GPT-6 Sol and Luna pricing dropped by half on 22 September, and OpenAI published benchmark numbers claiming Sol beats Claude Opus 5 on cost per task while landing within striking distance of Claude Fable 5 on coding. If your chatbot or automation stack fires the same model for every message regardless of how simple the task is, this is the week to check whether that's still the right call.
At WebEpex we build WhatsApp and web chatbots on Voiceflow and n8n for clients across the GCC, Europe, USA, Canada, and Indian SMB market, and we pulled up OpenAI's pricing page the same afternoon it went live to check it against what's running in production today.
GPT-6 Sol and Luna Pricing: What Actually Shipped on 22 September?#
OpenAI announced GPT-6 Sol and Luna rolling out in ChatGPT Work, Codex, and the API, with Luna also reaching Free and Go users in the desktop app. Sol is built for complex work, coding included. Luna targets high-volume clerical tasks: summarization, extraction, quick Q&A.
Pricing per million tokens, per OpenAI's own numbers:
| Model | Input | Output | Change vs GPT-5.6 |
|---|---|---|---|
| Sol | $2 | $10 | -50% |
| Luna | $0.10 | $0.50 | -50% |
On OpenAI's AutomationBench, Sol claims to beat Claude Opus 5 at roughly 9% of Opus 5's cost per task. On the DeepSWE v1.1 coding benchmark, Sol scores 68.8%, within 1.1 points of Claude Fable 5's best score, at what OpenAI says is about 80% lower cost per task. TechCrunch notes the launch landed roughly 90 minutes after Anthropic's own Opus 5.5 pricing move, the second same-day model-pricing collision this month.
Does This Change Anything for My Business?#
If you're not running any AI in your product or your support stack, no: skip to the next headline. If you already have a chatbot, a WhatsApp flow, or an internal automation calling an LLM API, the honest answer is: probably, but not urgently.
These are OpenAI's own benchmark numbers, not independently verified, so treat the exact percentages as a claim, not a fact. What's real and checkable is the sticker price: half of GPT-5.6 Sol and Luna, published plainly on OpenAI's pricing page. That part you can act on today without waiting for anyone to confirm the benchmarks. The benchmark claims are the part I'd hold at arm's length until someone outside OpenAI reruns them.
The Honest Pros and Cons#
The case for switching or adding Sol/Luna to your routing: cheaper tokens, a real factuality improvement OpenAI claims at roughly half the error rate of the prior generation, and a second high-quality option alongside Claude and Gemini for whatever task you're routing.
The case against moving fast: this is day two. Nobody outside OpenAI has reproduced the benchmark numbers, prompt behavior on a new model always shifts slightly even when the release notes say it's a drop-in swap, and a mid-task model swap on a live client chatbot is exactly the kind of change you test in a sandbox first, not push at 9am on launch day. We've been burned before by a "drop-in" model upgrade that quietly changed how a bot handled edge-case intents (cheap tokens don't help if the bot starts answering weird).
How We're Handling It#
At WebEpex, we ran our usual first pass within hours of the announcement: pull the actual conversation logs from a couple of client Voiceflow builds, bucket the turns into "needs real reasoning" versus "routing, lookup, or a canned answer," and price out what a Luna-tier model would cost against what's running today for the second bucket. That's the same exercise we ran when Fugu Max's routing layer shipped and when DeepSeek's flash pricing dropped: model pricing news is now frequent enough that this has become a standing checklist item, not a one-off.
We're not moving any live client flow onto Sol or Luna yet. What we have done is add both to the shortlist we test against before a client's next chatbot rebuild, the same shortlist that already includes the routing-layer approach we wrote up after GPTBots and Jev shipped this week. My read: the interesting move isn't which single model you pick, it's whether your architecture lets you swap the cheap layer in and out as prices move, because at this rate, this won't be the last 50% cut this quarter.
What I'd Tell a Client Asking About This#
If you're under a few hundred conversations a month, do nothing: the savings won't clear the cost of testing. If you're above that and every turn hits the same model regardless of complexity, it's worth an afternoon: export a week of logs, tag what actually needed reasoning versus what didn't, and price the second bucket against Luna's $0.10/$0.50 rate. You'll know within a day whether it's worth touching. Don't migrate a production flow on launch week, watch it for two.
One More Thing#
If you want a second pair of eyes on whether your current setup is paying the "expensive model tax" on routine turns, send me what you're running and I'll tell you straight (takes two minutes, no pitch attached). cal.com/webepex/growth-review
