Worth passing on?
Select any line in the story to share just that line. Or keep it for later and find it under Saved.
Read the original at The Register
theregister.com
Fetching the latest from every source…
Anthropic Claude subscription plan provides more value than OpenAI's, study says
The rivalry between Anthropic and OpenAI, the two leading US AI model makers, is no contest when it comes to assessing which one gives you more value for your subscription money, according to industry research firm SemiAnalysis.
4 min read
Key points
- The rivalry between Anthropic and OpenAI, the two leading US AI model makers, is no contest when it comes to assessing which one gives you more value for your subscription money, according to industry research firm SemiAnalysis.
- "Anthropic is an overwhelmingly better deal, offering ~5x the API-equivalent value across the board," SemiAnalysis claims in reference to mid-tier models GPT-6.1 Sol and Claude Opus 5.5.
- Company researchers contend that API-equivalent value is the best figure for capturing the value of a subscription plan, though they allow that other measures like raw token volume may be more appropriate in some instances.
The story
The rivalry between Anthropic and OpenAI, the two leading US AI model makers, is no contest when it comes to assessing which one gives you more value for your subscription money, according to industry research firm SemiAnalysis. Following changes to OpenAI's $200 per month subscription plan that effectively halved its value, SemiAnalysis looked at the various subscription plans offered by AI companies, because that's still the main way that individuals and small businesses buy AI services. AI subscriptions are heavily subsidized – customers get far more tokens for their money than they would paying metered API rates. And SemiAnalysis argues that it's necessary to delve into subscription pricing to evaluate AI shops financially because subscription usage consumes a disproportionate amount of company compute resources. Claude subscriptions, the biz claims, take up about 42 percent of Anthropic's inference compute while generating just 10 percent of its revenue. And those subscriptions provide considerable value. "Anthropic is an overwhelmingly better deal, offering ~5x the API-equivalent value across the board," SemiAnalysis claims in reference to mid-tier models GPT-6.1 Sol and Claude Opus 5.5. "You could argue that this is unfair for [OpenAI] because 6.1 Sol is much cheaper per token than Opus 5.5, but the gap is still massive even if you switch to comparing the number of tokens." OpenAI did not immediately respond to a request for comment. SemiAnalysis' calculation relies on measuring token output rates, converting those into tokens per 5-hour usage window and per month, and then computing the API-equivalent value. Company researchers contend that API-equivalent value is the best figure for capturing the value of a subscription plan, though they allow that other measures like raw token volume may be more appropriate in some instances. API-based comparisons can be slow to account for the rapidly shifting capabilities of models that come to market. The debut of a new open source model, for example, can alter what's possible at a given price. For those with access to subscription plans, the subsidized prices clearly offer real value, so much so that there's an underground market for reselling tokens from pooled subscription accounts in many parts of the world. But there are other ways to measure AI value. Artificial Analysis, for example, tracks cost per task – the estimated dollar cost of completing a specific task. By that yardstick, GPT-6.1 Sol ($0.72) is far more affordable than Claude Opus 5.5 ($5.98), at least for those paying API rates. But that ratio may not be true for different tasks – benchmark performance may not reflect your workload. As measured by tokens, Claude Pro ($20 per month) provides 2.9 billion tokens per month, compared to ChatGPT ($20/mo) at 1 billion tokens per month, according to SemiAnalysis. But models may use more or fewer tokens than others for the same task, and when the output is functional, like code, AI output may achieve a similar result across different models but with differing degrees of efficiency. Anticipating AI costs gets more complicated for businesses, which can't count on subscription subsidies and may not have experience with AI cost management. In a July report, McKinsey said 93 percent of enterprises exceeded their AI budgets. "I think all businesses are seeing their AI costs rise," said Dhruv Amin, co-founder of AI agent biz Skydive, in an interview with The Register. Since the beginning of the year, he said, AI has been getting better at solving certain business problems, and companies have started deploying AI more widely, then noticing their token bills going through the roof. "The labs typically subsidize their prosumer plans with subscription pricing massively," said Amin. "And I think it's part of their marketing funnel because more people are able to use the tools, get familiar with the tools, and then bring those tools to work." But for larger companies, he said, two problems arise. You have more employees than can fit on a subsidized plan and, at scale, costs add up fast. "A huge way that enterprises solve this is being able to make a workload work well on the frontier and then quickly switch it to an open source model that they control and that they can serve at a much lower fraction of the cost," he said. "And so for us internally we were able to drop our AI costs 75 percent plus just by moving them all to Skydive Glide, which is our model router that for the given workload tries to find an equivalent open source model that can handle that workload without a quality drop. And so today we run a lot of our workloads on GLM-5.3 Flash." Amin said that companies adept at AI cost management maintain evaluation benchmarks that reflect their workloads. Companies use those metrics to test new models as they come out in whatever AI harness they're using, to ensure consistency. Looking ahead, he said, he expects companies will trust more of that work to service providers. "So what this means is that if you deploy an agent on top of us, we capture in the background the data of how that agent's actually running in production," he said. "And then we use that to construct the eval data set on the fly." ® Updated Oct 6 at 2322 GMT to clarify that Skydive is the platform that sells the agents, while Anything is the umbrella parent company.
