Complete speculation here: Mythos and Spud are the first generation of polished GPT4.5-sized reasoning models.
Speculation suggests new models from Anthropic and OpenAI could bring a major performance jump at a premium price.
A viral theory circulating in AI communities speculates that the rumored models 'Mythos' (from Anthropic) and 'Spud' (from OpenAI) represent a significant shift. They are posited to be the first consumer-facing models built at the scale of the legendary, unreleased GPT-4.5. The original GPT-4.5 was known for its deep, nuanced reasoning and ability to make high-level abstract connections—qualities that some feel were sacrificed in later models like GPT-4o for better usability and cost. The speculation suggests that by applying today's advanced reinforcement learning from human feedback (RLHF) techniques to a model of that size, the result would be a 'polished' AI with both profound reasoning and practical utility.
This performance leap, however, is expected to come at a steep price. Running such large models is computationally expensive, leading to predictions of API costs as high as $25 per million input tokens and $80 per million output tokens. This would effectively turn a Claude subscription into a strictly rate-limited service. The theory further posits a strategic divergence: Anthropic, with limited compute, might make 'Mythos' API-only or heavily restricted, while OpenAI could leverage its scale to offer 'Spud' more broadly, even at a loss, to capture the market. The core concern is that this marks a return to an era where 'size matters,' potentially locking the most capable AI behind a paywall that only corporations can afford, widening the access gap for individual consumers and researchers.
- Models rumored to be built at GPT-4.5 scale, combining its abstract reasoning with modern RLHF for a major performance jump.
- Expected to be extremely expensive to run, with speculated API costs of ~$25/M input tokens and ~$80/M output tokens.
- Could create a two-tier access system, where the most powerful AI is only viable for corporate use via API, not consumers.
Why It Matters
This signals a potential future where the most advanced AI capabilities become prohibitively expensive, primarily serving enterprises over individual users.