OpenAI has priced its next-generation GPT-6 Astra model at $10 per million input tokens and $50 per million output tokens — the same base rates Anthropic unveiled days earlier for Fable 5.1. With both companies preparing for public listings, a dual strategy is emerging: hold the line on base pricing while leveraging peripheral cost levers like cache read pricing to win enterprise customers.
The price freeze is no accident. Both OpenAI and Anthropic are on the cusp of stock market debuts, and a base-rate price war would directly hit profitability. Investors are demanding sustainable earnings capacity, and leadership at both companies recognizes that a base-price battle would weigh on their IPO processes.
The competitive focus has shifted from base rates to peripheral costs. In its Fable 5.1 update, Anthropic slashed cache read pricing by 75%, from $1 to $0.25. For developers building agent-based workloads with heavy repetitive context usage, this delivers substantial cost savings — roughly 25% efficiency gains on typical tasks and up to 45% on complex multi-step operations. OpenAI has yet to respond on this front, leaving Astra at a relative disadvantage for cache-heavy applications.
OpenAI is also applying a clear dual-track strategy internally, separating premium and standard models. Astra, with its 1-million-token context window, occupies the $10/$50 premium tier for long-horizon agent tasks. Meanwhile, Sol, the standard frontier model, is offered at $4/$20, creating a 2.5x price gap between flagship and standard tiers. High-value, complex enterprise use cases are funneled to the premium tier, while volume-driven demand for mass integration is absorbed by the standard tier.
This financial architecture is designed to satisfy the competing demands of public-market investors. The premium tier demonstrates high-margin revenue potential that justifies outsized valuation multiples, while the standard tier provides the high-usage metrics needed to validate massive infrastructure investments in GPU clusters and data centers.
A different dynamic is unfolding at the utility tier. Google’s Gemini 3.8 Flash is available at $0.75/$3.75 through year-end, while Meta’s Muse Spark 1.3 holds at $1.25/$4.25. These models target high-throughput, low-latency tasks where per-token cost is the primary decision factor — not the sophisticated agent workloads Astra or Fable address. Frontier labs are effectively ceding this segment and concentrating on the high-margin, high-complexity market.
Astra’s actual pricing structure is far more complex than the headline rates suggest. For long-context usage exceeding 272,000 tokens, input costs double to $20 and output rises to $75. Additionally, OpenAI operates separate tiers: a Fast mode at $20/$100 on short-context benchmarks, and a Batch/Flex tier at $5/$25. For long-context, Fast jumps to $40/$150 and Batch/Flex to $10/$37.50. Enterprise operators are now forced to make granular trade-offs between latency and cost.
Meanwhile, Anthropic’s IPO push is becoming more concrete. The company has selected Morgan Stanley and Goldman Sachs as underwriters and is targeting a listing just before the U.S. midterm elections in November. Market chatter pegs the valuation at up to $2 trillion (approximately 2,692 trillion won). If achieved, it would surpass SpaceX’s $1.77 trillion valuation from its June listing to set a record for the highest-ever market debut.
Behind Anthropic’s aggressive listing timeline is a reversal of fortunes in the enterprise market. According to Ramp, a U.S. corporate spending data analytics firm, Anthropic accounted for 43.5% of enterprise payments on its platform in July, ahead of OpenAI’s 39.7%. Leveraging its Claude Code coding tool, Anthropic first overtook OpenAI in April and has maintained its lead since. Anthropic is estimated to have posted roughly $559 million in adjusted operating profit in the second quarter — its first quarterly profit on record.
The rivalry between the two companies is now expanding into a capital-raising race to secure computing capacity. With OpenAI having confidentially filed for a U.S. listing in June, Anthropic is seizing the initiative by firming up its November IPO schedule. China’s Moonshot AI also filed for a Hong Kong IPO on the 3rd, seeking to raise approximately $3 billion (about 4 trillion won).
Yet the cost structure behind the impressive numbers remains a burden. Anthropic’s annualized revenue has surged to around $65 billion (approximately 87.5 trillion won), but analysts note that real cash flow remains tight once infrastructure investments for next-generation AI models and data center operations are factored in. With the $2 trillion valuation already pricing in growth expectations, the challenge of proving out performance remains.
The trajectory of the AI economy hinges on how companies manage peripheral cost levers. Further experimentation with batch tiers and aggressive optimization of cache read pricing are expected to continue. The attention of operators and investors must shift from headline per-token rates to the total cost of ownership (TCO) for specific recurring workloads. As OpenAI and Anthropic march toward their IPOs, the pressure to protect base-rate margins while optimizing peripheral levers will only intensify.