Anthropic is now worth more than OpenAI. On May 28, 2026, it closed a $65 billion Series H at a $965 billion post-money valuation. That edges past OpenAI's $852 billion. The engine behind the number is Claude Code, the coding agent whose run-rate revenue crossed $47 billion earlier that month.
Here is the part nobody puts on the slide. The exact same monthly AI workload that costs you around $2,500 on Claude Opus and $3,000 on GPT-5.5 costs about $348 on DeepSeek.
You are paying the premium. They are becoming a trillion-dollar company.
This is the AI API pricing war, and it is the single most important line item on your 2026 infrastructure bill.
The $965B number, and where it comes from
Anthropic's Series H raised $65 billion. Roughly $15 billion of that was previously committed capital from hyperscalers, including $5 billion from Amazon announced in April. It was co-led by Altimeter, Dragoneer, Greenoaks, Sequoia, Capital Group, Coatue, and D1 Capital Partners.
OpenAI's last raise was a $122 billion round in March at an $852 billion valuation. So Anthropic didn't just catch up. It passed the company that defined the category.
What changed between Anthropic's Series G in February and now? One thing, mostly: developers kept paying for tokens. Claude Code adoption climbed across enterprise customers, and run-rate revenue hit $47 billion. The round landed the same day Anthropic shipped Claude Opus 4.8, tuned for agentic tasks and coding.
Translation: the valuation is built on output tokens. Your output tokens.
DeepSeek just made the math impossible to ignore
On May 23, 2026, DeepSeek locked in a permanent 75% price cut on its V4-Pro model. Not a promo. A new floor. After the discount window closed on May 31, the standing rate became one quarter of the old price.
The numbers that matter: V4-Pro output now sits at $0.87 per million tokens, down from $3.48. Cache-hit input dropped to fractions of a cent. The headline is the output price, because for any agent that writes code, drafts content, or returns long responses, output is where your bill actually lives.
The per-token math, with no marketing in the way
Published list pricing as of late May 2026, per million tokens:
- DeepSeek V4-Pro: ~$0.87 output
- Claude Opus 4.7: $25 output
- GPT-5.5: $30 output
Now scale it to a real workload. Say your product generates 100 million output tokens a month — a mid-size agent in production, nothing exotic.
| Provider | Monthly Cost (100M output tokens) |
|---|---|
| DeepSeek V4-Pro | ~$348 |
| Claude Opus 4.7 | ~$2,500 |
| GPT-5.5 | ~$3,000 |
That is a 7x gap to Claude and roughly 9x to GPT. Annualized, you are looking at $4,176 versus $30,000 versus $36,000 for the identical token count.
Zoom out across the whole market and the spread is almost comical. Between the cheapest open models and the priciest frontier APIs, the gap now hits 300x on input and 450x on output.
So why does anyone pay the premium?
Because sometimes it's worth it. Frontier models still win on the hardest agentic tasks. Claude Opus 4.8 holds an edge on multi-step coding, long-horizon planning, and self-correction — the stuff where a 3% accuracy bump prevents a production incident that costs far more than the token spread.
But here's the trap: most workloads are not that. Classification, summarization, data extraction, first-draft generation, routing, internal tooling — the bulk of real API traffic is routine. Paying frontier rates for routine work is how the $965B valuation gets funded.
The pattern that wins in 2026 is routing by task: cheap model for the 80% that's routine, frontier model for the 20% that's hard. Teams doing this cut their bills 60-80% without users noticing a quality drop.
What this means for your stack right now
Three concrete moves:
1. Audit your output-token spend, not your input. Output is 5-10x the price of input on premium models and it's where the bill compounds. If you don't know your output-to-input ratio, you don't know your real cost structure.
2. Benchmark the cheap model on your actual tasks. Not on a leaderboard — on your prompts, your data, your eval set. DeepSeek V4-Pro and other open-weight models clear the bar for a shocking share of production work. The only way to know is to run it.
3. Build a router, not a religion. Loyalty to a single lab is the most expensive habit in AI engineering. The cost-effective architecture sends each request to the cheapest model that passes your quality gate.
The pricing war isn't slowing down. DeepSeek's cut forces a response. When the floor drops 75%, the premium players either justify the gap with capability or quietly follow the price down. Either way, the developer who's paying attention wins.
Full analysis with FAQ: news.skila.ai/articles/
Top comments (0)