New Models from Anthropic and Meta, Nvidia Still Under Fire

JensenGate: Nvidia Pressed Commerce Department to Halt Megaspeed Investigation

In Washington, pressure is mounting following the espionage scandal surrounding Nvidia CEO Jensen Huang to tighten export controls on AI chips. This creates a new political battleground for Nvidia, of all things on an issue central to the company’s core business. The catalyst is Huang’s admission of having had an affair with Alice Huang. She is an executive at Nvidia customer Megaspeed, and US authorities consider her an asset for Chinese intelligence. She is not related to Jensen Huang.

Huang, 63, stated on Sunday that there was “no excuse” for his actions. Nvidia’s board of directors has launched an independent investigation. It is examining whether the head of one of the most critical companies of the AI boom was the target of a foreign intelligence operation. The board stated it has “no reason to believe” that national security or Nvidia secrets were compromised. Nevertheless, the inquiry is set to evaluate all available information.

The backstory dates back to last year, when questions first arose about Megaspeed, a Singapore-based major buyer of Nvidia chips. The suspicion was that the company was rerouting portions of its shipments to China in violation of US export restrictions. Both Nvidia and Megaspeed rejected the allegations. Megaspeed stated it complied with all applicable laws, including US export controls. Last month, reports surfaced detailing a “close relationship” between Jensen Huang and Alice Huang. President Trump was reportedly briefed on the intelligence agencies' assessment, but took no action.

According to new reports, Nvidia directly intervened in the investigation against Megaspeed. In late summer 2025, Nvidia’s general counsel Tim Teter reportedly phoned Jeffrey Kessler, who oversees chip export controls as Under Secretary of Commerce. According to insiders, Teter wanted to know whether Kessler’s Bureau of Industry and Security had been involved in the arrest of a Megaspeed executive by Singaporean authorities. Another report concludes that Nvidia apparently successfully pressured the Commerce Department to halt an inquiry into potential chip smuggling by Megaspeed. The contact was reportedly facilitated by a Trump-appointed official.

At the center of the question of what China might have achieved through Alice Huang lies the close relationship between Huang and Trump. Since the start of Trump’s second term, the two have met at least seven times. Huang flew on Air Force One and took phone calls from the president during public appearances. → The Information, The Guardian, Bloomberg Law

Synthszr Take: Export controls were previously negotiated over compute capacity and performance thresholds—arenas where Nvidia won on technical arguments and direct access to the president. The scandal shifts the debate from the spec sheet to a question of trust. In that realm, lawmakers no longer need proof of smuggled chips; a phone call from the general counsel to Jeffrey Kessler is enough. Huang’s at least seven meetings with Trump were Nvidia’s most powerful lobbying asset; now they serve as circumstantial evidence. The next round of restrictions will therefore hit intermediaries like Megaspeed in Singapore first.

Anthropic Releases Claude Fable 5.2 and Mythos 5.2

On Monday, Anthropic released Claude Fable 5.2 and Claude Mythos 5.2. API prices dropped by 20 percent to $8 per million input tokens and $40 per million output tokens. Their generation 5.1 predecessors cost $10 and $50. Both models are based on the same foundational model. Fable features stricter safeguards for cybersecurity and biology and is available to all paid Claude subscribers, as well as in Claude Code, on the Claude Platform, and via AWS, Google Cloud, and Microsoft Azure. Mythos remains restricted to vetted organizations in cybersecurity and life sciences, accessible via Anthropic’s verification programs.

Alongside pricing, Anthropic emphasized speed. According to the company, the new models generate output up to 30 percent faster than their predecessors. An optional Fast Mode is designed to accelerate token generation by up to 2.5x and costs double the standard rate at $16 and $80. Even after the price cut, Fable remains twice as expensive as Opus 5.5, which costs $4 and $20. OpenAI’s GPT-6 Astra remains at $10 and $50 for prompts up to 272,000 tokens. The naming convention is notable: Opus, Sonnet, and Haiku already carry the version number 5.5.

Regarding cost per task, Anthropic points to its system card. In the FrontierCode 2.0 evaluation, a task with Fable 5.2 on the High effort setting cost an average of $6.72. Opus 5.5 on the Max setting reached $12.82 for an almost identical result. The highest setting, Max, pushed Fable’s costs to $44.70, yielding a score lower than the cheaper Xhigh setting. In the Claude apps and Claude Code, Medium is now the default setting. Anthropic continues to recommend Opus 5.5 for most coding and knowledge tasks. Fable is meant to serve as an advisor consulted by smaller models on complex problems, or as an orchestrator delegating simpler sub-tasks to them.

On benchmarks, according to Anthropic’s own comparison, one of the two new models leads in 13 out of 15 reported evaluations. On FrontierCode 2.0, Fable 5.2 achieves 49.1 percent, compared to 32.4 percent for Fable 5.1, 45.5 percent for Opus 5.5, and 42.8 percent for GPT-6 Astra. On FrontierSWE v2, Fable hits 69.9 percent, ahead of Astra at 65.5 and Opus at 62.3 percent. Google’s Gemini 4 Carbon leads the business workflow benchmark AutomationBench at 53.9 to 51.4 percent. Opus 5.5 maintains a narrow lead in the Elo ranking of the knowledge work benchmark AA-Briefcase. On the Artificial Analysis Intelligence Leaderboard, Fable 5.2 claimed the top spot from Gemini 4 Carbon on launch day. → Anthropic, VentureBeat, Mashable, The New Stack

Synthszr Take: The 20 percent discount on the list price is the smaller number. The more important one appears in the system card: Fable on High solves a FrontierCode task for $6.72, while Opus 5.5 on Max takes $12.82 for virtually the same outcome. The more expensive model is thus the cheaper one per task. This benefits customers with enough volume to systematically evaluate effort levels and routing. Smaller customers keep paying by the token, while large enterprise customers buy outcomes: Fable on Max costs $44.70 and delivers less.

Meta’s Muse Ember Matches Claude Haiku 5.5 Performance at One-Third the Cost

Meta launched Muse Ember on Monday, a low-cost, proprietary model. According to VentureBeat, it is designed for multi-step agentic tasks such as support tickets, test runs, and minor code fixes. In independent evaluator Artificial Analysis’s Intelligence Index, Ember scored a rounded 36 points at maximum reasoning effort, matching Anthropic’s Claude Haiku 5.5. Unrounded, the scores stand at 35.8 and 36.2. Per benchmark task, Ember cost roughly 14 cents, compared to 41 cents for Haiku. Artificial Analysis attributes the difference partly to cheaper read rates for Prompt Caching and partly to Ember consuming only about a third as many output tokens. → VentureBeat

Synthszr Take: On the rate card, Ember, Haiku, and GPT-6 Luna are identical: $0.10 input, $0.50 output per million tokens. The 3x factor comes from cheaper cache reads and consuming one-third the output tokens. That is exactly how the cloud price wars unfolded: once AWS, Azure, and Google normalized instance pricing, competition shifted to discount tiers and egress fees. Buyers must therefore compare cost per completed task, because the pricing tier above 200,000 tokens shows where Meta claws back its margin.

This week, the boundaries between tabloid drama and geopolitics are blurring: Nvidia’s Jensen Huang admits to an affair with an alleged Chinese spy while his board simultaneously initiates a national security review. At the same time, Apple’s new chief John Ternus pushes for significantly greater urgency in AI agents, and Iranian hackers target European supermarket refrigerators of all things. The common thread: trust, not just technology, is currently the truly scarce resource.

Instagram Lets Anyone Build Story AI Effects via Prompt and Links to the Creator

Instagram launched Effect Builder, a web-based tool allowing users to create AI effects for Instagram Stories solely via text prompts. The tool is part of the Edits video editing app and, according to the November 9, 2026 announcement, requires no design or coding skills. The workflow comprises three steps: users describe their desired look via prompt, test it against their own photos, and submit the effect. According to the company, every effect undergoes a safety review before appearing in the Stories effect browser. Users applying someone else’s effect will see the creator’s Instagram handle and can navigate directly to their profile. → Edits App

Synthszr Take: Instagram compensates new effect creators with a profile link, and that suffices. Snapchat and TikTok built their effect studios as pro-grade tools whose manual craftsmanship kept libraries scarce and valuable. A prompt removes that scarcity, meaning the volume of effects will explode. Scarcity will then reside solely in effect-browser placement—curated and controlled by Meta. The pre-release safety review thus becomes leverage over the ultimate currency: attention.

Apple Now Sells AI Usage in Logic Pro and Keynote as a $50 Pack

For the first time, Apple is offering additional usage capacity for its AI features in Logic Pro and Keynote. A so-called Intelligence Usage Pack can be purchased for a one-time fee of $49.99, as spotted by 9to5Mac and reported by The Verge. The packs apply to Apple Intelligence features with a usage limit that access cloud models from OpenAI or Google Gemini. In Logic Pro, this affects the built-in AI assistant, which can set up a new song or configure a vocal track via prompt. According to Apple, how quickly the limit is reached depends on the complexity of conversations with the assistant. The packs are tied to the purchasing Apple account and cannot be shared via Family Sharing. → The Verge

Synthszr Take: With the Usage Pack, Apple makes it clear that its proprietary AI is procured commodity goods, carrying unit costs owed to OpenAI and Google. Every prompt costs money, and no flat fee can absorb that indefinitely. The bill gets passed to the customer, billed according to the “complexity of conversations”—a metric no one can forecast in advance. For studios and agencies, software shifts from a fixed overhead to a variable cost center. Once Keynote follows, quarterly deck preparation will run on a metered connection too.

Cognition Warns AI-Driven Cyberattacks Are Entering a Dark Age

Alex Stamos, Chief Information Security Officer at AI developer Cognition, described a cybersecurity industry overwhelmed by a wave of AI-assisted attacks on The Verge’s Decoder podcast. He anticipates a “dark age.” In his assessment, criminal groups and state-sponsored actors from Russia or North Korea are using AI to discover and exploit vulnerabilities simultaneously. In October alone, alongside critical vulnerabilities in PHP and Node, there was a string of successful breaches, including several targeting so-called Neocloud providers. Stamos characterized the Binance crypto exchange breach as the largest theft in human history. → The Verge

Synthszr Take: The Fireworks chain shows where the clock runs out: compromised API keys remained active long enough to ship a malicious plugin to 5,000 enterprises via Cursor. AI slashes the cost of an exploit, driving their frequency up. Defensive play continues at the cadence of change advisory boards, maintenance windows, and monthly patch cycles. Response time thus becomes a critical procurement criterion. An AI tool whose vendor fails to revoke compromised keys within hours has no place in an enterprise supply chain, regardless of its productivity gains.

New Cyberattack on Fireworks AI Jeopardizes a Billion-Dollar Business

AI infrastructure provider Fireworks AI has once again been targeted in a cyberattack, according to The Information. The outlet views the incident as a threat to a business valued at one billion dollars. Fireworks runs an Inference platform hosting trained AI models, enabling enterprises to access open-source models via API. The phrase “new hack” in the headline suggests this follows an earlier security breach. → The Information

Synthszr Take: Fireworks sells the exact same open-source models as any competitor, which meant sales previously turned on token price and latency. Following this second incident, every sales pitch will open with a security questionnaire. That stretches sales cycles, ties up legal teams, and necessitates costly audits. The customer acquisition cost climbs while token pricing remains capped—because the exact same model is accessible just one API endpoint away. Fireworks bears that spread alone, handing every untarnished competitor a ready-made sales pitch.

France Accuses Iran’s Revolutionary Guard of Hacking European Supermarket Refrigeration

French authorities hold Iran’s Islamic Revolutionary Guard Corps (IRGC) responsible for a cyberattack on refrigerated displays and freezers in supermarkets, according to the Wall Street Journal. The attack caused food spoilage across Europe last month. French prosecutors plan to request international arrest warrants for the leadership of the IRGC cyber unit believed to have executed the breach. Foreign Minister Jean-Noël Barrot announced on Monday that the action “will not go unpunished.” According to the paper, US authorities confirmed that the tactics match established methods of Iranian hackers. The resulting damage to food supplies and equipment amounts to several million dollars. → Wall Street Journal

Synthszr Take: A few million dollars in spoiled goods, plus blood banks, answered by arrest warrants against officers who will never set foot in Europe. That calculus makes cooling systems an attractive target for Tehran. The attack disrupts daily life and rattles nerves, yet remains well below any threshold that would warrant military retaliation. Consequently, the burden of defense shifts from the state to retail chains and refrigeration manufacturers. Unhardened connected refrigeration becomes a direct liability risk, and insurers will price it into their premiums.

Trump Defends Relaxed AI Rules Post-Midterm Debacle, Citing China

In his first televised interview following the midterm defeat, Donald Trump defended his hands-off approach to AI regulation. His argument: China has no guardrails, and stricter mandates would disadvantage the US in global competition. The conversation with journalist Ross Douthat took place Friday at the White House. Alongside artificial intelligence, the discussion covered immigration, corruption, Iran, prospective successors, and Trump’s own mortality.

The backdrop is a severe defeat for Republicans in last Tuesday’s midterms. The party lost control of both the House of Representatives and the Senate. It also lost statewide offices in conservative strongholds like Texas, Kansas, and South Carolina, which Democrats had not won in decades. Unlike following the 2020 election, Trump made no claims of election fraud and appeared largely willing during the interview to accept the loss.

Trump primarily blamed the candidates for the outcome, stating that he himself was not on the ballot and that his popularity could not simply be transferred to nominees voters disliked. At the same time, he acknowledged that the war, rising gasoline prices, and widespread disapproval regarding Israel played a role. Douthat pointed out that Trump had personally endorsed candidates such as Ken Paxton in Texas and Darline Graham in South Carolina. Graham lost to a center-left challenger in a deeply red state.

Shortly beforehand, Vice President JD Vance had cited the war against Iran on social media as the primary cause of the defeat. Trump noted he did not disagree with that assessment, adding that Vance had asked for his view in advance and Trump gave him free rein. → CBS News

Synthszr Take: “China has no guardrails” is a line that requires zero subject expertise and turns every technical nuance into a loyalty test. Technical regulation would have to articulate which model requires which audit. The geopolitical competition frame simply defines who loses. Since Tuesday, that frame has hit a wall: the House and Senate have new majorities, and laws are drafted there, not in television interviews. AI providers banking their product roadmaps on perpetual Washington inaction are now planning around a majority that no longer exists.