OpenAI Releases GPT-6 Sol and Luna: 50% Cheaper API Pricing and Benchmarks

MarkTechPost’s September 22, 2026 write-up lays out OpenAI’s mid-tier drop: GPT-6 Sol and GPT-6 Luna sit below GPT-6 Astra, trained with similar methods, live in the API as gpt-6-sol and gpt-6-luna — API-only, no self-host weights.

List prices: Sol $2 / $10 per million input/output tokens (was $4 / $20); Luna $0.10 / $0.50 (was $0.20 / $1.20 — output cut closer to 58%). Astra stays at $10 / $50. OpenAI credits better caching and inference. Availability: ChatGPT Work and Codex for Plus through Edu; Luna on desktop for Free and Go; ChatGPT consumer rollout gradual; not yet fully in Chat at write-up.

We’re reading Sol/Luna as a market-structure move: push Astra-class methods down the price curve so the default workhorse changes.

Sol is the complex coding / professional tier; Luna the high-volume everyday tier. OpenAI’s reported scoreboard leans on cost-per-task as much as raw accuracy — Sol at xhigh on AutomationBench 1.0.6 at 33.2% for $0.27 per task versus Claude Opus 5 max at 26.9% for 11.1× that cost; Agents’ Last Exam 56.4% for Sol at max, beating Opus 5’s best at 60% lower cost; DeepSWE v1.1 Sol 68.8% (1.1 points behind Fable 5 at xhigh) at ~80% lower cost; Luna 66.6% on DeepSWE comparable to Opus/Fable at medium while costing 93–96% less. Internal factuality: Sol about half the mistakes of its predecessor on flagged ChatGPT error conversations; Luna at higher effort matches GPT-5.6 Sol at ~1/100 the cost.

Useful AI loves intelligence that gets cheaper and less wrong. Price cuts tied to caching and inference — cached input discounts up to 90%, shared prefixes within a 30-minute window, Prompt Caching Dashboard, miss diagnostics, explicit breakpoints, mid-conversation effort changes without breaking cache — are healthier than temporary promo cliffs. GitHub reportedly cut the share of prompt tokens needing fresh processing by more than 50% for Copilot.

The harm / downside: lab-reported benches and internal flagged-error sets are not customer incident rates. Same-week frontier drops train buyers to rebenchmark constantly and can compress the digestion time safety and procurement need. If “half the mistakes” only holds on OpenAI’s flagged set, enterprises will feel the gap in production.

I’m watching independent DeepSWE / AutomationBench / factuality replications, whether half-price tiers stick after the launch week framing fades, and whether cache hit rates in the wild match the Copilot anecdote.

Sana Hassan for MarkTechPost. GPT-6 family now three tiers: Astra (hardest), Sol (complex coding/pro), Luna (fast volume). OpenAI also carried Astra’s clearer, shorter communication style into these models.