OpenAI passes Anthropic in OpenRouter spend after Astra launch
GPT-6 Astra led dollars spent last week, while cut-price GPT-5.6 Luna dominated token volume across the two model families.
By Ryan Merket · Published
Primary source: Peter Walker on X
Why it matters
OpenAI's first weekly spend lead over Anthropic on OpenRouter since 2024 shows Astra winning premium workloads while Luna's pricing drives scale. One model family is competing for both margin and volume.
OpenAI models drew more spending than Anthropic models on OpenRouter last week, according to data posted on X by Peter Walker (@PeterJ_Walker), OpenRouter's head of insights. GPT-6 Astra took the largest share of dollars, while the far cheaper GPT-5.6 Luna processed the most tokens by a wide margin.
https://x.com/PeterJ_Walker/status/2099914428376564027
The split captures the two markets OpenAI is pursuing at once. Astra is selling expensive, frontier-level inference for difficult jobs. Luna is priced for applications that need enormous volumes of routine model work. Counting tokens crowns Luna; counting money crowns Astra.
Walker said OpenAI last surpassed Anthropic in weekly spending on OpenRouter on February 26th, 2024. For the week of September 7th through September 13th, Astra accounted for 19% of dollars spent across the two companies' models, ahead of Anthropic's Claude Opus 5 at 16%. OpenAI's GPT-5.6 Sol and Luna each captured 10%, followed by Anthropic's Claude Sonnet 5 at 7% and Claude Fable 5.1 at 6%, according to Walker.
Those figures measure traffic routed through OpenRouter, rather than either lab's total API or subscription business. Anthropic and OpenAI also serve customers directly, and neither publishes a model-by-model breakdown comparable to OpenRouter's data.
Astra sells scarce capability
OpenAI released Astra on September 4th, giving the model little more than a week to reach the top of Walker's spending table. OpenRouter lists Astra at $10 per million input tokens and $50 per million output tokens. OpenAI describes it as its flagship model for long-running work spanning software engineering, research, computer use and professional tasks.
Astra's rates are 50 times Luna's input price and about 42 times its output price. That gap means Astra can lead spending without approaching Luna's token volume. Its early wallet share shows that some OpenRouter customers were willing to move paid workloads to the new model almost immediately, though the chart does not establish whether those workloads stayed with Astra after launch-week testing.
The timing also matters. New frontier releases routinely receive an initial burst of evaluations, migrations and side-by-side tests. Walker's data establishes that Astra converted that attention into billed usage during its first full week. A longer series will be needed to separate durable production traffic from launch demand.
Luna wins the volume contest
Luna's token lead follows an aggressive price cut. OpenAI reduced Luna's API rates by 80% on July 30th, lowering them to $0.20 per million input tokens and $1.20 per million output tokens. The model had been released on July 9th.
OpenAI positions Luna as the GPT-5.6 option for cost-sensitive, high-volume jobs. Its model documentation points to a 1.05 million-token context window, reasoning controls and support for tools, giving developers a lower-cost model that can still participate in agent workflows.
The price reduction made it economical to use Luna for classification, chat, lightweight agents and repeated subtasks where Astra's higher capability would be difficult to justify. Luna's resulting token lead is a distribution win: developers appear to be finding enough work for the inexpensive model to keep it running at vastly greater volume.
OpenRouter's public catalog currently shows Luna with substantially more cumulative traffic than Astra. That total includes Luna's two-month head start, so it cannot be treated as a direct launch comparison.
Token rankings need a warning label
OpenRouter's ranking methodology counts prompt and completion tokens processed through its API. It excludes requests that customers keep private and does not cover calls made directly to model providers.
OpenRouter also warns that token volume does not measure request counts, spending, quality or completed work. Models vary in verbosity, caching behavior and tokenization. Its dataset documentation says token counts are reported using each upstream provider's tokenizer, which limits direct comparisons between laboratories.
Wallet share carries its own limitation. A costly model accumulates spend faster by design, even if it handles fewer jobs. Astra's lead therefore shows where OpenRouter customers paid the most, while Luna's lead shows where the largest measured volume flowed. Neither figure proves which model completed more useful work.
That gap is becoming central to model procurement. Teams choosing an API increasingly need cost per completed task, including retries and failures, rather than a raw token price or leaderboard position. Walker's chart shows why: OpenAI can simultaneously win the premium end of spending and the commodity end of consumption with two products priced more than an order of magnitude apart.