ElevenLabs is letting employees sell $300 million of vested shares at a $22 billion valuation, twice the $11 billion mark reached in February after the voice AI startup raised $500 million.
Wellington and T. Rowe Price co-led the transaction, buying shares with the intention of holding them after ElevenLabs goes public. It is the company’s second employee tender, following a $100 million sale at a $6.6 billion valuation in September 2025. ElevenLabs, founded in 2022 and based in New York and London, makes ultra-realistic human voices and sound effects.
Why it matters: The second tender gives ElevenLabs another retention tool as it competes with fast-growing AI startups for employees, while Wellington and T. Rowe Price are positioning for a return tied to a future public listing.
Zhipu AI’s open-weight GLM-5.3 nearly matches Anthropic’s Claude Mythos Preview at autonomous exploit development—and lacks comparable safeguards. On ExploitBench, GLM-5.3 produced working exploits in 50 of 410 attempts, versus 56 for Mythos; on Anthropic’s binary-exploitation test, it took control in 4% of tasks, versus 6%.
With a human expert, GLM-5.3 found unknown browser JavaScript vulnerabilities and chained them into an attack that read any file, including an SSH key. Anthropic reported the flaws; driver and firmware findings remain under review.
GLM-5.3-Flash turned a newly disclosed Chrome bug into an attack in eight hours of model time and 20 minutes of human attention, costing $20.40 through Zhipu’s API. CAISI called GLM-5.3 the most cyber-capable open-weight model to date.
Why it matters: Zhipu’s downloadable weights let an experienced team strip GLM-5.3’s refusal behavior for about $1,200, while Anthropic keeps Mythos behind vetted access; defenders therefore face an open model whose harmful-request refusal rate can fall from over 90% to 2–12%.
OpenAI’s Dots are launching exclusively for ChatGPT Pro subscribers, whose plans start at $100 a month. The agent runs on GPT-6 Astra and competes with Meta’s free Muse, available to anyone who downloads the app or visits its website.
In September testing, a Dot named Toolie searched ChatGPT history, flagged a pending data-request follow-up, suggested dinner spots, and offered to refine reporting pitches. Unlike a standard chatbot, it can message users without prompting and handle recurring tasks. Setup suggested linking Gmail and Google Drive, but the tester kept the agent from automating meaningful parts of life.
Why it matters: OpenAI must prove that proactive help is worth both $100 a month and access to personal accounts, while Meta removes the price hurdle with Muse and leaves permissions as the sharper test of whether users will delegate recurring tasks.
DeepMind’s SynthID Bio embeds a verifiable signature in AI-designed proteins that survives the move from digital model to synthesized molecule. It subtly guides amino-acid choices in sequences or adjusts atomic coordinates in predicted 3D structures without compromising biological function in laboratory tests.
Across binders targeting VEGF-A, the SARS-CoV-2 spike RBD, and PD-L1, watermarked designs matched unwatermarked versions on hit rate, binding affinity, and natural sequence diversity. A version built into AlphaFold 3 preserved prediction accuracy and delivered near-perfect detection after digital noise or minor coordinate changes.
The approach could help DNA-synthesis providers review unfamiliar AI-generated orders and databases flag mislabeled synthetic entries. Deliberate tampering remains a challenge.
Why it matters: Twist Bioscience could use SynthID Bio to focus manual review on unfamiliar DNA orders rather than rely on the assumption that novel sequences are natural, while the Protein Data Bank, UniProt, and GenBank could flag mislabeled AI-generated entries.
Google released Gemini 4 Argon, its first frontier model in more than seven months, putting it back in competition with OpenAI and Anthropic. Google says it can autonomously find, validate, and patch critical software vulnerabilities; trusted cyber defenders in Fairwind and Google's internal teams get initial access.
Argon raises the output limit from 64,000 to 1 million tokens and costs $2 per million input tokens and $10 per million output tokens. Google will expand access, starting with paying API customers and Google AI Ultra subscribers, but has not set a date.
Artificial Analysis scored it 53, tying GPT-6 Astra and Claude Fable 5.1 but trailing Claude Opus 5.5 at 58. Vals AI ranked it first at 68.9%; Google's own benchmarks put Argon ahead in most tests.
Why it matters: For API buyers comparing frontier models, Argon's $2 input rate comes with an average 62,000 output tokens per task—more than twice GPT-6 Astra's 27,000—so token efficiency matters more than the headline price.