each request into its own task on its own branch, runs it on a per task crew of open models via OpenRouter, provider keys, or local Ollama, and claims first place among ten harnesses on DeepSWE with DeepSeek V4 Flash. It's an early preview with anonymous telemetry on by default, which CODEAF TELEMETRY=off or DO NOT TRACK=1 turns off…
Spotted something we missed? Start a thread.