Altman claims AI cost declines are 20x faster than Moore's law

Why it matters

OpenAI is using its flagship model to optimize the infrastructure that serves it, then directing the savings toward cheaper agent workloads while preserving Sol's premium pricing.

Sam Altman (@sama) said AI costs are falling faster than Moore's law as he amplified OpenAI's latest efficiency claim on July 31st. "i see your moore's law and i raise you 20x," Altman wrote after OpenAI said GPT-5.6 Sol had helped cut the cost of serving itself by 20%.

Altman's line captured the bet underneath OpenAI's July 30th announcement: increasingly capable models can lower the cost of building and operating the next generation of AI systems. The disclosed result was a 20% reduction, rather than a 20-fold improvement. OpenAI also attributed a greater than 15% gain in token-generation efficiency to experiments conducted with Sol.

OpenAI said GPT-5.6 Sol autonomously rewrote and optimized production GPU kernels inside a human-led engineering process. OpenAI also tasked Sol with designing and running hundreds of experiments, monitoring training runs and intervening when hardware failures or training instability appeared.

Those are OpenAI's internal production figures. OpenAI has not published the underlying serving-cost baseline or an independent assessment, limiting direct comparisons with efficiency claims from other model providers. The figures still offer a concrete example of a frontier model performing engineering work on the infrastructure used to run that same model.

OpenAI passes selective savings to customers

OpenAI paired the infrastructure disclosure with steep price reductions for the smaller members of the GPT-5.6 family. Luna's API price fell 80% to $0.20 per million input tokens and $1.20 per million output tokens. Terra fell 20% to $2 per million input tokens and $12 per million output tokens.

Developer Nic Dunz (@nicdunz) said Luna Max matched GPT-5.4 Full on one benchmark at roughly one-thirteenth the token price, four months later. The narrow comparison does not establish parity across workloads, but it illustrates how quickly lower prices can change the economics of a capable model. Another developer, @signulll, replied that the price drop had finally made setting up the API worth it.

Sol's standard API price did not change. Instead, OpenAI introduced a Fast mode that promises processing speeds up to 2.5 times faster than standard requests at twice the price, without changing the underlying model. Fast mode replaces OpenAI's Priority Processing option, while existing requests tagged for priority processing will continue to work.

The selective cuts show how OpenAI intends to package its efficiency gains. Sol remains the premium reasoning tier, while the sharply cheaper Luna becomes the volume model for routine agent work, testing, classification and well-specified implementation tasks. OpenAI can preserve Sol's premium price while using lower-cost models to compete for workloads where developers care more about throughput and unit economics than maximum reasoning performance.

OpenAI also said the lower Terra and Luna prices will reduce how quickly the models consume credits in Codex and ChatGPT Work. Subscription prices and total quota budgets remain unchanged. That distinction matters for developers who experienced GPT-5.6 as a more capable model family without necessarily receiving more total subscription capacity.

The cuts arrived three weeks after OpenAI made GPT-5.6 generally available. At launch on July 9th, OpenAI priced Luna at $1 per million input tokens and $6 per million output tokens, Terra at $2.50 and $15, and Sol at $5 and $30. OpenAI said then that internal agentic token use had increased roughly 22-fold over the prior six months, while the share of research compute devoted to internal coding inference had grown 100-fold.

That internal growth creates a direct financial incentive to improve inference efficiency. Every reduction in serving cost gives OpenAI more room to increase usage, lower prices or protect margins as agents run for longer periods and make more tool calls. OpenAI chose to spend much of that room on Luna, repositioning the smallest GPT-5.6 model for high-volume deployment.

The timing also puts pressure on competitors selling small and mid-tier models. Axios reported that cheaper Chinese open-weight models have forced OpenAI and Anthropic to defend the economics of their proprietary offerings. OpenAI's answer is a tiered system in which Sol handles planning and uncertainty while Luna executes cheaper, repetitive steps.

Altman's Moore's law reference points to the larger strategic claim. OpenAI wants software-driven optimization to compound alongside gains from chips and data centers. If models can repeatedly reduce their own inference costs under human supervision, OpenAI can improve its economics between major hardware generations. The immediate result is narrower: an internally reported 20% serving-cost reduction, a greater than 15% token-efficiency gain and lower prices concentrated in the models OpenAI wants developers to run most often.