Anthropic targets $2T valuation in IPO that could surpass SpaceX's record

Anthropic eyes $2T IPO that could surpass SpaceX's record $86B debut The Claude maker confidentially filed Form S-1 and could go public as soon as the end of August, working with Morgan Stanley, Goldman Sachs and JPMorgan. Bloomberg reports Anthropic expects to match or beat SpaceX's $86.2B offering, with preliminary Q2 revenue above $11.5B. 12

Broadcom seeks over $60B in debt to finance AI chips for Anthropic Bloomberg reports talks with lenders including Apollo and Blackstone could reach $100B, with a special-purpose vehicle leasing hardware to Anthropic. The financing extends the AI XPV partnership's plan to fund 20GW of compute for leading labs by 2028. 3

OpenAI cuts GPT-5.6 Sol API prices by more than 20% Starting August 21, Sol drops to $4 input and $20 output per million tokens for three months, following cuts to Terra and Luna. The move comes as Anthropic and Chinese labs undercut OpenAI on enterprise API pricing; Amazon Bedrock also announced reduced pricing. 456

DeepSeek releases V4-Flash-Vision-Exp multimodal model with Files API The experimental model matches V4-Flash on text while sharply improving multimodal agent benchmarks, nearing Opus-4.8. Image inputs are billed as tokens (max 384 per image) at V4-Flash prices, and a free Files API lets developers reuse uploaded images. 7

Microsoft receives first production Nvidia Vera Rubin systems Satya Nadella confirmed delivery of Vera Rubin NVL72 systems, pairing 88-core Vera CPUs with Rubin GPUs featuring 288GB HBM4. Nvidia claims up to 5x faster inference and 3.5x better training versus Blackwell, with CoreWeave reporting 10x throughput per megawatt. 8

Anthropic puts Mythos 5 in defender tools, adds $35M open-source fund Claude Mythos 5 now scans code in Claude Security and partner products, returning patches and alerts without giving users direct model access. Anthropic also committed $35M in credits to open-source security work ahead of Europe's Cyber Resilience Act reporting deadline. 9

Google's open Gemma models pass one billion downloads Google DeepMind says developers have published over 100,000 Gemma variants in two years, with NASA and startups running Gemma in orbit. The figure trails Alibaba's Qwen claim of 3 billion downloads but excludes Android and Chrome integrations. 10

Google DeepMind pushes AI code generation with formal proofs in Lean A new 'Verified Code Generation' effort wants AI systems to produce machine-checkable proofs alongside code, building on AlphaProof's Lean work. Formal verification could catch hidden errors that testing misses, but depends on accurate specifications. 11