📜 The Primary Source

Thursday, November 19, 2026

12 stories · Standard format

Generated with AI from public sources. Verify before relying on for decisions.

🎧 Listen to this briefing or subscribe as a podcast →

Today on The Primary Source: Google's Gemini 4 Neon arrives with a headline price that obscures its heavy 'thinking' token consumption, while OpenAI drops GPT-6.1 Sol to push agentic costs to a new floor. We also cover banks quietly shifting AI-agent liability to account holders, NYC's unprecedented use of the False Claims Act in housing enforcement, and a linear-time breakthrough on the Four-Color Theorem.

Frontier AI (Practitioner)

Google Launches Gemini 4 Neon at $3/$15 Per Million Tokens — But Thinking Tokens Consume 90% of the Real Bill at Default Settings

Google launched Gemini 4 Neon on Wednesday at $3 per million input tokens and $15 per million output tokens with a 95% cache discount ($0.15/M), four configurable thinking levels, and a 1M-token context window. Independent API testing found that at the default HIGH thinking setting, thinking tokens comprised 90% of the total bill — making the effective cost-per-task higher than Carbon through November 30 (when Carbon's introductory rates expire). At XHIGH thinking, Neon's per-task cost of $6.44 exceeds Carbon's $4.90. After November 30, Neon runs approximately 25% cheaper than Carbon at standard rates. Third-party benchmarks rank Neon #3 on Artificial Analysis Intelligence Index (61 at XHIGH), #1 on Artificial Analysis Coding Agent Index (58.6), and #2 on Vals Index (70.9%).

Google's 32% cost-per-task advantage over Carbon — the headline claim — is calculated at Google's chosen thinking level and benchmark selection. At default HIGH thinking in real API calls, the gap inverts until November 30. The practical decision for agentic workloads requires modeling your actual thinking-token ratio per task type, not comparing $3 to $4 per million. Neon's coding-agent #1 ranking and legal/tax benchmark leads are relevant for routing decisions, but the per-task cost audit comes before the benchmark ranking in the buy decision. Watch the December 1 pricing reset: that is when Neon's 25% discount over Carbon actually materializes.

Verified across 3 sources: aiweekly.co · Google · AI Chat Daily

OpenAI Ships GPT-6.1 Sol at $2/$10 With 95% Cache Discount — Near-Astra Performance at One-Fifth the Price

Following up on Tuesday's release of GPT-6.2 Astra and Sol, OpenAI shipped GPT-6.1 Sol on Thursday at $2 per million input tokens and $10 per million output, with cached input at $0.10/M — a 95% reduction from standard input pricing. Per OpenAI's own benchmarks, Sol matched GPT-6.2 Astra's DeepSWE v1.1 score at roughly one-fifth the cost per task; on OSWorld 2.0 offline computer-use tasks it came within 2.1 percentage points of Astra at approximately one-seventh the per-task cost. The $0.10/M cached input rate confirms the halved cache pricing for the Sol line we noted earlier.

The $0.10/M cached-input rate is now a cross-vendor competitive floor: Anthropic's Opus 5.5 charges $0.20/M for cache reads, Qwen4-Max $0.075/M, and Sol now $0.10/M. For agentic workflows where cached context dominates token spend, Sol at one-fifth Astra's list price with near-parity coding performance makes the default-to-Sol, escalate-to-Astra pattern economically defensible. OpenAI's internal benchmark numbers require independent corroboration before treating them as settled, but the verifiable pricing updates the routing arithmetic.

Verified across 2 sources: OpenAI · Kimbodo

Banks Rewrite AI Agent Liability Terms — Citi, PNC, and Citizens Shift Account-Holder Risk on Third-Party Agent Activity; AlphaProof Code Formally Verifies Signal Protocol in Four Weeks

Citi, PNC, and Citizens updated customer terms on Wednesday to hold account holders responsible for activity that third-party AI agents perform in their accounts — including activity after credentials are leaked or the agent is compromised. Citi carves out an exception for AI agents Citi itself provides. Separately, DeepMind released AlphaProof Code, a multi-LLM agentic pipeline in which Gemini Carbon agents generated over 6 million lines of Lean 4 to formally verify Signal's end-to-end encryption protocol in four weeks, flagging approximately 1% of code for human review versus years of manual work.

The bank liability shift is the more consequential piece: by writing the "your agent, your risk" answer into contracts before courts settle it, Citi and peers are creating a two-tier structure that benefits bank-provided agents over third-party ones. Any fintech or AI-consultant advising clients on automated financial workflows now has a material new legal exposure to disclose — the client bears loss from agent compromise, not the bank, unless the agent is bank-sourced. The AlphaProof Code result is evidence that formally verified software assurance at cryptographic-protocol scale is becoming tractable, with implications for compliance-heavy regulated deployments where audit trails must be machine-verifiable.

Verified across 1 sources: AI Weekly

Agent Architectures & Tooling

AHE-Beta: Requirement-Centered Harness Evolution Lifts Pass@1 from 21.9% to 33.3% While Cutting Tokens 40.8% — Gains Transfer Across Model Families

Researchers published AHE-Beta (arXiv, Wednesday), a requirement-centered loop for evolving agent harnesses that simulates user issues, implements generalizable fixes, and merges only verified solutions. On held-out tasks, AHE-Beta improved pass@1 from 21.9% to 33.3% while using 40.8% fewer tokens than the base harness — outperforming Claude Code, Meta-Harness, and prior AHE. In a blind human A/B test, real users preferred AHE-Beta in 31 of 48 valid pairs and rated it higher on all five measured dimensions. Gains transferred to GLM-5.3 and DeepSeek-V4.1-Flash without further harness evolution.

The transfer finding is the load-bearing result: a harness optimized on one model family improves performance on structurally different models without retraining. That means harness investment depreciates more slowly than model investment — a harness built today survives the next two or three model generations. The 40.8% token reduction alongside a 50% relative accuracy gain also inverts the usual cost-quality tradeoff assumption. The concrete implication for practitioners designing production agentic systems: harness iteration budget deserves parity with model selection budget, not residual attention.

Verified across 1 sources: arXiv

LangGraph Deep Agents Ships Async Subagent Steering — Mid-Flight Task Redirection Without Restart Overhead

LangGraph's subagent design, documented Thursday, quarantines heavy computation — reading 40 files, long searches, diffs — into separate agents that return only conclusions, protecting the parent's context window from intermediate tool call bloat. AsyncSubAgentMiddleware on Agent Protocol servers adds four live steering tools: start_async_task, check_async_task, update_async_task, and cancel_async_task. The update_async_task tool enables supervisors to redirect running jobs mid-flight without restart. Three delegation modes are supported: synchronous parallel task calls, isolated mode (subagent sees only task description), and fork mode (subagent inherits parent conversation history).

Mid-flight task steering turns long-running subagent jobs from fire-and-forget operations into steerable work items — a qualitative shift for workflows where early output reveals the task needs reframing before completion. The isolated/fork mode choice is a concrete context-budget lever: isolated minimizes context overhead but requires complete task descriptions; fork reduces description cost but exposes the subagent to parent history that may be irrelevant noise. The pattern also compresses the cost of error recovery: catching a wrong direction at turn 5 and redirecting via update_async_task costs one steering call, versus canceling and restarting a 20-turn job.

Verified across 1 sources: Vadim's Blog

Independent Print Publishing

USPS January 2027 Rate Package: Periodicals-Class Mail Is Exempt From 5.9% and 11% Surcharge — Parcel and Package Services Are Not

A Walsworth analysis of the USPS January 17, 2027 rate filing explicitly confirms what we tracked in late October: market-dominant mail categories, including First-Class and Periodicals, are exempt from both the 5.9% shipping base increase and the new 11% permanent fuel surcharge. Those hikes apply strictly to competitive products like Priority Mail, Ground Advantage, and Parcel Select. International rates across all categories will still rise 7.9% on average.

For Kav Magazine and other independent periodicals, this formally confirms that the January Q1 2027 cost spikes we've been tracking will not hit your core distribution postage. The distinction matters for near-term P&L modeling, though the operational risks we outlined yesterday—including the Canada mail suspension and PMG Steiner's threat of degraded service standards—remain fully live ahead of the anticipated July 2027 rate hike.

Verified across 4 sources: Walsworth · GeekSpin · USPS · WWD Sourcing Journal

Frum Community & Rockland Local

NYC Uses False Claims Act Against a Landlord for the First Time — Barry Singer Faces $10M+ in Treble Damages Across 11 Bronx and Brooklyn Properties

Mayor Mamdani's administration filed a False Claims Act lawsuit Thursday against landlord Barry Singer, the first time NYC has deployed that statute against a building owner. The suit seeks over $10 million in damages across 11 properties in the Bronx and Brooklyn, alleging Singer knowingly avoided $1.05 million in charges for emergency repairs and pest control. The filing covers 456 violations at 2500 Webb Avenue (183 immediately hazardous) and 261 violations at 3030 Heath Avenue (105 immediately hazardous). Under the False Claims Act, penalties run up to $28,619 per violation plus daily civil penalties of $1,000 per violation — structurally different from standard HPD enforcement.

The False Claims Act's treble-damages mechanism turns HPD's existing violation count into a multiplier — 717 violations across two buildings alone, at $28,619 each before trebling, represents potential statutory exposure of roughly $60 million before daily penalties. That arithmetic is new. Prior enforcement relied on repair orders, fines, and receiver appointments; the FCA adds federal-style fraud liability to routine NYC housing violations. For any multi-family operator with open HPD violations and unpaid city charges, the Singer case establishes that the administration is willing to use the heaviest available enforcement tool — and that the "chronic violator" designation is no longer a reputational label but a liability category.

Verified across 1 sources: NYC Mayor's Office

Small Multi-Family Real Estate

DHCR Fact Sheet 40 Update Shifts Concession Litigation From Categorical to Fact-Specific — 12,000 Units and $1 Billion in 421-a Credits at Stake

Yesterday we noted that DHCR's reissued Fact Sheet 40—which restores the distinction between temporary concessions and preferential rents—is undercutting the recent wave of overcharge litigation. Today, detailed reporting from The Real Deal and Bulwark quantifies the stakes: the guidance reversal directly addresses at least 21 pending lawsuits against landlords including Two Trees, Black Spruce Management, and Blackstone, spanning 24+ properties, 12,000+ units, and $1 billion in 421-a tax credits. As we covered, landlords can now defend concessions on factual grounds (pandemic occupancy, construction) rather than facing categorical liability.

We've already detailed the strategic burden shift from categorical liability to fact-specific defenses. The new $1 billion figure attached to the 421-a tax credit exposure explains why tenant attorneys will keep pursuing those specific properties: the 421-a program creates independent compliance obligations that plaintiffs can still target, ensuring these 21 lawsuits will be heavily litigated despite the DHCR update.

Verified across 3 sources: Bulwark · The Real Deal · The Real Deal

Personal Finance Mechanics

Wise Outlines Lower-Price Growth Strategy at Morgan Stanley Conference — Take Rate Cut to 49bps, 50%+ of Revenue Now Non-Cross-Border

Wise CFO Emmanuel Thomassin outlined the company's growth strategy Thursday at the Morgan Stanley European Technology, Media and Telecom Conference: the take rate was cut to 49 basis points from 52 basis points a year prior, platform pricing is approximately 40 basis points, and over 50% of net revenue now comes from non-cross-border sources including interest income and interchange fees. The company moves $240 billion annually against a stated $45 trillion addressable market, with 78% of transactions running through direct integrations with nine global payment systems. Customer deposit holdings grew 28% to $44 billion. Wise's OCC federal banking charter application was rejected in November, as previously reported; a new application is in preparation under state licenses.

Wise's revenue mix shift — over half from deposits, interest, and interchange — means the company's economics are now materially interest-rate sensitive at $44 billion in customer holdings. With the 10-year holding above 5.14% (covered separately), Wise's interest income buffer grows, partially insulating it from the take-rate compression it is deliberately engineering. The strategic logic is a price-volume bet: cut unit margin, capture more of the $45 trillion TAM, and harvest interest income while rates remain elevated. The OCC rejection adds execution risk to the US banking charter path, though state licenses cover current operations. The first-half FY2027 numbers confirm the thesis is tracking — 23% revenue growth on 24% volume growth — but operating expenses at 28% growth exceed revenue, compressing the margin story.

Verified across 2 sources: Investing.com · Kalkine

Jewish History from the Archives

Lithuanian State Historical Archives Announces 23-Million-Image Digitization of Grand Duchy Records — Free Public Access via FamilySearch, Russian Occupation Named as Urgency Driver

The Lithuanian State Historical Archives announced Wednesday a four-to-five-year digitization project covering approximately 23 million images across 1,576 inventories in 1,043 archival fonds, funded by FamilySearch at no cost to the archives. Prioritized collections include nearly 16,000 Old Acts books from the Grand Duchy of Lithuania, completion of church register digitization for all denominations, noble-status confirmation cases, and Vilnius residents' passport cards from 1915–1917. The project represents a 4.6-fold scale-up from the 5+ million images digitized in thirteen years prior to 2026. All results will be stored in Lithuania's cloud-based EAIS system and accessible through FamilySearch free of charge. The preamble explicitly names Russian occupation — referencing archive removals in 1915 and 1939 — as an existential threat justifying urgency.

The Grand Duchy church registers, court records, and noble-status files are primary sources for reconstructing Jewish demographic and social history in the Lithuanian region through the late Imperial period. The free public access mandate via FamilySearch converts what was archive-visit-required research into distributed remote scholarship — lowering the entry cost for genealogical and historical work on communities whose physical traces are increasingly at risk. The LDS baptism-for-the-dead association surfaces as a persistent institutional question even where archivists confirm no religious activity during digitization; Jewish genealogical organizations have historically negotiated access conditions around this. The five-year timeline and 23-million-image scale make this the largest single expansion of Eastern European Jewish archival access since Project Zikaron.

Verified across 2 sources: Lithuanian State Historical Archives · LRT

Recreational Math & Computation

Four-Color Theorem Gets Linear-Time Algorithm After 30-Year Stasis — Reducible Configuration Abundance Is the Structural Insight

An international team including Mikkel Thorup (University of Copenhagen) and Carsten Thomassen (DTU) presented an optimal polynomial-time algorithm for four-coloring any planar map at FOCS'26 on Wednesday, winning Best Paper. The algorithm reduces computational complexity from quadratic to linear time by discovering that planar graphs contain far more reducible configurations ('handles') than previously known, enabling simultaneous multipoint coloring. A map with 10× more regions now requires only 10× more computation rather than 100×. The team reports the structural insight extends beyond planar surfaces to toroidal and higher-genus surfaces.

The key advance is not the four-coloring result itself — that has been known since 1976 — but the discovery that an "astonishing number" of reducible configurations exist throughout planar graphs, enabling a fundamentally different algorithmic approach. The generalization to surfaces with handles suggests the abundance-of-handles insight may apply across a family of graph-coloring problems in topology, with downstream applications in network design, circuit layout, and computational geometry. The FOCS Best Paper designation and the 30-year gap since the last major algorithmic advance on this problem signal this is a genuine structural result, not an incremental optimization.

Verified across 3 sources: Phys.org · arXiv · Popular Science

OEIS A229457 and A229604: Claude Code Attempts First Formal Proof of Defective 3-Coloring Recurrences Open Since 2013

A formal verification project registered on GitHub Thursday is working to prove two empirical recurrence relations from OEIS: A229457 (defective 3-colorings with exactly one mistake on n×6 arrays, a 14th-order recurrence) and A229604 (exactly two mistakes, a 21st-order recurrence over 94 color-defect classes). The orchestrator, Claude Code (Opus), enumerated class counts and verified both recurrences against Hardin's b-files for n ≤ 210. The project carries a mathematical deadline of December 29, 2026; the orchestrator assigns P(Proved) = 0.85 for A229457 and 0.70 for A229604, citing certificate size risk for the latter.

These sequences represent the first open width in Potts antiferromagnet strip recurrences beyond the solved widths 2–5; a formal proof would close a 13-year gap in the combinatorial geometry literature and establish a reproducible template for proving OEIS sequences that carry empirical recurrences without formal verification. The 94-class lumping for the width-6 row transfer matrix is the technical bottleneck — certificate size grows with recurrence order, and a 21st-order recurrence generates a large formal object. The December 29 deadline makes this a live tracking item.

Verified across 3 sources: GitHub (The Omega Institute) · OEIS · OEIS


The Big Picture

Headline Token Prices Are Increasingly Fictional Without Thinking-Level Assumptions Three separate candidate analyses this cycle — Google Neon's 90% thinking-token overhead, Qwen4-Max's 77% output-token inflation versus Opus 5.5, and the dual-curve LLM price tracker showing frontier output rising while commodity capability deflates — converge on the same point: published per-token rates describe a scenario that rarely matches production. Practitioners who price agentic work off list rates without modeling thinking-token ratios, output verbosity, and cold-cache reload costs are using a systematically wrong number.

Harness-Level Engineering Now Outperforms Model Swaps on Both Cost and Accuracy AHE-Beta's requirement-centered loop lifted pass@1 from 21.9% to 33.3% at 40.8% fewer tokens, transferring gains to GLM and DeepSeek without retraining. Google Antigravity's mathematical analysis shows that evaluation gates at 95% accuracy correct sequential step-failure from 36% to 99.75% end-to-end success. The cascade-routing paper achieves 33% GPU-hour reduction by routing difficulty, not tasks. Taken together, the marginal return from switching frontier models is now smaller than the return from disciplined harness design — which is durable across model generations.

Third-Party AI Agent Liability Is Being Institutionalized Simultaneously by Banks and Courts Citi, PNC, and Citizens updated customer terms to shift liability for AI-agent-initiated transactions to the account holder — including post-compromise activity — while carving out bank-provided agents. The same week, NYC deployed the False Claims Act against a landlord for the first time, establishing treble-damages exposure for systematic evasion. Neither action is directly about AI agents, but the pattern is identical: institutions are front-running the ambiguity about who owns the action by writing their answer into contracts and precedent before courts settle it.

Northeast Multifamily Supply Withdrawal Is Converting Into Rent Floor Recovery October 2026 national rents rose year-over-year for the first time since June 2023, led by two-bedroom units, against a 26% drop in multifamily completions. The Northeast recorded a 49.3% completion drop — the steepest of any region. Meanwhile, the DHCR Fact Sheet 40 clarification shifts overcharge litigation to fact-specific grounds, Lee, Massachusetts saw 27–31% reassessment-driven tax increases, and rent-stabilized buildings in New York are trading at 43% below 2019 benchmarks. The supply withdrawal and regulatory relief are directionally constructive for small-landlord operating margins, but the tax and financing headwinds are compressing them from the other side.

Archival Digitization of Eastern European Jewish Records Is Accelerating Under Wartime and Funding Pressure The Lithuanian State Historical Archives announced a FamilySearch-funded project to digitize 23 million images across 1,043 archival fonds — a 4.6× scale-up from thirteen years of prior output — explicitly citing Russian occupation as a destruction risk. JewishGen simultaneously added 100,809 Ukrainian records covering ten localities from 1834–1926. Both initiatives reflect the same forcing function: deteriorating physical access and geopolitical instability are converting the "eventually digitize" posture into urgent infrastructure work, with a free public access mandate making the output usable for distributed scholarship rather than gated to institutions.

What to Expect

2026-11-23 — Spring Valley Planning Board public hearing at 6 p.m. on two-lot subdivision at 32 Twin Ave and proposed house of worship at 3 Tenure Ave, Village Hall.
2026-11-25 — FCC Order 26-67 TCPA consent-revocation rules take effect: category-specific opt-out withdrawals and exclusive opt-out method disclosure required. All seven standard opt-out keywords must be recognized by SMS senders.
2026-11-30 — Gemini 4 Carbon and Argon introductory pricing expires; Carbon doubles to $4/$20, Argon doubles — Neon becomes 25% cheaper than Carbon at standard rates from this date.
2026-12-01 — Rockland County Legislature public hearing at 7:05 p.m. on proposed eminent domain acquisition at 41 Spook Rock Road for Wesley Chapel Dam rehabilitation (Capital Project No. 3403).
2026-12-18 — FCC SIM Swap and Port-Out Fraud Order (47 CFR 52.37) compliance deadline: wireless providers must offer no-cost account lock/freeze options and provide advance notice of port-out requests.

Every story, researched.

Every story verified across multiple sources before publication.

🔍

Scanned

Across multiple search engines and news databases

1042
📖

Read in full

Every article opened, read, and evaluated

188
⭐

Published today

Ranked by importance and verified across sources

12

— The Primary Source

🎙 Listen as a podcast

Subscribe in your favorite podcast app to get each new briefing delivered automatically as audio.

Apple Podcasts
Library tab → ••• menu → Follow a Show by URL → paste
Overcast
+ button → Add URL → paste
Pocket Casts
Search bar → paste URL
Castro, AntennaPod, Podcast Addict, Castbox, Podverse, Fountain
Look for Add by URL or paste into search

Spotify isn’t supported yet — it only lists shows from its own directory. Let us know if you need it there.