AI Daily Brief · Nov 6, 2026 | Claude Personal Finance; GPT-6 Cyber; METR on Anthropic R&D speedup; Gemini Omni 1.2; Cursor live apps

Over the past 24 hours, consumer features and security models landed together. Anthropic opened a U.S. Personal Finance beta for Pro, Max, and Ultra, connecting banks and cards through Plaid with read-only access. OpenAI formally launched GPT-6 Cyber—ExploitGym at 48.5%, for select Daybreak Red customers—continuing the Linux kernel research thread. Independent lab METR estimated that Anthropic’s pace of capability gains on well-specified tasks is about twice its prior period, with much of the speedup likely from AI-assisted R&D. Google released Gemini Omni 1.2 Flash with up to 15-second video and a claimed ~30% price cut. Cursor shipped describe-to-deploy live apps with shareable links; Thinking Machines introduced Inkling 2 Small. On the agent side, Perplexity launched Teammates, a Claude Code email-in-User-Agent issue hit Hacker News, Antigravity refreshed its UI while remote agents keep rolling out, and Skills/orchestrator repos dominated GitHub Trending.

Today’s highlights

• Claude Personal Finance: U.S. beta, Plaid read-only account linking, Pro / Max / Ultra [Company claim].
• GPT-6 Cyber: ExploitGym 48.5%, select Daybreak Red access; OpenAI says 81 previously unknown high-severity Firefox bugs [Company claim].
• METR: Anthropic’s well-specified-task progress ~2× prior pace; AECI ~1.9×; much of it likely AI-assisted R&D [Third-party assessment].
• Gemini Omni 1.2 Flash: up to ~15s video, ~30% cheaper vs 1.1 Flash, about $0.07/s at 720p [Company claim].
• Cursor: paid plans can describe a tool, then build, check, and deploy a shareable live app [Company claim].

1. Anthropic launches Personal Finance in Claude (U.S. beta, read-only bank links) [Company claim]

Product: Anthropic says Personal Finance is rolling out in beta in the United States. Users connect accounts they choose and ask money questions in everyday language from their own balances and transactions. Access is read-only; Claude cannot move, manage, or invest money. The company says there are no ads or sponsored recommendations, and the feature is not designed to give investment advice. It is rolling out gradually on Pro, Max, and Ultra in the U.S.

Privacy: Linking goes through Plaid. Conversations in the Personal Finance hub are not used for training. Accounts can be unlinked anytime. TechCrunch, The Verge, and Engadget also covered the read-only limits and plan gating.

Sources: @claudeai 1 (~36.6K likes, ~7.0M views) · @claudeai 2 · Introducing Personal Finance in Claude · Claude Help · TechCrunch · The Verge

2. OpenAI launches GPT-6 Cyber: ExploitGym 48.5%, select Daybreak Red access [Company claim]

Capabilities: OpenAI says GPT-6 Cyber is built for authorized security research. Versus GPT-5.6 Cyber, ExploitGym rises from 32.1% to 48.5% (about a 51% relative gain). In an alignment eval, the model avoided exploiting out-of-scope services in 99.93% of tests without its deployment prompt, and 100% with that prompt. It powered the recently disclosed Linux kernel work and, OpenAI says, helped find 81 previously unknown high-severity Firefox vulnerabilities. Security teams at Mozilla, Palo Alto Networks, Calif, Cisco, and OpenAI are already using it. Access starts with select Daybreak Red customers and expands over time.

Also: @gdb highlighted training to stay inside the operator’s specified scope (~1,247 likes) and announced a strategic partnership with JPMorganChase on enterprise engineering, security, and customer use cases (~672 likes). A Cyber at the Frontier livestream is set for Nov 12.

Sources: @OpenAI 1 (~18.7K likes) · @OpenAI 2 · @OpenAI 3 · @gdb 1 · @gdb 2 · OpenAI: Introducing GPT-6 Cyber · The Verge · System Card

3. METR: Anthropic’s well-specified-task progress roughly 2× prior pace [Third-party assessment]

Findings: From Aug 12 to Oct 29, 2026, METR ran a high-access evaluation at Anthropic. Subjectively, as of October 2026, Anthropic is likely improving capabilities on well-specified tasks at about twice the rate seen from March 2024 to October 2025 (rough uncertainty band ~1.3×–3×). Empirically, AECI growth since October 2025 is about 1.9× (95% CI ~1.5–2.8×). METR is confident the acceleration was not driven by faster training-compute scale-up; interviews and surveys are consistent with AI use roughly doubling progress; it did not see clear bottlenecks that would halt further rapid progress in the next 12 months (many details redacted by Anthropic).

Reactions: @RyanGreenblatt called it a major update toward ASI within 1–2 years and said he previously under-estimated RSI speedup (~1,149 likes) [Opinion]. Community accounts relayed METR’s figures and said Sam McCandlish agrees with much of the picture, with Anthropic planning its own progress report soon [Media / secondhand].

Sources: @RyanGreenblatt · @kimmonismus 1 · @kimmonismus 2 · @kimmonismus 3 · @testingcatalog · @dotey · METR assessment · AP · The Verge

4. Google ships Gemini Omni 1.2 Flash: up to 15-second video, ~30% cheaper [Company claim]

Product: Google says Gemini Omni 1.2 Flash supports native video generation and editing up to about 15 seconds, multi-turn conversational edits, stronger reference adherence, and more natural speech. It claims pricing is about 30% lower than Omni 1.1 Flash, at roughly $0.07 per second for 720p. Rollout spans the Gemini API, AI Studio, Cloud, and the Gemini app (Ultra subscribers can use the full 15-second length).

Sources: @GoogleAI · @OfficialLoganK (~2,253 likes) · @GoogleAIStudio · @GeminiApp · Google Blog · Google Cloud Blog · The Verge

5. Cursor: describe a tool, get a live shareable app [Company claim]

Feature: Cursor says users can describe the tool they want; Cursor builds it, sets up a database, runs deterministic and agentic security checks, and deploys with a shareable link. Internal teams already use it for dashboards, customer decks, ops tools, onboarding, and developer portals. Available on paid plans.

Sources: @cursor_ai 1 (~3,120 likes) · @cursor_ai 2 · Cursor Changelog

6. Thinking Machines releases Inkling 2 Small; AA Intelligence Index score 31 [Company claim]

Specs: Thinking Machines describes Inkling 2 Small as a MoE with about 594B total parameters and 17B active, with image input and variable effort. It scores 31 on the Artificial Analysis Intelligence Index. Fine-tuning and inference start on Tinker with closed weights; a weight release is planned after external safety testing and red-teaming. Artificial Analysis says that if weights ship as planned, it would be among the highest-scoring open-weight models developed outside China on its current rankings [Third-party claim].

Sources: @thinkymachines (~4,118 likes) · @ArtificialAnlys · Introducing Inkling 2 Small

In brief

• Perplexity Teammates [Company claim]: personal agents for ongoing Max work—goals instead of one-off prompts; proactive use of Computer tasks, skills, memory, and tools; teammate collaboration and Slack. Web first; native apps later. (@AravSrinivas · @perplexity_ai · Perplexity blog)
• Claude Code / model User-Agent email leak [Community report]: HN post (~1,335 points) points to GitHub issue #107757. @mitsuhiko says the pattern often follows 401/403 or a server request, and Claude Code makes it more visible because email is frequently in context (~198 likes). (HN · issue · @mitsuhiko)
• Antigravity redesign and remote agents [Company docs]: UI refresh; Remote Agents run in Google-managed environments for personal Pro/Ultra, rolling out gradually. (@_mohansolo · @ai_for_success · docs · Changelog)
• Hot GitHub Skills / orchestrators: cloudflare/security-audit-skill (+2,524 today, ~63K stars), mattpocock/skills (+2,058, ~316K), google/skills, stablyai/orca, Paradev, NVIDIA/auto-ontology. (security-audit-skill · mattpocock/skills · google/skills · orca · Paradev · auto-ontology)
• LangChain / Mastra nits [Company claim]: async subagents and latency metrics. (@LangChain · @mastra)
• AI stock selloff and SI politics [Media report / political statements]: post-midterm regulation and data-center worries; levels per original market sources. @emollick argues a drawdown would not greatly change the AI trajectory [Opinion]. (@emollick)
• Qwen: The becoming begins. [Unverified / teaser]: ~7,316 likes; no companion spec page in this window. (@Alibaba_Qwen)
• GPT-6.2 “today” chatter [Unverified / community rumor]: community impatience; no matching OpenAI launch note at verification time. (@theo · @kimmonismus)
• Show HN: Stroq: local audit of what a coding agent did and why (~88 points). (Show HN · GitHub)

---

About this column and how we verify

1. AI-assisted editing: This AI Daily Brief is AI-assisted and was prepared with help from Grok Bot for the Beijing-time window 2026-11-06 00:00 through about 22:28.
2. Tweet-by-tweet checks: Every cited post was fetched individually to confirm it exists, the author matches, and the timestamp falls in this edition’s window; this edition also ran a second-pass check of all selected IDs through the X API. Like and view counts are as of editing time (~22:28 Beijing time).
3. X News summaries are leads only: X News blurbs are Grok-generated and used only to find stories; facts follow original posts, official pages, and media. When a summary conflicts with a primary source, the primary source wins.
4. Tag meanings: [Company claim] is an official post or product page; [Third-party assessment] is an independent public evaluation; [Opinion] is a personal judgment; [Unverified / teaser] and [Unverified / community rumor] lack a matching official spec or launch note; [Media report / political statements] and [Media / secondhand] are press or relayed accounts; [Community report] is a public issue plus developer commentary; [Company docs] are product docs; [Third-party claim] is a leaderboard operator’s own statement.