Despite achieving performance parity, OpenAI has raised prices by 2.5 times against the trend: Can its "pay-per-performance" model break through the ceiling for enterprise payment?

On September 3, OpenAI released GPT-6 Astra. According to the latest comprehensive ranking by Artificial Analysis, in independent evaluations, Astra tied for first place with Anthropic's Claude Fable 5.1, regaining parity in the technological narrative.

Astra and Fable 5.1 offer comparable performance, but their pricing strategies are polarized.

Instead of continuing the price reduction strategy of its previous flagship models, OpenAI set the API price of Astra at 2.5 times that of its predecessor, GPT-5.6 Sol , and began to explore "pay-per-task" pricing. At a point where performance is only on par with, rather than crushing, competitors, this is more like a bold test of the market's bottom line for paid services.

According to the official statement, Astra achieved better results with fewer output tokens in some tests, and the estimated API cost per task was actually lower.

Anthropic, on the other hand, took the opposite approach: Fable 5.1 did not adjust the base unit price, but reduced the cache read price to $0.25/million tokens, a 75% reduction from Fable 5 , which is expected to reduce the cost of typical workloads by about 25% and the cost of highly agented tasks by up to 45%.

At a time when both companies have secretly filed for IPOs and are aiming for trillion-dollar valuations, this divergence points to the core issue of commercializing large models: enterprise customers’ willingness to pay for “extra intelligence” is approaching its limit, and whether the premium narrative of “more expensive but more powerful” can hold true will determine who can truly monetize the model’s capabilities into sustainable revenue.

After catching up, the price doubled against the trend.

OpenAI did not follow the strategy of continuously lowering prices of previous flagship models when pricing Astra; instead, it re-established a price premium for its top-of-the-line capabilities.

Its long-term intention goes beyond simply increasing the price per token—shifting from charging based on quantity to charging based on the results of task completion, and the business model is expected to evolve into a pricing system based on project complexity, time consumption, or output value.

This pricing is supported by an unprecedented scale of investment. Astra used over 100,000 GPUs for training at its Stargate, Texas facility, far exceeding GPT-5.6 Sol;

The Astra ARC-AGI-3 achieved a 99.9% general inference score (compared to 7.8% in its predecessor), making it the most notable single-feature improvement in this release. The high hardware costs and energy consumption directly drove up the subsequent API pricing.

Anthropic reverses price cuts, jeopardizing its cost-effectiveness.

Fable 5.1's input and output prices remain unchanged from its predecessor at $10 and $50 per million tokens, respectively; the adjustment focuses on the "cache hit" stage, i.e. the billing when the model rereads previously processed content, with a reduction of 75% to $0.25 per million tokens.

Anthropic states that for typical workloads, the expected cost of using Fable 5.1 will be about 25% lower than that of Fable 5; for highly agent-based tasks, the savings can be as high as 45%.

This "price-for-volume" strategy corresponds to the judgment that enterprise customers' willingness to pay for "additional intelligence" is nearing its limit: once leading models reach a certain capability threshold, enterprises are more concerned with cost-effectiveness than simply paying a premium for higher benchmark scores.

The dual test of valuation and willingness to pay.

Both companies have now secretly filed for IPOs, aiming for a trillion-dollar valuation. Anthropic's valuation has reached $965 billion after its latest round of financing, surpassing OpenAI's $852 billion.

Under the pricing logic of the capital market, the key to success is no longer who gets the most first-place rankings on several benchmarks, but who can more stably convert expensive model capabilities into sustainable revenue.

OpenAI tested the acceptance of "pay-per-performance" by raising prices by 2.5 times, while Anthropic used price reductions to increase the penetration rate of enterprise workflows—whoever succeeds first is more likely to gain the upper hand in the next stage of competition.

The focus will now shift to the actual implementation progress of Astra's "pay-per-performance" model, and the real choices enterprise customers face between "more expensive but potentially more effective" and "cheaper and sufficient".

Risk Warning and DisclaimerInvesting involves risk; please exercise caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Any investment decisions made based on this information are at your own risk.