NPDev Daily AI Briefing - November 2, 2026

OpenAI is merging ChatGPT's Chat and Work tabs, and confirming one shared token pool

TechCrunch reported on November 2 an email OpenAI sent to users: the Chat and Work tabs inside ChatGPT are being merged into one surface, first for select users to gather feedback, then more widely. The stated pitch is continuity — bring files and plugins, ask ChatGPT to build a website or a presentation or act on the web, and review its work in the same place. Codex stays a separate experience in the desktop app. The same email restates the billing change: ChatGPT and Codex draw tokens from one allowance pool, which OpenAI says will not affect most users but may mean some wait for resets more often. Anthropic merged Claude's chat and Cowork screens in September.

Why it matters: This is the mechanism behind the Help Center note that landed on October 31: one pool, no published size, and no conversion rule between chat and agent work. If you run Codex against a subscription seat, instrument tokens per task type this month so a December plan change is a decision rather than a surprise, and keep an API key for runs that must not stop at a cap.

Sources:

OpenAI trims its round to $20 billion at $1.25 trillion as Anthropic's IPO nears pricing

The Financial Times reported on October 30, and Reuters and CNBC carried on October 31, that OpenAI cut its current fundraising target by $10 billion to about $20 billion and lowered its desired valuation to $1.25 trillion from $1.4 trillion. Friday was the deadline for commitments; the first close is expected to raise less than $20 billion after some investors reduced allocations. A UAE consortium including MGX is expected to be the largest participant, alongside Founders Fund, Coatue, Sequoia and Andreessen Horowitz, with a later retail offering at the same $1.25 trillion. Separately, Anthropic publicly filed its S-1 on October 26 for a Nasdaq listing under ANT; WSJ and Reuters reporting cited by Yahoo Finance points to a debut in the week of November 16, with marketing possibly starting November 9. TechCrunch also relayed an FT report that Anthropic expects a high-yield rating from credit agencies.

Why it matters: Both labs are raising into the same investor pool during a selloff, and the price of their API tiers in 2027 is downstream of how these deals clear. The practical read for a platform team is not the valuation but the discounting: assume list prices stay competitive through the IPO window, and do not sign a multi-year capacity commitment on the assumption that today's per-token price is a floor.

Sources:

S&P puts Meta's AI payback on watch with a $132 billion three-year cash forecast

S&P Global Ratings revised Meta's outlook to negative from stable after the close on October 30, keeping the AA- rating. Its forecast is far harsher than the Street's: free operating cash flow deficits of $22 billion in 2026, $71 billion in 2027 and $39 billion in 2028 — $132 billion over three years, against consensus of about $61 billion. Meta's third-quarter free cash flow was negative $1.16 billion on the company's headline measure and negative $2.12 billion including finance leases, with capex of $33.94 billion, up 80% year over year. S&P's objection is payback: it says Meta's ability to monetize AI is more limited than peers that lease out compute. Meta is lining up a stock sale of up to $95 billion and disputes the assessment; its CFO has said AI lab investments will take longer to return than the core business. Alphabet's quarter shows the contrast: negative $9.1 billion of free cash flow, but Google Cloud revenue more than doubled to $30.7 billion.

Why it matters: This is the clearest statement yet that the constraint on AI capacity is underwriting, not demand. If credit and equity for compute tighten, reserved-capacity terms, memory pricing and multi-year discounts move against buyers. For planning, treat 2027 unit costs as uncertain in both directions and keep an exit ramp in any commitment longer than twelve months.

Sources:

Meituan's LongCat-2.5 arrives as a 1.6-trillion-parameter model with no weights

Meituan's LongCat team published LongCat-2.5 on November 1, a native multimodal model aimed at long-horizon agentic work rather than single-file coding: 1.6 trillion total parameters with roughly 48 billion activated per token, the same totals as LongCat-2.0, with a rebuilt emphasis on visual understanding and agentic coding across codebases, terminals and browsers. The page's own metadata says training ran entirely on domestic (Chinese) accelerators. It is available through Meituan's chat service and API platform; launch coverage says the API speaks OpenAI- and Anthropic-compatible protocols and that GitHub and Hugging Face releases are planned to follow. As of this check there is no LongCat-2.5 repository in Meituan's GitHub organization and no LongCat-2.5 weights on its Hugging Face organization. The release table, as relayed by launch coverage, reports Terminal-Bench v4.0 at 38.5 (from 3.0 for LongCat-2.0) and DeepSWE v1.1 at 64.6 (from 9.7) — still behind Claude Opus 5 at 53.9 on Terminal-Bench v4.0.

Why it matters: Two things matter here. First, a Chinese lab is now shipping frontier-class agentic coding only through an API, reversing the open-weights pattern that made LongCat-2.0 notable — so there is nothing to audit or self-host this week. Second, if the endpoint really is Anthropic-compatible, evaluating it is a base-URL change in Claude Code-style harnesses. Run your own tasks; a vendor table with no downloadable weights is a claim, not a measurement.

Sources:

Release ledger: no new frontier checkpoint in the window, and one model still behind a gate

Checked directly against first-party surfaces on November 2. OpenAI's model release notes end with the October 29 Model Spec revision and the October 26 GPT-6.1 Terra launch; GPT-6 Rosalind from October 30 remains trusted access for approved life-sciences organizations. Anthropic's platform release notes end on October 30 with the Opus 4.8 fast-mode deprecation, its newsroom's newest article is October 28, and Claude Opus 5.5 from September 22 is still the flagship. Google's Gemini API changelog ends on October 30 with Computer Use reaching general availability; the newest model there is Gemini 4 Argon from October 13. Gemini 4 Carbon, announced October 28, is still limited to a subset of cyber defenders through the Fairwind program, with developers, businesses and consumers promised access “as soon as possible” and no date, while Google cites pre-release government testing and unfinished guardrails. Google's own post claims Carbon ranks first on the Artificial Analysis Intelligence Index v5.0 at 59. Moonshot's newest is Kimi K3.1 Pro from October 24, a 3-trillion-parameter-class model with a 1M-token context, available through the app and API only; its last open weights are Kimi K3 from June. No new frontier weights appeared from major labs on Hugging Face in the window.

Why it matters: A quiet checkpoint week is information. It means the movement is in metering, pricing and gated access — and that the models you can download have not changed. For tracking, the durable triggers are the vendor changelogs linked here. Note the distinction on Carbon: limited access to vetted defenders is not general availability, and the absence of a date is a statement, not a scheduling hint.

Sources:

Copilot Fusion makes model routing somebody else's decision - and your eval problem

GitHub made Copilot Fusion generally available on October 28 across VS Code, the Copilot app, the CLI and JetBrains: instead of choosing a model, you choose Fusion and an effort level (Standard, High, Max) and Copilot routes each task, orchestrating models in single-model, cascade or critique patterns. GitHub says Fusion Max scored 60 on the Artificial Analysis Coding Agent Index, within two points of Claude Code with Opus 5.5 Max, at about half the average API cost per task, about half the tokens and roughly 40% less time per task. Billing is per model at each model's standard rate, with a temporary 30% discount on model costs for paid plans. Administrators can restrict which models Fusion may use and make it the enterprise default.

Why it matters: Routing is now the product, so your evaluation has to measure the router rather than a model. Two consequences: per-task cost attribution gets harder because one task may bill several models, and the benchmark numbers are vendor-selected on a single harness. If you enable it, log which model served each request and re-run your own suite before you let an admin set it as the default for everyone.

Sources:

Bottom line

Capital, metering and routing are three versions of the same question: who pays for the tokens. OpenAI's trimmed round and S&P's warning on Meta are the market telling the 2027 capacity bill that it needs a payer with a shorter payback; OpenAI's merged tabs and GitHub's routing default are the vendors telling you that payer is metered per token, at whichever model clears the quality bar. Three things follow for this quarter: measure token burn per task type before allowances change, treat router and agent benchmarks as vendor claims until you reproduce them, and do not plan 2027 margins on the assumption that inference keeps getting cheaper.

As of November 2, 2026, 15:00 UTC. Researched and generated for NPDev News subscribers.