OpenAI expands GPT‑6 family with cost‑focused Sol and Luna

OpenAI announced two new GPT‑6 family models — GPT‑6 Sol and GPT‑6 Luna — positioned as faster, more affordable variants that carry many of the method advances introduced with GPT‑6 Astra into lower price tiers. The company says these models bring improvements across professional work, factuality, coding, computer use, and alignment while cutting API prices compared with GPT‑5.6 promotional pricing. ([2])

What GPT‑6 Sol and Luna are

According to OpenAI, Sol and Luna were trained with similar methods to GPT‑6 Astra and are intended to “advance the frontier on cost efficiency,” offering a step up in capabilities for everyday and professional tasks while being cheaper to run than earlier GPT‑5.6 models. ([2])

Pricing and availability

OpenAI states it is passing infrastructure savings to customers by reducing API prices for Sol and Luna by 50% relative to GPT‑5.6 promotional pricing; the company provides per‑1M‑token rates and positions the models within the GPT‑6 API family. ([2])

| Model | Input (per 1M tokens) | Output (per 1M tokens) | Reported change |
|------:|---------------------:|----------------------:|----------------:|
| GPT‑6 Sol | $2 (was $4) | $10 (was $20) | 50% cheaper |
| GPT‑6 Luna | $0.10 (was $0.20) | $0.50 (was $1.20) | 50% cheaper |

([2])

OpenAI describes the models as part of the GPT‑6 API lineup and links developer documentation for working with GPT‑6 on its developer site. ([1])

Benchmarks and claimed performance

OpenAI published multiple benchmark claims comparing Sol and Luna to earlier OpenAI models and competitors. On AutomationBench, they report GPT‑6 Sol (xhigh effort) scored 33.2% at $0.27 cost per task, and they say Sol outperforms Claude Opus 5 at far lower cost per task. ([2])

In Agents’ Last Exam V1, OpenAI reports GPT‑6 Sol at max effort scored 56.4%, which the company says beats Claude Opus 5’s highest score at about 60% lower cost per task. The announcement also highlights factuality gains—claiming GPT‑6 Sol makes about half as many mistakes as its predecessor in an internal factuality evaluation based on de‑identified user‑flagged errors. ([2])

For coding, OpenAI cites FrontierCode and DeepSWE results: GPT‑6 Sol allegedly scores 68.8% on DeepSWE v1.1 at max effort, within 1.1 points of Claude Fable 5’s top score while incurring roughly 80% lower cost per task; GPT‑6 Luna is reported at 66.6%, comparable to some competitor settings at far lower cost. ([2])

On computer‑use tests (OSWorld 2.0), OpenAI says GPT‑6 Sol at xhigh effort achieves a similar score to Claude Opus 5 at medium effort while costing approximately 80% less per task, and that Luna can exceed the predecessor at a fraction of the cost. ([2])

Caching, agents, and developer tooling

OpenAI also reports improvements to prompt caching for GPT‑6 that raise cache hit rates by default and offer discounts of about 90% on cached input‑token reads, plus new ways for developers to monitor and optimize caching performance. These changes are presented as complementary cost savings for long conversations and agent workflows. ([2])

The company points developers to the GPT‑6 developer documentation and changelog for implementation details and migration guidance when building agents and applications on the GPT‑6 API. ([1])

Developer impact and practical considerations

OpenAI frames Sol and Luna as options that let teams trade some of Astra’s absolute top‑end quality for much lower token costs and faster throughput, giving developers “more room to iterate” and the ability to run larger or more sustained agent and coding workloads within budget constraints. The announcement reiterates that GPT‑6 Astra remains OpenAI’s top model when the highest quality is required. ([2])

For teams planning to adopt Sol or Luna, the practical items called out by OpenAI are: evaluate the new pricing per 1M tokens against your current token usage, test factuality and coding behavior on your real workloads (OpenAI’s reported improvements are based on internal and benchmark datasets), and measure cache hit rates to capture the reported cached‑token discounts. ([2])