NPDev Daily AI Briefing - November 1, 2026

OpenAI will merge ChatGPT and Codex into one token allowance in December

OpenAI's Help Center article “Understanding ChatGPT and Codex usage limits” was updated on October 31 with a short policy change: starting in December, Plus, Pro and Business plans get a single usage allowance shared across ChatGPT and Codex, measured in tokens, as the agentic capabilities of Work move into the main ChatGPT experience. Instant chat does not draw on it. The article gives no date beyond “December”, no allowance size for any plan and no conversion rule, and says users will see combined remaining usage and reset times in Settings → Usage, with waiting, upgrading or buying credits as the options at the cap. There was no blog post or API changelog entry — the only other signal in the window is the October 29 release note that doubled the Work and Codex allowance on Pro 100 and Pro 500 and cancelled the previously announced Pro 200 reduction.

Why it matters: Budget consequences hit developers first. If you run Codex on a personal or Business seat, agent work now competes with chat for one pool whose size OpenAI has not published. Measure token burn per task type before December so a plan change is a decision rather than a surprise, and keep an API key as the fallback for work that must not be throttled mid-run.

Sources:

Banks and retailers are blocking AI agents — and 37 of 51 do it by accident

CNBC reported on November 1 that JPMorgan Chase, Charles Schwab, Amazon, Neiman Marcus and Kohl's have moved to block personal agents such as Meta's Muse, with Jamie Dimon saying the bank blocks external agents acting for retail customers on security grounds, while Walmart went the other way and published terms admitting agents at the same prices humans see. An audit of 51 large US retailers' published access rules by Billions Network, an identity-verification company, found that only 11 have taken a deliberate position: five, including Amazon, Kohl's and Neiman Marcus, explicitly restrict agents, and 37 block shopping agents largely by accident through older bot rules written for scrapers. None offers a path for a verified agent to complete a purchase. CNBC also cites Accenture research that 71% of customers would welcome an assistant like Muse inside their primary bank's app.

Why it matters: User-agent rules and WAF configs have quietly become a product decision, and most sites have already made one without noticing. If you operate a site, audit bot rules for legitimate agent traffic before a competitor advertises that it accepts agents. If you build agents that transact, assume checkout is blocked and design for a graceful handoff plus an identity story — signed agent identity, spend limits, auditable consent — rather than for stealth.

Sources:

VOLTA infostealer routes around refusals with Claude, Kimi and Gemini

Cisco Talos published an analysis on October 31 of VOLTA-Autodoxer, a self-deleting Windows infostealer whose optional AI stage turns a stolen credential log into a ranked dossier. After harvesting browser credentials, cookies, cards, wallet seed phrases, Discord, Telegram, Steam and Roblox sessions, Microsoft OAuth tokens, SSH and cloud configs and every saved Wi-Fi password via WlanGetProfileList, it enriches the bundle using the victim's own sessions — searching Gmail for Amazon, DoorDash, Uber Eats and PayPal receipts to confirm a home address, pulling order history, valuing Roblox inventory through a live Rolimons lookup — writes Identity_Report.txt and POSTs it to OpenRouter's chat-completions endpoint with a hand-assembled JSON body and a generic Mozilla user agent. The hardcoded default is anthropic/claude-opus-4.6 at temperature 0.2 under a “forensic intelligence analyst” system prompt asking for “WHAT THEY'RE WORTH” and “THEIR OPSEC” sections. If the reply contains phrases such as “I can't” or “I'm not able”, it fails over to moonshotai/kimi-k2 and then google/gemini-2.5-flash, and saves the answer as AI_Analysis.txt. Talos found no attacker-owned infrastructure, did not track real-world use, and does not claim any provider's safeguards failed.

Why it matters: The interesting part is not a jailbreak — it is failover. Refusal-based safety is now a routing problem an operator solves in a few dozen lines, and every endpoint in the chain is a legitimate third-party API billed to someone's key. Practical responses: alert on headless Chrome launched by an unexpected parent process (Talos says an application-control policy on that path would block the App-Bound Encryption bypass), treat model-API egress as a detection surface, and after a hit assume every session token on the box is burned and every plaintext seed phrase is gone.

Sources:

Intel's worst day since the 1970s drags the AI trade down

Intel fell 29% on Friday, October 30 — its worst session since the 1970s — after guidance came in below expectations with the company citing a plant shutdown in Israel and continued PC-market softness. The drop took roughly $12 billion off the value of the US government's 10% stake and about $6 billion off Nvidia's position; Nvidia itself fell 4%, Broadcom more than 5% and Qualcomm 11%. The Nasdaq Composite slid 2.9%, capping a week in which Alphabet had its worst day on record, down more than 12%, and Amazon held 2026 capex at $220 billion while posting a second consecutive quarter of negative free cash flow. D.A. Davidson downgraded Intel to neutral on Friday under the title “Not fixed yet”.

Why it matters: Two repricings in one number. Intel had run up about 150% this year on demand for CPUs in agentic workloads, and that thesis just lost its guidance support; separately, the financing chain behind AI capacity is being re-rated week by week. The second-order effect matters more than the ticker: memory and CPU pricing, lead times and reserved-capacity terms get harder, so stress-test the assumption that 2027 inference is cheaper per token in the plan you are about to submit.

Sources:

DeepSeek open-sources a coding agent built to need almost no supervision

DeepSeek published dsh-superagent-prototype on GitHub on October 31: MIT-licensed TypeScript, roughly 6,000 stars and 570 forks within a day, and labelled in its own README as a research prototype rather than an official product. It is a bundle for DeepSeek Harness that watches one GitHub repository and, optionally, Feishu. When an authorized person asks for work, it opens a work item, runs that item in its own agent session and directory, and answers in the original conversation — typically with a pull request — while a web panel shows what needs a human. The team's stated principle is “take as little human attention as possible, and give as much value as possible”: the agent acts only when asked, or when a rule the team wrote allows it, and speaks only when someone must act or can use the result. It accompanies a design paper, “Exploring Super-agent Design: Towards a More Autonomous Agent System”, and publishes no benchmark results.

Why it matters: Read it as an architecture proposal, not a tool to deploy. Authorization rules, per-work-item session and directory isolation, and the constraint that the agent speaks only when a human must act are the parts worth copying into your own stack. The caveats are in the README, in DeepSeek's own words: the public code is not identical to the internal version, and some designs in the paper are not implemented.

Sources:

Mistral's trillion-parameter Le Chonk is downloadable, with a licence catch

Mistral released the weights of Mistral Large 4, “Le Chonk”, on October 29 — three weeks after its API preview and after additional reinforcement learning that it says improved coding, defensive cybersecurity and reasoning efficiency. It is a 1-trillion-parameter natively multimodal mixture of experts with about 49 billion active parameters per token, 52 billion including embeddings, and a 1M-token context window, trained from scratch on 3,800 Nvidia Grace Blackwell GPUs in Mistral's European datacenters and natively fluent in more than 160 languages. FP8 and NVFP4 checkpoints ship with an embedded DSpark speculative-decoding head, vLLM and SGLang are supported, and community GGUFs already exist. API pricing is $0.70 per million input tokens, $0.07 for cached input and $2.10 for output. The licence is a modified MIT: free to use, modify, fine-tune and commercialise, but running a model-as-a-service business requires a separate commercial agreement. The headline benchmark, 19.7% on Cognition's private FrontierCode 2.0, is a vendor-selected result.

Why it matters: This is the most capable model most teams can actually run on their own hardware today, and the licence carve-out decides whether you can build a product on it. Self-hosting a 1T mixture of experts is a multi-node exercise: plan for NVFP4 on Blackwell or FP8 elsewhere, use the speculative head, and re-run your own evals instead of importing the vendor score. If you resell inference, read the licence before you size the cluster.

Sources:

Release ledger: no new frontier checkpoint this weekend

Checked directly against first-party surfaces on November 1. OpenAI's API changelog still ends at GPT-6 Rosalind on October 30, a trusted-access life-sciences family. Anthropic's newsroom ends at the October 28 AI for Science expansion, its API release notes at the October 30 Opus 4.8 fast-mode deprecation, and Claude Opus 5.5 from September 22 remains the flagship. Google's Gemini API changelog ends at Computer Use reaching general availability on October 30, with Gemini 4 Argon from October 13 as the newest model. Meta published no new model page. Moonshot's Kimi K3, the 2.8-trillion-parameter open-weight model, is unchanged since the summer. No new frontier weights appeared on Hugging Face from any of the major labs in the window.

Why it matters: A quiet weekend is information, not a gap. The week's movement was in packaging and access — quotas, licences, bot rules, agent harnesses — rather than checkpoints, and that is where the engineering decisions are this week. For launch tracking, the durable triggers are the vendor changelogs and model catalogs linked here, not aggregator headlines.

Sources:

Bottom line

Capability did not move this weekend; the metering and permission layer did. One shared token budget for chat and coding agents, bot walls in front of checkout, a malware pipeline that treats a refusal as a routing condition, and an open-source agent whose selling point is that it does not need you. Plan for all four: per-task token accounting, an agent-identity story for anything that touches a third party, egress monitoring that covers model APIs, and evaluation runs you own rather than benchmarks you inherit.

As of November 1, 2026, 15:00 UTC. Researched and generated for NPDev News subscribers.