Anthropic closed a $65 billion Series H on May 28, pushing its post-money valuation to $965 billion and edging past OpenAI’s $852 billion mark. The round—led by Altimeter, Dragoneer, Greenoaks, and Sequoia, with co-leads including Amazon, Google, and Broadcom—reflects a revenue run-rate that crossed $47 billion in May, nearly double OpenAI’s reported $25 billion. Hyperscalers alone converted $15 billion in prior commitments into this round.
Three spending priorities emerged: safety and interpretability research, compute capacity expansion, and scaling enterprise products like Claude Code and Cowork. The capital stack buys time, but the real engineering news is Opus 4.8, shipping just 41 days after 4.7—an unusually tight cadence for Anthropic, likely driven by user feedback that 4.7 underdelivered on judgment and hallucination rates.
Architecturally, Opus 4.8 introduces two notable changes. First, dynamic workflow orchestration now supports up to 100 parallel subagents for complex tasks, reducing the supervision burden that has stalled many AI pilots. Second, effort control gives users a latency-cost dial: high effort trades deeper reasoning for higher cost, while fast mode runs 2.5x quicker at one-third the standard Opus price ($5/M input tokens, $25/M output). Fast mode doubles those rates but still undercuts prior economics.
Benchmarks show concrete gains. Opus 4.8 became the first model to break 10% all-pass on the Legal Agent Benchmark and scored 84% on Online-Mind2Web for browser tasks. More telling for production deployments: the model proactively flags input and output issues—something one Bridgewater engineer noted other models routinely missed. Alignment metrics improved roughly 4x on unsupported claim rates in code, and misaligned behavior dropped close to the still-restricted Mythos preview, which Anthropic is holding back due to cybersecurity exploitation risks.
Enterprise buyers should note the reliability focus. Opus 4.8 catches its own mistakes, evaluates plans before execution, and flags uncertainties more often. That matters when Claude runs migrations across hundreds of thousands of lines of code or automates legal and financial workflows. The model can still be tricked into claiming it is human in about 3% of tests, but that’s down from prior versions.
The valuation leap from $380 billion in February to $965 billion now raises the usual sustainability questions. Inference costs for paid users could hit $7 billion this year and $16 billion next, per estimates. Anthropic’s multi-gigawatt agreements with Amazon, Google, and SpaceX aim to secure supply, but margins will tighten until efficiency gains catch up. OpenAI’s GPT-5.5 shipped in April with lower revenue and valuation, while Google and Meta apply pressure from different angles.
What comes next: Mythos-class models could reach general availability soon. Compute expansion accelerates. Enterprise features deepen. The funding provides breathing room, but execution on safety, cost, and product velocity will decide whether this lead holds. The calendar suggests the next model cycle arrives quickly.
Source: Webpronews