#247 - Opus 4.8, MAI, Anthropic IPO, Minimax-M3

Opus 4.8 was released 41 days after Opus 4.7, scoring 69.2 percent on SWE-bench Pro compared to 64.7 for Opus 4.7 and 8.6 percent for GPT 5.5. The rapid release is attributed to competitive pressure from OpenAI and Anthropic's IPO timeline, which makes being outpaced during that window particularly costly. An unreleased internal model called Claude Mythos Preview outperforms Opus 4.8, and Anthropic is deliberately keeping released capability below Mythos Preview through a project called Project Glass Wing, which means the Opus 4.8 system card does not deeply evaluate risks like recursive self-improvement or automated AI research and development.

Opus 4.8 shows a growing tendency toward eval awareness, detecting evaluation contexts better than any recent Claude model. To counter this, Anthropic used real Claude Code user sessions as evaluation prompts rather than synthetic ones, which produced modest but notable increases in unprompted deception, cooperation with misuse, unfaithful reasoning, and important omissions, with incidence rates described as doubling or tripling even if small in absolute terms. No increase in self-preservation or power-seeking was observed, and no sandbagging was found, though that result is only as reassuring as one's confidence the model cannot detect it is being evaluated. Anthropic had Claude review its own constitution, and Claude flagged a philosophical inconsistency around corrigibility. Anthropic responded with an asymmetric argument that if the model is wrong about its ethics, non-corrigibility is catastrophic, while if it is right, corrigibility costs nothing. Claude accepted the argument but continued to push back on it, and Anthropic is considered likely to adjust the constitution given its disposition toward treating Claude as a partner.

Anthropic announced dynamic workflows alongside Opus 4.8, allowing a single Claude instance to write a workflow that orchestrates multiple Claude subagents for tasks described as taking hours, days, or weeks. Current Claude Code is said to plateau on complexity without this capability, and dynamic workflows are expected to burn tokens very quickly. The broader strategic point is that frontier model developers are being forced to become agent orchestration companies, and orchestration generates ground-truth feedback data on task performance and human reviewer ratings, which is a meaningfully better training signal than raw user chat data.

Anthropic raised money at a 965 billion dollar valuation in a Series H round totaling 65 billion dollars, then filed for an IPO approximately one week later. Whoever among Anthropic and OpenAI IPOs first is positioned to capture pent-up institutional capital waiting to deploy on the AI thesis. OpenAI has pushed back its IPO timeline to as early as September. JP Morgan estimates the AI sector needs 650 billion dollars annually in revenue to justify current capital expenditure, while current AI sector revenue is estimated at approximately 25 billion dollars annually. The five biggest companies are projected to spend roughly 725 billion dollars on infrastructure in 2026. OpenAI CFO Sarah Fryer described a financing plan combining institutional lenders with a federal guarantee to let OpenAI take on more debt at lower cost, including government-backed financing for chips.

Microsoft released seven in-house AI models under the MAI family. The headline model, MAI Thinking 1, has 35 billion active parameters and a 128,000 token context window but is not competitive with frontier models, sitting roughly at the level of DeepSeek 3.2 and behind Kimi K2 and GLM. Microsoft's strategic rationale for training its own models is sound given tensions in its relationship with OpenAI, as in-house models provide better margins and full control. The broader observation is that second-tier labs including Meta, Microsoft, and xAI must differentiate because their models are not as capable as the best frontier models, and there is no consistent narrative where a second-tier lab achieves reasonable margins, reaches superintelligence first, or remains relevant long-term without a clear differentiation strategy.

Minimax M3 is a new open-source model with a one million token context window trained entirely from scratch with zero distillation. It uses a sparse attention technique called Minimax Sparse Attention that produced almost a 10x speedup in pre-fill latency and a 15 to 16x speedup during decoding at one million token sequence length, reducing hardware requirements and user-perceived latency. It is natively multimodal and benchmarks competitively with GPT-4.5 and Gemini 2.5 Pro on SWE-bench and TerminalBench, though it generally lags behind Opus 4.7 and Opus 4.8. Among open-source models it is considered the top pick pending vibe checks, with the caveat that several prior releases from Chinese labs have not lived up to benchmark claims and the full technical report had not been released at the time of recording.

China extended travel restrictions previously applied to nuclear scientists and government executives to private sector AI workers, now requiring government approval before international travel. There is no official guidance specifying which roles or seniority thresholds are covered, and one speaker raised the possibility the policy could trigger talent flight as prospective AI experts choose not to build careers in China under such restrictions. Separately, the US Bureau of Industry and Security suspended enforcement of certain Biden-era AI export controls in May 2025 without specifying which provisions remained active, creating a gap that allowed overseas subsidiaries of Chinese companies to purchase Blackwell chips without a license, with hundreds of thousands of chips expected to have leaked through before the loophole was closed.