DeAI Daily Brief — 12 November 2026

Today in DeAI: OpenAI pools ChatGPT and Codex usage into one token allowance, Akash begins moving AKT to Solana after a 97% governance vote, and Meituan publishes full LongCat-2.5 weights under MIT.

OpenAI moves Plus, Pro, and Business to a shared token pool across ChatGPT and Codex

Starting in December, usage on OpenAI's Plus, Pro, and Business plans will be measured in tokens and drawn from one shared allowance across both ChatGPT and Codex, per OpenAI's help center. Instant chat is excluded, and the same month six older GPT-5 and o3 snapshots leave the API on December 11. Why it matters: coding-agent traffic now draws down the same pool as chat, so budgeting a subscription in sessions rather than tokens no longer works. Full detail in our coverage.

Akash's AKT token migration to Solana is underway

Akash Network's Proposal 346 passed on November 2 with 97.3% of roughly 61.66M AKT cast in favor, per Solana Compass, and the SPL token and one-way migration portal are going live per Akash's blog. Precision worth keeping: Prop 346 authorizes the token migration only, as moving the marketplace itself requires a second cutover vote, and live compute leases continue on akashnet-2 until then. Why it matters: a major decentralized compute network is betting settlement needs high-frequency, low-cost transactions more than app-chain sovereignty.

Meituan publishes full LongCat-2.5 weights under MIT

The weights repo is live on Hugging Face: a 1.6T-parameter native multimodal mixture-of-experts model with about 48B active parameters, licensed MIT, one week after the model was introduced. Benchmark figures circulating with the release are Meituan's own self-evals until independent reruns land. Why it matters: near-frontier agentic capability with a fully permissive license, though at 1.6T total parameters it is a hosted-API or datacenter-scale decision, not a GPU-box one.

NVIDIA open-weights Alpamayo 2 Edge for onboard driving compute

NVIDIA released Alpamayo 2 Edge, a 2B-parameter vision-language-action reasoning model for real-time operation on NVIDIA DRIVE, per the NVIDIA developer blog, with weights on Hugging Face under OpenMDW-1.1 and Apache-2.0 code. Why it matters: the interesting shape for inference builders is license-plus-latency, not driving: a 2B reasoning model targeting hard real-time bounds on fixed hardware is the same trade-space as on-device inference generally.

Google ships Gemini Decision 1 and a Decisions API

Google's new model scores up to 1,000 closed-ended questions in a single pass at a claimed 150 to 250 ms, per Simon Willison's first-hand writeup; the latency and throughput figures are Google's own, verifiable by anyone with an API key. Why it matters: the decision-model category we tracked as a startup niche, with Fastino and open Jev replicas, now has a hyperscaler product, and bulk classification pricing is a signal for the most common agentic workload: routing, moderation, and gating.

Watching tomorrow

Whether OpenAI publishes the actual allowance sizes for each plan tier behind the December shared pool, since that single number decides whether Codex-heavy workflows fit.