NPDev News · Intelligence. In perspective.

AI News

NPDev Daily AI Briefing - November 7, 2026

Capital got colder and open weights got stronger in the same 48 hours: the Nasdaq's 5.5% slide and OpenAI's smaller-than-planned $11B close point to a repricing of AI capacity, while Kimi's weights reached Bedrock, Thinking Machines promised Apache 2.0 next week, and a $32M 65B model took the US open-weight coding crown.

Written by NPDev Labs with AI assistance and reviewed before publication. How we work.

AI trade cracks: Nasdaq sheds 5.5% and roughly $2.5T in one session

The AI trade cracked on Nov 5. The Nasdaq Composite fell 5.5% — its worst day in 19 months — and closed in correction territory, about 12% below its Oct 6 record, after Democrats' midterm wins raised the odds of data-center pushback. The Nasdaq-100 lost roughly $2.5T in market value, Nvidia fell 9%, AMD 17%, and all 11 S&P 500 sectors closed lower. Friday brought only a partial rebound, with the Nasdaq up 1.09%.

Why it matters: The buildout that sets GPU and inference prices is now politically exposed. UBS estimates about 30% of expected 2028 US data-center capacity additions could be at risk from construction pauses, so capacity planning deserves the same scrutiny as per-token pricing.

OpenAI closes $11B first tranche at $1.25T, well under its $30B ask

Bloomberg reported on Nov 6, and Reuters confirmed through a source, that about $11B was wired to OpenAI as the first close of a round valuing the company at $1.25T. The round may total about $20B, down from an earlier target of up to $30B at a $1.4–1.5T valuation. SoftBank and MGX participated; OpenAI declined to comment.

Why it matters: Even frontier labs are now pricing against a colder market. That tends to slow subsidy-heavy discounting and push more of the real serving cost onto customers, which is worth factoring into multi-year API commitments.

Anthropic opens a read-only Personal Finance beta in Claude

Anthropic began rolling out Personal Finance in beta on Nov 5 for US Pro, Max and Ultra subscribers. Claude links bank, card, loan and investment accounts through Plaid in read-only mode, builds categorized breakdowns from up to two years of transactions, and cannot move, manage or invest money.

Why it matters: The assistant is moving from generic advice to account-level context. That is the template most agentic products will follow, and it makes consent scope, read-only enforcement and audit trails the design questions rather than model quality.

A Claude Code convenience setting is leaking real credentials to public GitHub

Apiiro reported on Nov 5 that 1,122 of 104,185 public copies of .claude/settings.local.json contain at least one secret — 1,379 distinct secrets, including 379 database passwords — because choosing "Yes, and don't ask again" saves the full command verbatim. Anthropic's protection depends on a global git exclude on one machine, and Claude Code never adds the file to the project's own .gitignore. In 539 of 1,105 traced repositories, a commit listing Claude as co-author changed the file. gitleaks flagged 564 and TruffleHog 643 of the 1,122; 215 were caught by neither.

Why it matters: Audit your organization for this file, including deleted versions in git history, and rotate anything it touched. Prefer environment variables or a secrets manager, and approve command patterns rather than individual commands. Anthropic had not responded publicly as of Nov 7.

Kimi K3.1 Pro's open weights reach Amazon Bedrock

AWS said on Nov 6 that Kimi K3.1 Pro (Highspeed) is available on Amazon Bedrock through cross-Region inference, with prompt caching and the same security boundary as proprietary models. Moonshot published the checkpoint on Hugging Face on Nov 4: 2.8T total parameters, 104B active, 1M-token context, native image input, under a custom license. Microsoft Foundry lists $2.30 per million input and $11.50 per million output tokens.

Why it matters: Classify this as a new availability surface, not a new model. A frontier-class open-weight model inside a governed cloud boundary is a genuine option for teams that cannot send data to a consumer API, and the hosted price is a useful ceiling when you compare self-hosting.

Thinking Machines ships Inkling 2 Small, with Apache 2.0 weights promised next week

Released on Nov 5, Inkling 2 Small is a sparse MoE with 594B total and 17B active parameters, image input, variable reasoning effort and 1M-token context. It scores 31 on the Artificial Analysis Intelligence Index, up from 13 for its predecessor. It is available now on Tinker with closed weights; the lab says full weights follow under Apache 2.0 after external safety testing and red-teaming.

Why it matters: 17B active is cheap to serve and the vision-in-the-loop coding demos are the real pitch, but the lab's own tables put Terminal-Bench 4.0 at 13.1% — weak for terminal agents. Treat the open-weights date, not the index score, as the claim to verify.

A $32M, 65B dense model claims the US open-weight coding crown

Ricursive released Ripple-65B Preview under Apache 2.0 on Hugging Face: a 65B dense model pretrained on 12T tokens with up to 1M context. It reports 80.52 on TerminalBench 2.1, above Reflection Beam 501B (80.1) and well above Meta Muse Spark Open 816B (76.4), at a stated $32M total R&D cost, and says the model fits on a DGX Spark. It is served by Together, Fireworks, OpenRouter and GMI Cloud.

Why it matters: Benchmarks here are self-reported, including the Terminus 2 harness row at 80.5. Even discounted, a dense 65B at 4-bit (roughly 34GiB) is a credible local coding agent — and cost-per-capability, not parameter count, is now the open-weight contest.

Ant Group's Ling-3.2-flash adds native video at 560B total, 25B active

Ant Ling's developer changelog dates Ling-3.2-flash to Nov 5: 560B total parameters, about 25B active, a 1M-token window and native image, text and video input, with OpenAI- and Anthropic-compatible APIs and free access for the first two weeks. Weights are closed, unlike the MIT-licensed Ling-3.1-flash. Ant claims it surpasses Claude Opus 5 (max) on three release benchmarks — self-reported and not independently verified.

Why it matters: Another capable Chinese multimodal model at flash-tier activation cost. Because the API speaks Anthropic's protocol, swapping it behind an existing client is a small experiment; just do not repeat vendor benchmark numbers as measured results.

Labs to watch

Google shipped no new frontier model in this window: Gemini Omni 1.2 Flash went generally available on Nov 5 (covered in yesterday's briefing) and the Gemini Embedding 2 preview was deprecated on Nov 6. Meta made no model announcement; its only fresh item was a reported humanoid-robot prototype built with China's Fourier, which remains unconfirmed by Meta.

Bottom line

Capital got more expensive and open weights got cheaper in the same 48 hours. For engineers, that means planning for less aggressive API discounting and slower capacity growth while treating 65B-class open models as a real option for coding agents you can run yourself.

As of

November 7, 2026, 15:00 UTC