OpenAI, Anthropic and other AI firms roll out new multimodal models, agent tools and security‑focused deployments

OpenAI has expanded its product line with the launch of GPT‑Realtime‑2, the company’s most capable voice model to date. The new model achieves 96.6 % on Big Bench Audio and 96.1 % on Conversational Dynamics, and its context window has quadrupled from 32 K to 128 K tokens. GPT‑Realtime‑Translate now supports more than 70 input languages and 13 output languages, while GPT‑Realtime‑Whisper delivers low‑latency streaming transcription. The audio stack also tops Scale AI’s Audio MultiChallenge S2S leaderboard, and instruction retention has risen from 36.7 % to 70.8 % APR over the previous version. In addition, OpenAI released a Chrome plugin for Codex that allows the model to run across background tabs, use plugins where possible, and support multi‑tab parallelism. The company also announced GPT‑5.5 with Trusted Access for Cyber, previewing a GPT‑5.5‑Cyber tier for red‑team and penetration‑testing workflows. A new deployment unit, the Deployment Company, was unveiled with 150 engineers and a $4 B investment from 19 partners, aimed at embedding deployment services into enterprise workflows. Daybreak, a security‑focused model distribution, bundles GPT‑5.5, Codex, repository threat modelling, vulnerability discovery, patch generation and response automation, all under the Trusted Access for Cyber umbrella.

Anthropic introduced Natural Language Autoencoders (NLAs), a method that translates model activations into human‑readable text, and moved its Petri interactive behavioural‑evaluation tool to Meridian Labs while opening a public bug bounty on HackerOne. Goodfire launched a neural‑geometry research series arguing that networks think in manifolds, and Prime Intellect opened its Lab for reinforcement‑learning environments, where Ramp Labs trained a Fast Ask spreadsheet‑QA sub‑agent that outperforms Opus 4.6 at Haiku‑level latency. Nous/Teknium shipped Hermes Agent v0.13.0 with multi‑agent orchestration via Kanban, goal enforcement, disk‑usage optimisations and custom gateway channels. Cursor added an /orchestrate command and a PR‑review experience that cut skill‑token use by 20 % and cold‑start time by 80 %. LangGraph introduced delta channels for checkpoint diffs, and Deep Agents added sandbox backends with an auth‑proxy pattern to keep credentials out of prompt‑injectable sandboxes.

Thinking Machines unveiled full‑duplex multimodal interaction models that enable continuous audio, video and text to audio and text without turn‑based boundaries. Agent orchestration tools also advanced: aggit, a Rust CLI, offers S3‑backed artifact storage with stash, branch and restore functions; Deep Agents CLI can hot‑swap model providers mid‑conversation; LangChain added harness profiles for provider‑specific tuning; and Hugging Face released Hermes Agent support for local apps with native trace visualization. Cursor’s CLI now integrates with Microsoft Teams, reading full threads and opening PRs, while Claude agents gain a terminal control plane. The Coding Agent Index benchmark shows Opus 4.7 scoring 61 in the Cursor CLI, with GPT‑5.5 close behind; open‑weight models such as GLM‑5.1 and DeepSeek V4 Pro remain competitive but lag. Cost per task varies by a factor of 30, token usage by three, cache hit rates range from 80 % to 96 %, and time per task can be seven times longer, underscoring the importance of harness selection. OpenHands released an updated software‑engineering benchmark, highlighting the growing emphasis on measurable agent performance.