Daily AI digest, November 9

The biggest AI news of November 9 came from Anthropic and Nvidia. Anthropic released Claude Fable 5.2 and Mythos 5.2 a week before its planned IPO. Nvidia shares fell as much as 8.8% after its board opened an investigation involving CEO Jensen Huang. Google added two new TPUs, and Meta released its cheapest Muse model. Below are the big stories of Monday, November 9, plus the day's rumours.

Anthropic releases Claude Fable 5.2 and Mythos 5.2

Anthropic released Claude Fable 5.2 for all users and on AWS, Google Cloud and Microsoft Azure. It costs $8 per million input tokens and $40 per million output tokens. Anthropic says both models generate output up to 30% faster than the 5.1 versions.

Fable 5.2 scores 49.1% on FrontierCode 2.0, against 45.5% for Opus 5.5 and 42.8% for GPT-6 Astra. Mythos 5.2 is the same model with lighter cyber and bio safeguards. It is open only to verified security and life sciences teams and to Claude Security customers.

Anthropic also says a team of 1,000 Fable 5.2 agents formalized Fermat's Last Theorem in Lean in 30 hours. Our breakdown of the benchmarks, prices and safety tests is here.

Nvidia shares fall as the board investigates

Nvidia's board launched an independent investigation on Sunday after reports that Jensen Huang was the target of a foreign intelligence effort. The board said it has no reason to believe national security or Nvidia secrets were compromised.

Nvidia shares fell as much as 8.8% at Monday's open, from $186.64 to $170.24. By late morning they had recovered to about 4% down. We explain what analysts expect and what comes next in our article.

Google and Meta push on price

Google added TPU 8tx and TPU 8i Max to its eighth TPU generation, making four new TPU systems this year. TPU 8tx runs confidential inference on AMD EPYC hosts with up to 2x lower latency than TPU 8t. TPU 8i Max puts two TPU 8i chips in one package and doubles compute and HBM bandwidth. Google gave no release date or prices; our breakdown is here.

Meta released Muse Ember, its smallest Muse model, at $0.10 per million input tokens and $0.50 per million output tokens. It scores 36 on the Artificial Analysis Intelligence Index, about level with Claude Haiku 5.5, at about a third of the cost per task. It trails Haiku 5.5 on FrontierCode 2.0, as we explain in our article.

Fireworks AI discloses a data breach

Fireworks AI told customers on November 9 that an October 24 intrusion through a cloud provider reached its development machines. The attackers accessed billing data and, for some accounts, secrets, datasets and incomplete model weights, Bloomberg reports via RuntimeWire.

It is the second known breach at the inference provider in recent weeks. Fireworks investigated with Palo Alto Networks' Unit 42. The disclosure came two days after Fireworks named Meta's Dave Wehner as its CFO.

Smaller updates

Nineteen smaller stories came out on Sunday evening and Monday.

• ByteDance released Seedream 5.5, an image model that keeps edits stable for 100 rounds and cuts severe text errors to 0.4%. Details are here.
• Epic Games released the Unreal Engine 5.9 Preview, which it says was built mostly by AI agents in about a week. Claude Code and Codex can now run inside the editor; more in our article.
• Baidu introduced ERNIE Evolve, an automated research system that sped up ERNIE 5.5 Flash inference by up to 33.7%. We covered it here.
• Zyphra released ZONOS3-Mini, an Apache 2.0 text-to-speech model with about 100M active parameters that clones voices on a phone. More in our article.
• Google AI Studio can now build Google Maps apps from one prompt with a free demo key capped at 100 calls per API per day. Details are here.
• OpenAI started selling a $200 Codex edition of the ModRetro Chromatic handheld, which lets Codex write games onto a cartridge. We covered it here.
• Adobe became Disney's "Official AI Innovation Partner" in a multi-year deal announced at Adobe MAX. Disney will build custom Firefly models trained on its own properties.
• Mavria, a Houston startup building factory-made 5 MW data-center blocks, raised $130M, including a $100M round led by Paradigm at a $1B valuation.
• Mississippi regulators released penalty details against SpaceXAI for starting its Southaven data-center build without water permits. The maximum fine is $25,000 a day.
• A public notice puts SpaceX's planned supercompute site near Starbase, Texas, at $24B for its first phases, with about 748 MW from gas turbines. A hearing is set for December 9.
• President Trump told CBS "60 Minutes" on Sunday that new AI data centers build their own power plants and feed the grid. We checked the claim here.
• Korean startup DeepX taped out its DX-M2 edge chip, which targets 20B to 100B parameter models at 5 W. Samples are due in March 2027.
• Nature Will released Natura-1, a 50B-parameter model for dexterous robot hands trained on over 1 million hours of human video.
• ChatCut built its own GPU render engine so AI can write motion graphics as code without a browser. More in our article.
• Mistral showed Voxtral 3 captioning AI Engineer Paris live, using names and terms from the agenda. We explained context biasing here.
• Nous Research released Hermes Security, a free plugin that lets Hermes Agent find and test security bugs. More in our article.
• Elvis Presley's rights holders and Eminem's publisher Eight Mile Style filed separate lawsuits against Suno. Suno says it bans the use of artists' names and songs, Engadget reports.
• In Co-Design Bench, a test built from real user requests by Alibaba and Asian universities, the best agent fully solved only 26.4% of tasks. That agent was GPT-6 Astra, The Decoder reports.
• We had Sonnet 5.5 and GPT-6.1 Sol build a steam train scene in three.js. Sol cost $8.33 against about $46.28 for Sonnet; the results are here.

Six of the nineteen are tools for making images, video, sound or games.

Rumours

Three more stories circulated on November 9, but we could not read the primary report or confirm them with the companies.

• A Commerce Department probe of Singapore chip reseller Megaspeed was reportedly moved to the Justice Department after Nvidia intervened, The New York Times reports.
• The EU's draft Digital Fairness Act would reportedly target AI companion bots built to prolong engagement, with fines up to 6% of revenue, Bloomberg reports.
• A GitHub user claims updated Google support pages allow human reviewers to see Drive and Gmail content used by Gemini. Google has not commented.

Treat all three as unconfirmed until the companies announce them.

What the day was about

Monday's news showed the frontier labs racing on price as much as on power. Anthropic lowered Fable's price and sped it up, and Meta matched Haiku 5.5 at about a third of the cost. Google doubled what fits in one TPU rack slot.

Trust was the other theme. Nvidia lost as much as 8.8% in minutes over an investigation involving its CEO, and Fireworks admitted a second breach. On November 9 the models got cheaper and better, while investors and customers watched how safely the companies behind them operate.

Sources

• Introducing Claude Fable 5.2 and Claude Mythos 5.2 – Anthropic
• Our newest TPUs: 8tx and 8i Max – Google
• Introducing Muse Ember – Meta
• Fireworks AI discloses second security breach – RuntimeWire
• Adobe and The Walt Disney Company alliance – Adobe
• Mavria raises $130 million – Crypto Briefing
• SpaceXAI started data center build without water permits – Insurance Journal
• DeepX tapes out second-gen chip – EE Times
• Nature Will launches Natura-1 – Robotics and Automation News
• SpaceX's $24 billion supercomputer at Starbase – Crypto Briefing
• Suno sued again for copyright infringement – Engadget
• AI coding agents still need human cleanup – The Decoder