OpenAI Cuts GPT-6 Sol and Luna Prices in Half, but Independent Benchmarks Show Flat Performance

Price Cut, Not a Performance Leap

OpenAI released GPT-6 Sol and GPT-6 Luna on September 22, 2026, cutting API pricing roughly in half compared to their GPT-5.6 predecessors while delivering what independent testing describes as largely flat performance.

GPT-6 Sol now costs $2 per million input tokens and $10 per million output tokens, down from GPT-5.6 Sol's $4/$20. GPT-6 Luna costs $0.10 per million input tokens and $0.50 per million output tokens, down from $0.20/$1.20. Terra, previously the cheapest model in OpenAI's lineup, has been discontinued.

OpenAI attributes the cuts to improvements in caching and inference efficiency. The company says cached input tokens now receive a 90 percent discount, and it has shipped a new prompt caching dashboard along with tools that let developers adjust reasoning effort and tool availability without invalidating the cache.

Sol is positioned for recurring complex work — feature development, code review, debugging, data analysis. Luna targets high-volume, well-defined tasks like summarization and information extraction. Both are available via the API as gpt-6-sol and gpt-6-luna, and in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu tiers. Free and Go users get Luna only through the desktop app; neither model has reached standard chat yet.

OpenAI's Benchmarks vs. Independent Numbers

OpenAI's own comparisons emphasize cost-efficiency against Anthropic's Claude lineup. On DeepSWE v1.1, a long-horizon software engineering benchmark, GPT-6 Sol scores 68.8 percent at maximum reasoning effort — within 1.1 points of Claude Fable 5's 69.9 percent, which OpenAI says costs far more per task. On FrontierCode 1.1, Sol scores 49.3 percent for $2.14 per task, versus Claude Fable 5.1's 50.3 percent at $12.83.

Notably, GPT-6 Luna at maximum effort matches Sol's "xhigh" score of 66.6 percent on DeepSWE while costing $0.22 per task versus Sol's $1.00 — undercutting OpenAI's own flagship-tier model by 78 percent on cost for near-identical output.

Artificial Analysis, an independent benchmarking firm, reports a different picture on aggregate intelligence. Its Intelligence Index shows GPT-6 Sol rising modestly from 47 to 48 points, while Luna holds flat at 37. On the coding agent index, Sol gained 2 points and Luna lost 2. On GDPval-AA v2.1, a professional knowledge-work benchmark spanning 44 fields, Sol dropped roughly 100 Elo points and Luna dropped about 75, which Artificial Analysis traced to lower presentation quality and incomplete outputs rather than reasoning failures.

OpenAI's published benchmark set notably omits GDPval and Terminal-Bench 4.0, both standard in prior launches, and does not reference Anthropic's newer Opus 5.5, reportedly up to 40 percent cheaper than Opus 5 with stronger performance.

What This Means

GPT-6 Sol and Luna are a pricing move, not a capability upgrade. The headline number — 50 percent lower cost — is real and verified; the performance gains OpenAI advertises are narrower and, according to independent measurement, partially offset by regressions elsewhere. For developers choosing between OpenAI's own two new models, Luna's near-parity with Sol at a fraction of the cost raises a practical question about why Sol's higher price tier exists at all. The broader pattern — selective benchmark reporting, multiple reasoning-effort settings, and cost-per-task framing — makes cross-vendor comparisons increasingly difficult to trust without independent verification.