Ai2's StateBench cuts frontier models to 13% on long-horizon state
A reality check for anyone building long-horizon agents — plus ElevenLabs voice agents land in four contact centers and OpenAI merges Codex quotas

Copy markdown
Frontier models fall to 13% tracking long-horizon state
Ai2's new StateBench streams long runs of variable updates and asks a model for each final value: Gemini 3.8 Flash, Claude Opus 5.5 and GPT-6 Sol/Luna nail the easy cases but crater to 13% or lower on the longest 'general finite' problems (assignment by reference and swaps). The takeaway for agent builders — don't trust a transformer to hold evolving state over a long run — and the twist: tiny recurrent/hybrid models with Gated DeltaNet generalized near-perfectly to ~10x their training length. Dataset and code are open.
ElevenLabs drops voice agents into four contact centers
ElevenAgents added native connectors for NiCE CXone, Genesys Cloud, Amazon Connect and Five9, letting you deploy a voice agent straight into an existing call-center stack and hand off to a human when needed. If you've been eyeing support or IVR automation, that's the integration plumbing you no longer have to build yourself.
Heads-up: ChatGPT and Codex quotas merge in December
OpenAI is folding ChatGPT and Codex usage into a single shared token allowance across Plus, Pro and Business starting in December, with instant chat exempted from the pool. If you lean on Codex for real work, your heavy coding runs will now draw from the same budget as everything else — plan for it before the switch.