GPT-6 Sol and GPT-6 Luna Are Live — OpenAI Halves the Price of Its Best Models
OpenAI released GPT-6 Sol and GPT-6 Luna today, cutting prices by 50% versus GPT-5.6. Sol drops from $4/$20 to $2/$10 per million tokens. Luna drops from $0.20/$1.20 to $0.10/$0.50. Intelligence Index scores hold steady, hallucinations drop significantly, and both models are live on Multos now — Sol on Pro, Luna on Lite.
OpenAI shipped GPT-6 Sol and GPT-6 Luna on September 22, 2026, and both are available on Multos as of today. Sol moves to Pro tier. Luna stays on Lite — same access tier as before, but at half the price. The short version: OpenAI kept the intelligence roughly level with the 5.6 generation and cut costs in half. GPT-6 Sol costs $2 per million input tokens and $10 per million output, down from $4 and $20. GPT-6 Luna is $0.10 input and $0.50 output, down from $0.20 and $1.20. Cache reads carry the same 90% discount. The practical effect on a Multos session where 85% of input hits the cache: Luna blended cost drops to under $0.02 per million tokens. That is not a typo.
What Changed, What Didn't
Artificial Analysis's independent benchmarks tell the clearest story. GPT-6 Sol on the Intelligence Index: broadly level with GPT-5.6 Sol. The Coding Agent Index improved by 2 points — 57 versus 55 — with gains on Terminal-Bench 4.0 (43% vs 37%) and SWE-Atlas-QnA (58% vs 54%). GPT-6 Luna regressed slightly on the Coding Agent Index, down 2 points to 41, with lower scores on SWE-Atlas-QnA and DeepSWE v1.1. The number that stands out most is hallucination. GPT-6 Sol's hallucination rate dropped from 92% to 60% on AA-Omniscience. Luna went from 93% to 77%. Both models answer fewer questions — Sol attempts 83% of questions versus 99% for GPT-5.6 Sol — but the ones they answer are more likely to be correct. For anyone building apps where factual accuracy matters, that's a more useful model to deploy than one that confidently generates wrong answers. Cost per task tells the real economics. Artificial Analysis measured GPT-6 Sol at $1.06 per Intelligence Index task, versus $1.99 for GPT-5.6 Sol. GPT-6 Luna at $0.07 per task, versus $0.18 for GPT-5.6 Luna. Those are not just price cuts — they reflect that both models use slightly more output tokens per task, which means the price reduction per unit of work delivered is over 50%.
GPT-6 Sol for Pro Builders
GPT-6 Sol sits on Multos Pro alongside GPT-5.6 Terra, Kimi K3, and Gemini 3.8 Flash. At $2/$10 per million tokens it's now priced at what GPT-5.6 Terra used to cost — except Sol scores materially higher on coding benchmarks and has the full OpenAI Responses API feature set: hosted shell, computer use, MCP, apply patch, web search, and reasoning effort controls from none to max. For building on Multos: Sol is the right call when you need the full GPT ecosystem — structured outputs that match your schema precisely, apply patch workflows for multi-file codebases, or computer use for testing an interface. The 2-point coding improvement over 5.6 Sol is narrow, but the 50% cost reduction is not. If your Pro plan session used to spend most of its pool on Sol, it now goes roughly twice as far. One change worth knowing: Sol uses slightly more output tokens per task than 5.6 Sol (31K versus 29K average). For sessions with lots of short back-and-forth that works fine. For very long generation runs where you're near the output limit, it's worth knowing the model is naturally a bit more verbose than its predecessor.
GPT-6 Luna for Lite Builders — The Number That Changes Everything
$0.10 input, $0.50 output, $0.01 cached input. That is what GPT-6 Luna costs. For context: GPT-5.6 Luna used to cost $0.20/$1.20. The previous generation was already the most cost-efficient reasoning model at its capability level. GPT-6 Luna halves that again. On Multos Lite, where your monthly pool is 30 million tokens, Luna's blended cost at typical cache rates means those 30 million tokens go significantly further than they did last week. If you've been hitting your pool limit mid-month on Lite — this matters. The slight coding regression versus 5.6 Luna (down 2 points on the Coding Agent Index to 41) is real but narrow. Luna at 41 still comfortably outscores most open-weight models, runs the full Responses API toolset, and carries the same 1.05M context window. For tasks where Luna was already your default — routine coding, explanation, drafting, lightweight agents — it just got meaningfully cheaper with no setup change required. The hallucination improvement (93% down to 77%) is actually the more important change for builders. If you've been running Luna on tasks that involve looking things up, extracting information, or reasoning about factual content, the outputs are more trustworthy than the previous generation's. That's an upgrade in usefulness even before the price cut.
Building Games with GPT-6 on Multos
Game development is one of the better use cases for the GPT-6 family on Multos, specifically because of the Responses API's apply patch and hosted shell access. Here's a workflow that works well. Start a session with your game's full source in context — Phaser, Unity export scripts, whatever you're working with. Ask Sol or Luna to read the entire codebase before making any changes. GPT-6 Sol's stronger performance on SWE-Atlas-QnA (58%) means it handles large-repo comprehension better than its predecessor. Use apply patch to make changes. This is not just convenient — it's accurate. Apply patch means the model doesn't rewrite files from scratch; it surgically modifies exactly what it said it would. For game logic where small changes to physics or collision have downstream effects, that precision matters. For visual iteration: Multos's multimodal upload lets you paste a screenshot of the current game state. GPT-6 Sol's computer use capabilities mean it can reason about what it's looking at — not just the code that generated it. Describe the bug visually and it can trace it back to the code without you having to do the translation. For Luna specifically: the lower cost means you can run longer exploratory sessions without worrying about pool drain. Use Luna for the research phase — 'here is my game architecture, here are three approaches to implementing the inventory system, help me think through the tradeoffs' — and route to Sol for the actual implementation passes where the coding index improvement matters.
Where They Live on Multos
GPT-6 Sol: available on Multos Pro and above. You'll find it in the OpenAI group in the model dropdown. It replaces GPT-5.6 Terra as the headline model at Pro pricing. GPT-6 Luna: available on Multos Lite and above. Same position in the dropdown as before, same Lite tier. Just cheaper. Both models are served via the Multos managed OpenAI connection — no API key, no Responses API setup, no workspace config. You pick the model and it works. Auto Router uses Luna as a default for tasks where it's capable enough, and routes to Sol when the task complexity calls for it. Both GPT-6 models are now part of the routing pool, weighted by the updated cost curves.
Ready to Get Started?
GPT-6 Sol is live on Multos Pro. GPT-6 Luna is live on Multos Lite. Both available today from the model dropdown. OpenAI halved the prices — your pool goes roughly twice as far on the same tasks.
Start Building FreeRelated Articles
Claude Opus 5.5 Is Live on Multos — Fable-Class Work at 40% Lower Cost
Anthropic released Claude Opus 5.5 today and it's immediately available on Multos Elite. It matches Claude Fable 5.1 on most tasks, costs 40% less to run than Opus 5, and is 30% faster. On Terminal-Bench 4.0 it scores 66.4% — higher than GPT-6 Astra, higher than Opus 5, higher than Fable 5.1. Here's what that actually means for building apps and games on Multos.
Read more →MiMo V2.6 Flash and Pro Are Now on Multos — Frontier-Grade Agents at Flash-Tier Pricing
Xiaomi released MiMo V2.6 Pro and Flash on September 21, 2026. Pro scored 46 on the Artificial Analysis Intelligence Index — the highest any open-weight model has ever scored, at $0.13 per task. Flash costs $0.14/$0.28 per million tokens and trails Pro by only 4 points on DeepSWE. Both are live on Multos Lite via DeepInfra.
Read more →GLM-5.3 Flash: The Model That Was Hiding in Plain Sight
For six days, one of the most capable models on OpenCode and OpenRouter had no name. It was just called Ox Alpha. Then Z.ai revealed it: GLM-5.3 Flash, a 320B MoE model with a 1M token context, native multimodal input, MIT-licensed weights — and a price tag that makes you read it twice.
Read more →