#249 - Fable 5 ban, SpaceX Cursor + IPO, OSS Aplenty

The US government issued an order directing Anthropic not to allow any foreign national access to Claude 4, referred to as Fable Five, or its associated model Mythos Five, including foreign nationals located inside the United States. Anthropic had spent months coordinating with the White House on the Fable Five launch, with safety safeguards central to the release, and the two parties had agreed on deployment protocols before the White House reversed course. The trigger appears to have been Amazon CEO Andy Jassy contacting the White House to report jailbreaks found in Mythos Five, though the speakers note this was not a universal jailbreak but a specific instance in a specific codebase. Anthropic's position was that the exploit was a single highly directed case and that any frontier model could be made to do the same thing. The speakers argue no compelling case exists that jailbreaks from frontier models can ever be fully prevented, meaning the policy standard being applied cannot be met by any lab.

David Sacks publicly framed Fable Five as Mythos without guardrails, but the speakers counter that Mythos class model vulnerabilities were real and included critical weaknesses in load-bearing pillars of the internet and most critical infrastructure. The executive order on AI explicitly states its requirements are voluntary, and Sacks himself reportedly forced that voluntary language into the order, creating a direct tension with the administration's current enforcement posture. The speakers characterize the jailbreak rationale as a cover story and point to administration comments suggesting Anthropic's deference to the government was itself a factor. The situation creates a perverse incentive structure where safety investment becomes a liability. Andrei Karpathy, who recently joined Anthropic, would himself be unable to access Mythos Five under the order due to his status as a foreign national. A follow-up development is expected within days based on informal reporting suggesting conversations are ongoing about bringing Fable Five back online. The speakers warn that a genuine victory for AI safety advocates is premature because the foundational principle will not be truly tested until the same action is taken against OpenAI when it develops a comparable model.

SpaceX went public via IPO in the same week, listing at approximately 1.75 trillion dollars and rising above 2 trillion dollars after trading began. SpaceX simultaneously announced the acquisition of AI coding startup Cursor for 60 billion dollars, likely paid primarily in SpaceX shares. Cursor's market share in AI coding tools dropped from approximately 41 percent in June 2025 to about 26 percent in May 2026, driven largely by users switching to Claude Code and similar tools. Cursor never pre-trained its own models from scratch, relying on fine-tuning on top of open source models. Cursor had been using Colossus 2, SpaceX's large compute cluster, for training runs prior to the acquisition, and the speakers suggest SpaceX exercising its option to buy rather than declining indicates those training results justified the 60 billion dollar price.

Leaked OpenAI financial documents show revenue growing from approximately 3.7 billion in 2024 to approximately 13 billion in 2025, while total operating losses grew from approximately 9 billion to approximately 20 billion over the same period. Gross margins improved from 28 percent to 43 percent as the business scaled, though both figures remain well below typical SaaS gross margins of 70 to 80 percent, making the economics more utility-like. Anthropic margins are reportedly better than OpenAI margins and potentially by quite a bit. ChatGPT market share fell below 50 percent for the first time, reaching 46.4 percent by end of May 2026, while Gemini holds 27.7 percent and Claude holds 10.3 percent. The speakers characterized market share erosion as a bigger concern for OpenAI's IPO prospects than its spending trajectory. Anthropic converts 13 percent of its users to paid subscriptions, described as the highest conversion rate in the field, representing a meaningful headwind for OpenAI in high-margin enterprise revenue.

Moonshot AI released Kimi K2.7-Code, a coding-specific mixture of experts model with one trillion total parameters and 32 billion active per token, under a modified MIT license, priced at approximately one sixth the cost of Claude Opus 4.8. Open source coding models are not yet at the level of GPT 5.5 and Claude Opus 4.8 but are within striking distance and could start eating into frontier lab revenue if cost and speed improvements continue. Nvidia released Nemotron Ultra, a 550 billion parameter hybrid Mamba transformer model, and published a detailed 50-page technical report covering hyperparameters, data mixtures, and post-training recipe. The speakers noted Nvidia is deliberately steering the open source ecosystem toward its own hardware as an inference-side moat.

OpenRouter's new Fusion feature dispatches prompts to a panel of models in parallel and uses a judge model to synthesize results. Pairing Claude Opus 4.8 with itself outperforms solo Opus 4.8 by six to seven percentage points on benchmarks, and swapping one instance for GPT 5.5 yields an additional roughly 2.1 percentage point boost from model diversity. Because Anthropic and OpenAI will not sell multi-model ensemble routing themselves, OpenRouter holds a structural advantage as a provider to providers.