Anthropic's IPO materials claim Mythos did nearly all the kernel work to bring Claude up on AMD MI355X chips, a job it says used to take months. At a JPMorgan event before the listing, Dario Amodei called buying the right amount of compute the company's hardest business problem. For xAI, Grok Bot staff say harness changes give subscribers about 7% more work per plan after a week of usage complaints. OpenAI had a quiet weekend.
Speaking with Jamie Dimon at JPMorgan's Manhattan headquarters on Nov. 13, Anthropic's CEO said it is hard to forecast how much compute to buy a year or two ahead: too little leaves customers unserved, too much strains the finances. According to press reports of the briefing, he said strong gross margins and cash on hand cushion that risk, that securing compute is a main reason for listing now, and that the company aims to underpromise in its IPO.
Why it matters: Anthropic's filings show about $900 billion of commitments, mostly for compute, so how it manages that bet is central to the IPO pitch.
Grok Bot team member Sam Sokolin said on Nov. 13 that efficiency changes to the agent's harness let subscribers do roughly 7% more work before hitting limits, and up to 15% more on routine runs; Elon Musk highlighted the change on Nov. 14. It follows a week of complaints that paid allowances ran out within hours, and a weekly limit reset on Nov. 9.
Why it matters: Usage limits have been the main complaint about Grok Bot, and this is a modest step rather than a fix.
Try it: Grok Bot shows usage broken down by bot and source, which helps spot the tasks that use up the allowance.
Grok BotHappened Nov 13, 2026 · Sources: TeslaNorth
In investor materials for its IPO, Anthropic says Claude Mythos ported or wrote nearly all of the GPU kernels needed to run Claude on AMD MI355X chips in about three days, its first use of AMD hardware. It says bringing up a new chip platform has usually taken dozens of engineers six to 12 months. The figure covers prompt processing at a usable performance level and has not been checked independently.
Why it matters: If it holds up, AI-written kernels make it easier for labs to move between Nvidia, AMD, Google and Amazon chips instead of being locked to one vendor.