#251 - Mythos Back, Sonnet 5, Etched, LongCat

Anthropic's Mythos model was blocked from global release by the US Commerce Department using export controls, and after negotiations the company agreed to add new classifiers targeting cybersecurity tasks. Jeremy Harris argues these safeguards are likely things Anthropic would have implemented anyway, and the outcome looks more like giving the administration a political win than delivering genuine marginal security value. Mythos Five remains restricted to specific US government entities involved in defending critical infrastructure, while Fable Five was cleared for general commercial release. The US government also made a preemptive voluntary hold request to OpenAI for GPT 5.6, described as the first such request to a frontier lab, and that hold had not been lifted even after Fable was cleared, giving Anthropic recovered lead time. Harris states it is effectively guaranteed that US adversaries including China will find ways to jailbreak Mythos and similar models, and no one knows how to stop jailbreaks at scale. The specific jailbreak that triggered the original Mythos concern has still not been publicly identified or explained.

One speaker argues the administration acted with even-handedness toward both Anthropic and OpenAI, while the other describes the handling as catastrophically incompetent and driven by a desire to set precedents of control rather than genuine safety logic. A principled approach would impose the same release delay on competitors that was imposed on the first mover, but if a blanket capability threshold triggers a hold only for whoever reaches it first, there is no incentive to keep pushing the frontier. Anthropic is now scaling up coordination with the US government including peer model access for evaluation and is co-drafting a consensus framework with Amazon, Microsoft, Google, and others for assessing AI jailbreak severity.

Claude Sonnet 5 launched at $2 per million input tokens and $10 per million output tokens through August 31, rising to $3 and $15 after that date, making it cheaper than Opus 4.8, GPT-5.5, and Gemini 3.1 Pro at launch pricing. Sonnet 5 shows significant improvement over Sonnet 4.6 on agentic coding benchmarks, misaligned behavior is down, and on knowledge work benchmarks it exceeds Opus 4.8, meaning it is beginning to cannibalize the higher-tier model. According to Artificial Analysis, however, Sonnet 5 uses more tokens per task than Fable 5 did, making it more expensive on a cost-per-intelligence-unit basis despite lower per-token pricing. GPT-5.6 scores above Sonnet 5 on Anthropic's own cyber evaluation benchmark, which Anthropic presents as evidence that model is more dangerous.

Etched has raised $800 million across four rounds and reports over $1 billion in signed customer contracts, with first rack shipments planned for summer. The company completed an A0 tape out on TSMC's N4P four-nanometer process in under three years from seed funding and shipped the chip on the first attempt without re-spins. Etched is pivoting from a transformer-only ASIC pitch to a full-stack systems and manufacturing company that can also run mixture-of-experts models including DeepSeek, Qwen, Mamba, and Llama. The company is arguing for a $5 billion valuation without having shipped a single unit, and speakers note that signed contracts are not the hard part, with delivery quantities, timelines, and yields all unproven. TSMC allocation is described as the key competitive factor, and TSMC has an incentive to support Nvidia competitors to sustain long-term demand for its nodes.

LongCat 2.0 from Meituan has 1.6 trillion total parameters with 4.8 billion activated per token, was trained on 35 trillion tokens, and benchmarks claim comparability to Gemini 3.1 Pro. Training and large-scale deployment were built entirely on AI ASIC super pods, described as the first publicly reported instance of this approach. Estimated training scale is approximately 40 to 80 megawatts, roughly comparable to DeepSeek V3 Pro at around 50,000 chips, though Chinese chips convert megawatts to compute less efficiently than Western chips, so raw power figures overstate relative compute. The stable long training run is interpreted as evidence of robust engineering repeatability. Chinese open source frontier models now include DeepSeek, Kimi, GLM, and LongCat, and GLM 5.2 has separately gained traction in the developer community as a viable competitor for coding tasks, described as a first for a Chinese model.

DeepSeek raised approximately 50 billion yuan, equivalent to $7.4 billion US dollars, in one of China's largest ever startup funding rounds, and plans to at least double headcount with an emphasis on data and development engineers, signaling a shift from pure research lab toward product organization. The Chinese Communist Party reportedly signaled to China's AI industry to stop recruiting from DeepSeek to protect it as a national champion. Taiwan's Keelung District Prosecutor's Office raided Supermicro's Taiwan office and two supply chain partners over the diversion of approximately 50 servers containing top Nvidia chips to China using falsified customs declarations, causing Supermicro shares to fall 8 percent. Taiwan is considering new legislation to restrict AI chip sales to all Chinese customers rather than only blacklisted firms, and the situation is being described as a live test of whether AI export controls actually work in practice.