OpenAI Ships GPT-6 Sol and Luna at Half the Token Price, but Benchmarks Show Little New Intelligence
OpenAI's two new GPT-6 variants match their predecessors' capability for half the cost, a pricing move aimed at Anthropic that independent evaluations say adds little in actual reasoning gains.
OpenAI released two new models, GPT-6 Sol and GPT-6 Luna, positioning them as cheaper alternatives that match the performance of the company's previous generation while cutting the per-token price in half. The launch, reported by The Decoder, is a direct play for cost-sensitive developers and enterprise buyers who have been weighing OpenAI's API against Anthropic's higher-priced Claude lineup. The pitch is straightforward: same output quality, half the bill.
Independent evaluations of the two models have found little movement in raw capability. Benchmarks cited in early analyses show Sol and Luna trading blows with their predecessors on reasoning, coding, and instruction-following tasks, with gains that fall inside the margin of run-to-run variance. That puts OpenAI in an unusual spot: it is competing on price rather than on a step-change in model intelligence, a shift from the capability-first framing that defined the GPT-4 and GPT-5 launches.
The timing appears to have been complicated by Anthropic, which launched Opus 5.5 around the same window. Reporting suggests OpenAI did not anticipate the simultaneous release, leaving the GPT-6 announcement to compete for attention against a rival model that reviewers have described as a more meaningful capability jump. For developers choosing a stack, the decision now hinges less on which model is smartest and more on which vendor offers the best price-to-performance ratio for a given workload.
The pricing itself is the substantive news. Halving token costs changes the economics of high-volume applications such as document processing, agentic pipelines, and customer support automation, where inference spend scales linearly with traffic. A model that is merely as good as its predecessor but costs half as much can double the feasible scope of a deployed product, or let a startup undercut an incumbent on unit economics without touching its architecture.
What remains open is whether OpenAI can sustain the discount. Cheaper tokens usually reflect distillation, quantization, or more efficient serving infrastructure, and each of those carries trade-offs in latency, context handling, or reliability under load. The company has not detailed the technical basis for the price cut, and until third-party cost and throughput tests land, buyers are taking the half-price claim on faith. The next signal will be whether Anthropic responds with its own price move or leans on Opus 5.5's capability lead to hold its premium.