Opus 5.5 usage is up 3.6x, but OpenRouter's top spot is still DeepSeek

Let's start with the OpenRouter data. Peter Walker posted a chart showing that the 7-day average token volume for Anthropic models on OpenRouter has risen 3.6x since Opus 5.5 launched. Feedback on Sonnet 5.5 and Haiku 5.5 is not bad either. When Andrew Curran reposted it, he said he doesn't take money from any lab and only talks about what he sees: over the past two weeks, he has never seen so many people on his timeline say they are switching to Anthropic. He added that even if someone catches up with Opus, Fable 5.5 is already ready.

The chart marks three milestones: Opus 5.5 on September 21, Sonnet 5.5 on September 27, and Haiku 5.5 on October 15. The curve rises after the end of September and continues climbing in October.

Peter Walker later gave a more direct number: the 7-day average token volume for Anthropic models is 3.6x what it was before Opus 5.5 launched. Someone in the comments asked whether cache reads count, and Peter replied that they do and that cache makes up a very high share; however, the cache share was also high before launch. That's a useful caveat. In an agent loop, context is resent every round, so token volume will naturally run ahead of billing. When looking at growth, you can't focus only on total tokens.

The data on the enterprise spending side is harder. The Ramp AI Index, which tracks transaction data from 85,000 U.S. businesses, shows that in the week ending October 11, Anthropic accounted for 60.1% of enterprise AI token spending on Ramp, up from 53.8% in the same period a month earlier. OpenAI fell from 42.4% to 36.4%. In the same week, total token spending rose 23.0% and total token volume rose 52.4%. Anthropic's blended price per million tokens fell from $1.00 to $0.80, a decline of 19.2%. Lower price, higher usage, and higher share—this combination is more informative than token volume alone.

But don't rush to write 'Anthropic has won the market.' OpenRouter's weekly leaderboard tells a different story. Ranked by token throughput, DeepSeek V4.1 Flash 1017 is first at 33.6T, up 801% week over week; Claude Opus 5.5 is at 8.27T, ranked seventh, up 34% week over week. Ranked by text request share, DeepSeek has 26.5%, Google 20.1%, OpenAI 18.5%, and Anthropic 5.5%. Anthropic's request share is up 25% week over week—not a low growth rate—but its absolute base is still small. On OpenRouter, a large number of free and low-cost models can push total token volume very high, and that is not the same test as enterprise paid spending.

A commenter posted another chart showing that Opus 4.6 once accounted for 38.8% of spending but only 27% of tokens, while 5.5 is the first version where the two bars are relatively aligned. This detail is more worth looking at than slogans like 'supremacy.' When spending share and token share line up, it suggests users aren't just using it for cheap tasks.

Andrew Curran's read on the reason is taste and creativity. He said that in September, new internal Fable/Mythos training produced something that was passed to Opus 5.5. When someone asked why Fable 5.5 hasn't been released yet, he replied that it is already ready and has been for about two weeks; Opus 5.5 is already winning now, and releasing it later preserves the advantage. He also cited Tibo's view: the period when OpenAI kept Astra internal was their biggest development advantage. That claim can't be verified from public data, but it explains why Anthropic isn't in a hurry to play all its cards.

Dissent is on the same timeline. E.A. Bruce said OpenAI and Anthropic are trading the lead, Astra 6.2 may jump back up in a month, and we've seen this movie before. Camiel pointed out that on Artificial Analysis's GDPval-AA, Opus 5.5 is at 1866 Elo, Muse Wonder Full at 1857, and Kimi K3.1 Pro at 1846—less than 20 points separate the top three, so the lead could flip quickly. M argued that Anthropic has a brand premium and that Sonnet often outsells OpenAI models with better performance and lower prices. SRCL complained that Opus 5.5's web search is lazy, and Andrew acknowledged that Grok and Gemini's combined search is stronger and can see things from a few minutes ago.

Put together, the picture isn't complicated: Anthropic has gained strong momentum in enterprise paid spending and coding-agent scenarios, and OpenRouter token volume signals the same direction. But total request share, benchmark gaps, cache share, and the search weakness all remain. Fable 5.5 not being released is the biggest variable. It's too early to say 'regime change,' and it's also inaccurate to say 'it's just cache inflation.' A more solid judgment is that Opus 5.5 has pushed Anthropic's paid base up another notch, and next we need to watch two things: when Fable 5.5 makes its move, and whether the competition's next update can close the 20-point gap.

Data sources:
Andrew Curran's tweet
Peter Walker's tweet
OpenRouter Rankings
Ramp AI Index
Lisan al Gaib's tweet