GPT-6 Sol and Luna reset OpenAI's API curve while Claude Opus 5.5 lowers Anthropic's premium tariff. For enterprise buyers, the meaningful denominator is becoming the cost of a verified completed task.
OpenAI and Anthropic pushed the frontier-model price curve lower on September 22, but the more important change is in what customers can sensibly measure.
OpenAI priced GPT-6 Sol at $2 per million input tokens and $10 per million output tokens, while Luna costs $0.10 and $0.50. OpenAI says those tariffs are 50% below the promotional API prices for GPT-5.6. Anthropic launched Claude Opus 5.5 at $4 per million input tokens and $20 per million output tokens, a 20% cut from Opus 5's $5/$25 list pricing. Some of Anthropic's performance claims imply a larger reduction in workload cost than the tariff alone because the company also says the model is more efficient.
The distinction matters. A 20% cheaper token is not the same thing as a 20% cheaper software migration, research task or claims-processing workflow.
The bill that matters includes failure
For an enterprise agent, total task cost is a stack: uncached model tokens, cached context, tool calls, retries, verification and human review. A cheaper model can lose the economic comparison if it fails twice, invokes more tools or requires an engineer to inspect every result.
OpenAI is explicitly encouraging that comparison. Its launch material publishes cost-per-task measurements for several evaluations rather than only benchmark scores. On AutomationBench, for example, OpenAI reports GPT-6 Sol at high effort reaching a 33.2% score at about 9% of the cost per task of Claude Opus 5. Those remain vendor measurements, not neutral production telemetry, but the choice of denominator is revealing: price competition is moving toward the cost of reaching an acceptable outcome.
Caching widens the gap between tariff cards and real workloads. OpenAI offers a 90% discount on cached input and says changes to its coding infrastructure have cut the share of GitHub content that needs fresh processing by more than half. For agents that repeatedly ingest the same repository, policy corpus or customer history, cache hit rate can move economics more than another small cut in nominal input pricing.
Routing becomes a procurement capability
The widening price ladder creates a second effect. If Luna can handle a high percentage of routine tasks while Sol or Opus is reserved for the hard tail, the buyer no longer has to choose one model for an entire workflow. The valuable layer becomes the router and verifier that decides when a cheap attempt is sufficient and when escalation is worth paying for.
That changes vendor leverage. Frontier providers can still defend premium pricing where reliability, latency or tool use produces a higher accepted-output rate. But they increasingly have to prove that advantage at the workload level. A benchmark lead that does not reduce retries, review or completion time is harder to monetize when capable models are available at a fraction of the tariff.
The unanswered provider-side number is gross profit per accepted workload. Falling list prices are not evidence by themselves that inference economics are deteriorating: serving cost can fall faster through model efficiency, hardware utilization, caching and routing. Nor does lower serving cost guarantee better margins if competition forces the savings straight through to customers.
For enterprise buyers, the practical procurement metric is now straightforward enough to demand: dollars per verified completed job at a stated latency and acceptance threshold, including retries and human review. Token price remains an input. It is no longer the decision.