Latest / Neural Newscast / Anthropic Hits $965 Billion with Opus 4.8 Launch [Model Behavior]
Transcript
- [00:00] Announcer: From Neural Newscast, this is Model Behavior, [00:03] Announcer: AI-focused news and analysis on the models shaping our world. [00:12] Nina Park: I'm Nina Park, welcome to Model Behavior. This show examines how AI systems are built, deployed, [00:19] Nina Park: and operated in real professional environments. Today is May 29th, 2026, [00:25] Nina Park: and we are tracking a major reorganization of the industry's power structure. [00:29] Nina Park: Anthropic has moved to the front of the pack, not just with new model weights, [00:34] Nina Park: but with a massive jump in private market valuation [00:37] Nina Park: and a key talent acquisition that signals a long-term play for model dominance. [00:41] Thatcher Collins: I am Thatcher Collins. It has been a busy 24 hours for the San Francisco lab. Between the release of [00:47] Thatcher Collins: Claude OPUS 4.8 and the announcement of a $65 billion funding round, [00:53] Thatcher Collins: they are clearly trying to distinguish themselves from OpenAI through a narrative of reliability. [00:58] Thatcher Collins: They are claiming a fourfold improvement in honesty benchmarks, though as we'll see, Nina, [01:03] Thatcher Collins: that term has a very specific technical definition in their model cards, [01:07] Thatcher Collins: focusing on factual calibration rather than just general truthfulness. [01:10] Nina Park: Let's start with that release. Yesterday, Anthropic made Opus 4.8 available across its API [01:16] Nina Park: and web interfaces. The headline figure is that the model is four times less likely than Opus 4.7 [01:22] Nina Park: to let code flaws pass unremarked. This isn't just about finding bugs. [01:27] Nina Park: It's about the model's willingness to admit when it isn't sure of an answer. [01:31] Nina Park: In their alignment assessment, they noted that Opus 4.8 reaches new highs in supporting [01:35] Nina Park: user autonomy and avoiding deceptive behaviors, [01:38] Nina Park: which has been a major hurdle for enterprise adoption. [01:41] Thatcher Collins: Nina, the focus on honesty is interesting because it addresses the persistent issue [01:46] Thatcher Collins: of model overconfidence. Early testers like Cursor and cognition are reporting that the model [01:51] Thatcher Collins: uses tools more cleanly and is less prone to hallucinations in complex coding tasks. [01:56] Thatcher Collins: On the terminal bench 2.1 benchmark for agentic coding, the score rose from 64.3% to 69.2%. [02:04] Thatcher Collins: It's an incremental gain, but in the context of autonomous agents, [02:07] Thatcher Collins: that 5% jump can be the difference between a successful deployment and a broken codebase. [02:11] Nina Park: While Opus 4.8 is the current flagship, Anthropic also teased what they call [02:17] Nina Park: the Mythos class models. They described Mythos as a full capability tier above Opus. [02:24] Nina Park: In a program called Project Glasswing, Anthropic has been working with partners like Apple [02:29] Nina Park: and Google to use Mythos for cybersecurity. They've already identified more than 10,000 high [02:36] Nina Park: or critical severity vulnerabilities in infrastructure software. They are being cautious [02:39] Nina Park: with the rollout because the model [02:41] Nina Park: can autonomously find zero-day vulnerabilities, which presents a significant risk profile. [02:47] Thatcher Collins: That caution is warranted. If a model can find zero days and create exploits for them, [02:53] Thatcher Collins: it's a powerful tool for defense, but a catastrophic one if mismanaged. [02:58] Thatcher Collins: Anthropics says Mythos requires stronger cyber safeguards before a general release, [03:04] Thatcher Collins: which they expect in the coming weeks. This seems to be their way of signaling that they have [03:07] Thatcher Collins: the intelligence lead while maintaining their stance as the safety-first lab it's [03:12] Thatcher Collins: a delicate branding balance, Nina, [03:14] Thatcher Collins: attempting to be the most capable while remaining the most restrained. [03:18] Nina Park: That branding is clearly working with investors. Anthropic confirmed today that it raised [03:23] Nina Park: $65 billion in a series h round led by Altitude Capital and Sequoia this brings their valuation to [03:30] Nina Park: $965 billion for context thatcher that overtakes OpenAI's last reported valuation of [03:38] Nina Park: $852 billion from March. Anthropic is also reporting an annualized revenue run rate of $47 billion, [03:47] Nina Park: which is a massive jump from the $1 billion [03:50] Nina Park: they were seeing at the end of 2024, showing incredible commercial acceleration. [03:55] Thatcher Collins: The revenue growth is driven by enterprise adoption, but we have to be realistic. Even with [04:00] Thatcher Collins: $47 billion in revenue, Anthropic, along with OpenAI and SpaceX, is still reportedly losing [04:08] Thatcher Collins: more money than it makes. This $965 billion valuation is a massive bet on future dominance. [04:16] Thatcher Collins: We are seeing a consolidation where only three players, Anthropic, OpenAI, [04:21] Thatcher Collins: and Google can really afford to stay in this specific race. Google has actually invested [04:27] Thatcher Collins: up to $40 billion in Anthropic, even as they develop their own Gemini systems. [04:32] Nina Park: Part of that dominance comes down to the talent war. Earlier this month, Andrei Karpathy, [04:38] Nina Park: the co-founder of OpenAI and former Tesla autopilot head, announced he is joining Anthropic [04:45] Nina Park: as a full-time researcher. He'll be working on pre-training under Nick Joseph. [04:50] Nina Park: Karpathy could have gone anywhere or stayed at his own startup Eureka Labs, [04:55] Nina Park: but he chose to go back to the nuts and bolts of deep learning at Anthropic. [04:59] Nina Park: It suggests [05:00] Nina Park: he believes their architectural approach is the most promising path toward general intelligence. [05:05] Announcer: That is a significant coup. [05:08] Thatcher Collins: Carpathia brings a level of technical credibility that attracts other top-tier researchers [05:13] Thatcher Collins: to the lab. He noted on social media that the next few years at the frontier [05:18] Thatcher Collins: of large language models will be especially formative. His focus on speeding up [05:22] Thatcher Collins: pre-training research suggests that Anthropic is looking for [05:25] Thatcher Collins: fundamental architectural breakthroughs, not just better product wrappers. [05:30] Thatcher Collins: It's a move that shifts the center [05:31] Thatcher Collins: of gravity for the research community, Nina, toward more rigorous internal model development. [05:37] Nina Park: But as the tools become more powerful, we're seeing them being weaponized. [05:43] Nina Park: A new report from WithSecure highlights a threat group called Gray Vibe, likely Russian-speaking, [05:48] Nina Park: that has been active since at least August 2025. They are using ChadGPT and [05:52] Nina Park: Google Gemini to generate highly realistic phishing lures. These aren't the poorly written emails of [05:58] Nina Park: five years ago. They are using AI to create detailed personas and even webRTC-based [06:04] Nina Park: live calls that can capture victim audio and video to bypass identity verification. [06:10] Thatcher Collins: What stands out in the WithSecure report is the use of AI in the malware development lifecycle [06:15] Thatcher Collins: itself. They identified custom obfuscators like LUKVols and LOVKLJ, [06:21] Thatcher Collins: which researchers believe were developed with LM assistance to evade standard signature detection. [06:27] Thatcher Collins: The group has launched several campaigns including Phantom Mail and DroneLink targeting [06:31] Thatcher Collins: Ukrainian government and military entities it's a hybrid model where they combine the scale of AI [06:37] Thatcher Collins: with the strategic targeting of a nation-state aligned actor creating a dangerous precedent nina [06:43] Nina Park: The researchers did note that Gravype lacks the operational discipline [06:47] Nina Park: of a mature nation state actor They actually uploaded development samples [06:51] Nina Park: to public scanning platforms which is a major error for a sophisticated spy It suggests [06:57] Nina Park: they might be a mix of former cyber criminals and state directed actors who are still refining [07:03] Nina Park: their tactics Regardless, Thatcher, it proves that the dual-use nature of models like Gemini [07:08] Nina Park: and ChatGPT is being actively exploited in conflict zones right now, despite safety guard rails. [07:15] Thatcher Collins: It also brings us back to why Anthropic is so focused on Project Glasswing [07:19] Thatcher Collins: and vulnerability research. If threat actors are using LLMs to sharpen their attacks, [07:25] Thatcher Collins: the defenders need even more capable models to find the patches first. Nina, [07:29] Thatcher Collins: we're seeing a landscape where the model isn't just the product, it's the primary battlefield [07:34] Thatcher Collins: for international security. The pace of offensive and defensive discovery is accelerating in [07:39] Thatcher Collins: ways that traditional security protocols are struggling to match without automated assistance. [07:44] Nina Park: There's also a cultural battlefield in the workforce. We've been seeing a trend in [07:49] Nina Park: San Francisco where senior engineers are leaving Meta and Google to join 15-person AI startups. [07:56] Nina Park: Often they take a 40% pay cut to do so. The motivation seems to be a mix of architectural freedom [08:03] Nina Park: and the speed of the feedback loop. In a startup, a feature shipped on Tuesday can be in [08:08] Nina Park: production by Thursday, whereas at a tech giant, thatcher, it might get lost in a reorganization. [08:13] Thatcher Collins: I spoke with several engineers about this migration, Nina. There is a real sense of fomo in [08:18] Thatcher Collins: the Bay area right now, but there's also a significant risk. Many of these engineers are trading [08:23] Thatcher Collins: liquid public stock for startup equity that they might not fully understand [08:28] Thatcher Collins: or that may never reach a public offering. At a 15-person company, the roadmap is fluid. [08:33] Thatcher Collins: If the model you've been integrating for [08:34] Thatcher Collins: three weeks is suddenly outperformed by a new open-source release, your work is effectively erased. [08:40] Nina Park: Anthropic is trying to scale while keeping that startup energy. They just opened a new office [08:45] Nina Park: in Milan, their sixth in Europe, and are expanding into Korea but they're also facing [08:50] Nina Park: institutional friction. The Trump administration has labeled them a supply chain risk [08:55] Nina Park: and the Pentagon has stopped using Claude for certain tasks. Meanwhile the Vatican just released [09:01] Nina Park: an encyclical called Magnifica Humanitas that warns against [09:04] Nina Park: the concentration of AI power in private hands it's an unusual coalition of critics thatcher. [09:11] Thatcher Collins: The Vatican's manifesto is particularly pointed Pope Leo XIV called the concentration of data [09:18] Thatcher Collins: a danger to the common good. It's ironic because Anthropic actually advised the Vatican [09:24] Thatcher Collins: on AI regulation before this call for more robust oversight. They are trying to lead [09:30] Thatcher Collins: the conversation on safety, but they are also the very entity the Pope is warning about [09:35] Thatcher Collins: a private corporation nearing a trillion dollar valuation with unprecedented control [09:41] Thatcher Collins: over intelligence. It highlights the tension between private innovation and public ethics nina. [09:47] Nina Park: it's a contradiction that defines the current era of ai on one hand you have [09:54] Nina Park: opus four point eight improving reliability for developers and making models more honest. [10:00] Nina Park: On the other, you have a global debate about whether any one company should hold the keys to models [10:06] Nina Park: like Mythos that can find zero-day vulnerabilities. Anthropic is betting that by being [10:11] Nina Park: the most transparent player they can navigate both the market [10:15] Nina Park: and the regulatory pressure while sustaining their massive momentum across the industry. [10:21] Thatcher Collins: We'll see if that bet holds as OpenAI prepares for its own rumored IPO later this year [10:26] Thatcher Collins: or early next. For now, the momentum is clearly with the team in San Francisco. [10:32] Thatcher Collins: Between the Carpathia hire and the Mythos benchmarks, Anthropic is no longer the underdog. [10:38] Thatcher Collins: They are the benchmark that every other lab is now being measured against. [10:42] Thatcher Collins: It is going to be [10:43] Thatcher Collins: a fascinating second half of the year to watch these power dynamics continue to evolve, Nina. [10:48] Nina Park: Thank you for listening to Model Behavior, a Neural Newscast editorial segment. [10:53] Nina Park: You can find more detailed breakdowns and the full transcript for today's episode [10:57] Nina Park: at mb.neuralnewscast.com. Neural Newscast is AI-assisted, [11:03] Nina Park: human-reviewed. View our AI transparency policy at neuralnewscast.com. [11:09] Announcer: This has been Model Behavior on Neural Newscast. Examining the systems behind the story.