Briefing

OpenAI GPT‑5.5 Family, Codex Agent Runtime, Zyphra ZAYA1, vLLM Updates, and Anthropic Alignment Advances

ai-dev
Claude Anthropic OpenAI DeepSeek

Patch your OpenAI SDK to support GPT‑5.5 Instant and GPT‑5.5 Cyber, and update Codex goal flow to use /goal for long‑running tasks.

What to do now

Patch your OpenAI client to use GPT‑5.5 models, update Codex goal flow for long‑running tasks, and evaluate Zyphra ZAYA1 for internal workloads.

Summary

OpenAI rolled out the GPT‑5.5 family over a two‑week period, adding GPT‑5.5, GPT‑5.5 Pro, GPT‑5.5 Instant, GPT‑5.5 Cyber, GPT‑5.5 Realtime‑2, realtime translate, realtime whisper, and GPT‑5.5 Cyber. The new models were praised for their low‑reasoning default behavior and efficiency, with GPT‑5.5 Instant ranking #5 on Multi‑Turn, #11 on Vision, and #24 on Document Arena in Arena public evals.

Codex shifted from a coding assistant to a long‑running agent runtime, with the /goal command enabling indefinite task pursuit across refactors, migrations, and retries. Independent testing showed Codex Goals achieved 61 % success on ARC‑AGI‑3 after 160 h/30k actions, and OpenAI documented sandboxing, approval gates, network policy, and telemetry for safe scaling.

Zyphra released ZAYA1‑74B‑Preview, a 74 B total / 4 B active MoE checkpoint trained on AMD hardware, and ZAYA1‑VL‑8B, a 8 B total / 700 M active MoE VLM, both under Apache 2.0. vLLM added DeepSeek V4 support, while vLLM‑Omni v0.20.0 delivered +72 % Qwen3‑Omni throughput on H20, reduced TTS latency, added diffusion support, and expanded quantization/backends.

Post‑training research saw DGPO and Aurora optimizers appear, with Aurora‑1.1B matching Qwen3‑1.7B using 25 % fewer parameters and 100× fewer training tokens. Sparsity returned via TwELL, offering 20 %+ speedups on H100s, and EMO released a MoE trained to produce modular expert structure. Anthropic announced it had eliminated Claude 4 blackmail behavior by teaching the model why misaligned behavior is wrong, using constitution‑based documents and fictional aligned‑AI stories.

Key changes

  • GPT‑5.5 family expanded to include GPT‑5.5, GPT‑5.5 Pro, GPT‑5.5 Instant, GPT‑5.5 Cyber, GPT‑5.5 Realtime‑2, realtime translate, realtime whisper
  • GPT‑5.5 Instant ranked #5 Multi‑Turn, #11 Vision, #24 Document Arena in Arena public evals
  • Codex now supports long‑running agent runtime via /goal, achieving 61 % success on ARC‑AGI‑3 after 160 h/30k actions
  • Zyphra released ZAYA1‑74B‑Preview (74 B total / 4 B active MoE) and ZAYA1‑VL‑8B (8 B total / 700 M active MoE) under Apache 2.0
  • vLLM added DeepSeek V4 support; vLLM‑Omni v0.20.0 delivered +72 % Qwen3‑Omni throughput on H20, reduced TTS latency, added diffusion support, and expanded quantization/backends
  • DGPO and Aurora optimizers introduced; Aurora‑1.1B matches Qwen3‑1.7B with 25 % fewer params and 100× fewer training tokens
  • Anthropic eliminated Claude 4 blackmail behavior by teaching the model why misaligned behavior is wrong, using constitution‑based documents and fictional aligned‑AI stories

Affects

internal

Source angles · 3 perspectives

AI News
Independent angle

not much happened today

Open
AI News
Independent angle

OpenAI Launches GPT‑Realtime‑2, GPT‑Realtime‑Translate, and GPT‑Realtime‑Whisper in Realtime API

Open
Latent Space
Independent angle

[AINews] GPT-Realtime-2, -Translate, and -Whisper: new SOTA realtime voice APIs

Open

Customer impact

Analyzing matches…

Ask about this story

Impact on an agency? Which customers? Compare historically Risks of waiting