The One-Fifth Model
OpenAI chose DevDay to do something the industry has been waiting for: cut the cost of frontier intelligence by 80%. The new GPT-6.1 Sol model, launched September 29, is priced at $2 per million tokens (MTok) for standard input and $10 per MTok for output – exactly one-fifth the cost of GPT-6 Astra, the company’s most capable model. For developers building high-frequency applications, the most consequential number may be the cached input rate: $0.10 per MTok, a 90% reduction from Astra’s $1.00.
The pricing data, confirmed against OpenAI’s developer documentation and multiple corroborating sources, represents the sharpest cost compression in the frontier model tier to date. OpenAI described Sol as a “major upgrade” with strong performance across professional work, computer interaction, and agentic coding. It is not yet available in the main ChatGPT consumer product – the company is rolling it out first through ChatGPT Work and Codex, with plans to expand to Plus, Business, Enterprise, and Edu users.
Same Day, Opposite Moves
The pricing compression arrived with an asterisk. On the same day Sol was unveiled at OpenAI’s annual DevDay conference, the company pulled GPT-6.1 Astra from availability, citing safety and alignment concerns. This marks the second training halt for the company in three months. OpenAI CEO Sam Altman described the Astra decision as falling within the “normal course category” – language that does not quite resolve the structural tension between shipping agent products built on the Astra architecture while pausing the model development required to improve them.
Sol, then, is the model OpenAI chose to ship. Not its most capable, but its most cost-efficient. The message to developers is that near-frontier capability at frontier pricing is the commercial bet OpenAI is making – and that the safety pause on Astra does not extend to the economics of intelligence.
What the Numbers Mean for Developers
The pricing tiers tell a story about where OpenAI believes the market is headed. Standard input at $2/MTok (down from Astra’s $10) and cached input at $0.10/MTok (down from $1.00) are designed to unlock the high-volume, cost-sensitive use cases that Astra’s pricing made prohibitive. OpenAI also announced a new Ultrafast speed mode generating tokens up to eight times faster in Codex and up to six times faster in the API, bundled into the new $500-per-month Pro 500 tier.
CFO Sarah Friar told CNBC that when the company launched its $200 tier, “people thought we’d lost our minds.” Now OpenAI is pushing further upmarket while simultaneously compressing the cost of its API models – a dual strategy that assumes scale will compensate for margin pressure. The company reported a revenue run rate of approximately $68 billion annualized, with 70% quarter-over-quarter growth and its enterprise business doubling since July, per CNBC. OpenAI is in early discussions for a ~$30 billion funding round at a ~$1.4 trillion valuation.
Altman, for his part, signaled that the pricing compression is not finished. “Speaking of cost generally, come watch,” he told CNBC. “We got some cool stuff.” He added that he does not believe any open-source model on the market beats OpenAI on a cost-price comparison – a direct challenge to the Chinese open-weight models that have been gaining traction on inference cost.
The Anthropic Contrast
When viewed alongside Anthropic’s S-1 filing, submitted September 28, the implications of Sol’s pricing sharpen. Anthropic’s prospectus reveals $518 billion in cloud, compute, and infrastructure commitments – approximately 80% of which are binding and non-cancelable – against $42 billion in losses and a target valuation exceeding $2 trillion. For every dollar Anthropic earned in 2025, it committed roughly $113 to future infrastructure.
OpenAI’s ability to deliver a near-frontier model at one-fifth the cost of its own most capable offering, while maintaining a $68 billion revenue run rate, suggests that the cost of intelligence may be decoupling from the cost of compute faster than the compute landlord thesis assumed. If Anthropic’s massive infrastructure commitments are designed to produce capability advantages that a competitor can approximate at 20% of the price, the question becomes not whether the compute is needed, but how much of it is actually required.
What Remains Unresolved
Sol is not Astra. OpenAI has been careful to position it as near-frontier, not equivalent. The company has not published independent benchmark comparisons between Sol and Astra on the tasks that matter most for agentic workflows. The safety concerns that halted Astra have not been disclosed in detail, and it is not clear whether Sol inherits any of the same alignment risks.
The 1.2 billion weekly ChatGPT users that OpenAI cited at DevDay is a self-reported figure that has not been independently audited. The pricing math – even at a conservative 0.1% conversion rate to the $500 tier – generates impressive projections, but conversion assumptions remain untested at this price point.
What is clear is the direction. OpenAI is compressing the cost of frontier intelligence while its most capitalized competitor is locked into half a trillion dollars of non-cancelable infrastructure. Whether that advantage holds depends on whether Sol can deliver on the promise of near-Astra capability – and whether the safety architecture that paused Astra does not eventually catch up with the economics of shipping faster.
Note: OpenAI’s 1.2 billion weekly user figure is self-reported and has not been independently audited. GPT-6.1 Sol pricing verified against OpenAI developer documentation. Long-context tier (>272K tokens) pricing is $4/$15 per MTok with $0.20 cached input. Anthropic S-1 figures are from the prospectus as reported by Reuters – the filing has not yet appeared on SEC EDGAR.