The model landscape is compressing fast. Two years ago, running frontier-quality AI on a budget meant accepting real quality trade-offs. Today, OpenAI’s two newest releases suggest that gap is narrowing faster than most expected.
GPT-6 Soul and Luna: What Was Released
On the same day Anthropic shipped Opus 5.5 — one day after Grok 4.7 arrived — OpenAI released GPT-6 Soul and GPT-6 Luna. Neither is the company’s most capable offering; GPT-6 Astra holds that position. What these models bring instead is speed, economy, and a substantial reduction in cost.
GPT-6 Soul comes in at 2 dollars per million input tokens and 10 dollars per million output tokens — a 50 percent cut from its predecessor. GPT-6 Luna goes further: 0.10 dollars per million input tokens and 0.50 per million output, following an 80 percent cut that had already taken place earlier. That makes Luna roughly 40 times cheaper on input than Soul, and orders of magnitude less expensive than Astra.
What the Benchmarks Say
On Automation Bench — a benchmark measuring how well models handle complex automated tasks — GPT-6 Soul at extra-high effort scores around 33 percent, ahead of Claude Fable 5.1 and ahead of GPT-6 Astra at low effort, though still below Astra at full power. Luna reaches around 20 percent on the same benchmark at under 0.05 dollars per task.
The more striking result is on Frontier Code, which tests software development tasks. Luna scores 66.6 percent at 0.22 dollars per task. Soul scores comparably at 0.80 dollars per task. Astra scores comparably at 1.70 dollars per task. All three fall within roughly six to seven percentage points of each other — despite price differences of 10 to 40 times. For code work specifically, Luna’s value-for-cost ratio stands out.
The clear exception is computer use — navigating desktop interfaces autonomously. Luna scores 53 percent on the OS World benchmark against Astra’s 73 percent, a meaningful gap that matters if that kind of work is part of the picture.
The Pattern Behind the Release
What’s notable here isn’t just the numbers — it’s the timing and the framing. Both OpenAI and Anthropic cut prices on the same day, which reflects something that’s been building across the industry: once a frontier model demonstrates capability, the emphasis shifts toward efficiency. Making that capability accessible at lower cost becomes the priority, rather than pushing further on raw performance alone.
The cadence has become recognizable. A top-tier model releases first — expensive, maximally capable, aimed at the most demanding use cases. Weeks or months later, cheaper variants optimized for speed and cost follow. Soul and Luna fit that pattern exactly. Luna, in particular, is positioned as capable of handling 90 to 95 percent of typical tasks at its price point.
For practitioners building on top of these models, the practical question is no longer whether a capable model is accessible — it’s which tier to route each task to. Luna handles the bulk cheaply; Soul steps in where quality needs to be higher; Astra handles the small percentage of work where peak performance is genuinely non-negotiable. Getting that routing right is where the real cost savings accumulate.
Both GPT-6 Soul and Luna are available in ChatGPT, Workspaces, and Codex for Plus, Pro, Business, Enterprise, and EDU users. Free and Go users get Luna in the desktop app.
As this tiered structure matures across providers, the interesting question shifts: not which model is best, but which is best for this task at this cost. That is the question practitioners will be optimizing for — and the tools to answer it are getting sharper with each release cycle.
Text summarized and optimized using Anthropic’s models and reviewed by a human.