GPT-6 Sol and GPT-6 Luna for Marketers: A First-Week Take on the Two-Tier Reset (and the Lineage Question You Should Answer Before You Migrate)

I had my daily research agent reconfigured by Tuesday evening. That's the headline: OpenAI shipped GPT-6 Sol and GPT-6 Luna on September 22, and within 48 hours the most consequential change in my stack wasn't a smarter model — it was the cached-input price. A 90% discount on cached reads, combined with the new ability to change reasoning effort mid-conversation without nuking the cache, quietly reshapes which jobs are worth running how often. Pricing-per-token was the story. Pricing-per-task-after-cache is the actual story.

This post is the practitioner's first-week read on Sol and Luna specifically — what changed since GPT-6 Astra three weeks ago, what to give each new model, what to keep on its sibling or other peers, and the routing principle that falls out once you stop reaching for the flagship by default. If you want the July pre-release framing of the GPT-5.6 Sol/Terra/Luna trio that this generation descends from, that one's here.

First: the lineage question, because three people have asked me this week

If you've been using ChatGPT since June, you may remember a model called GPT-5.6 Luna. GPT-5.6 Luna is not the same model as GPT-6 Luna. Three things changed between them and they matter for your routing:

	GPT-5.6 Luna (July 2026)	GPT-6 Luna (September 22, 2026)
Pricing	$0.20 input / $1.20 output per 1M	$0.10 / $0.50 per 1M (50% / 58.3% off)
Training lineage	GPT-5.5 / GPT-5.6 generation	"Similar methods as GPT-6 Astra" — trained with the GPT-6 toolchain
Reasoning effort	Limited	Six levels: none / low / medium (default) / high / xhigh / max
Reasoning depth ceiling	Mid-tier	At higher effort, "matches GPT-5.6 Sol factuality at about one hundredth its cost" — per OpenAI

The other half of the lineage question: GPT-6 Terra is gone. The GPT-6 family shipped on September 3 with four models (Astra, Sol, Luna, Terra). Tuesday's announcement reduced it to three. The Sol slot dropped from $4/$20 to $2/$10 — which is roughly where Terra lived at $2.50/$15. OpenAI didn't add a mid-tier; it moved the mid-tier down. Your budget line that used to say "Terra" should now say "Sol."

The new GPT-6 family, as a three-tier rotation

Model	Position	Input $/M	Output $/M	Cached input $/M	Context	What it's actually for
Astra	Flagship — reasoning ceiling, Computer Use	$10	$50	$1.00	1.05M / 128K	Multi-step work across software surfaces, hard reasoning, browser-driving
Sol	Workhorse — agents, complex production	$2	$10	$0.20	1.05M / 128K	Sustained agent loops, sub-agent orchestration, high-quality professional work
Luna	Budget engine — extraction, classification, summarization	$0.10	$0.50	$0.01	1.05M / 128K	High-volume background jobs, content moderation, classifier reruns, bulk extraction

All three share the same 1.05M / 128K context envelope and the same April 20, 2026 knowledge cutoff. The only meaningful difference between them at the API level is the price-per-token and the per-request reasoning budget. Sol and Luna both accept text and image input, return text only, and are reachable via the Responses and Chat Completions APIs. Neither supports fine-tuning or Realtime.

That last row is the one to print out and stick next to your monitor. Same context, same modalities, same endpoints, different bill. Sol at xhigh or max will burn through reasoning tokens in ways Luna at medium won't — but if you're paying attention to the cached prefix, you can mix and match effort levels across turns of the same conversation without losing your cache. That single capability change is bigger than the price cut for anyone running agents.

What actually moved: price, cache, factuality

Three concrete things changed versus the GPT-5.6 generation. Each is a routing decision waiting to happen.

1. Permanent 50% price cut (not a promo). OpenAI confirmed to VentureBeat that the Sol and Luna rates aren't a limited-time offer — they're the new standard rate, funded by inference and caching improvements. This is the first time an OpenAI mid-tier model has shipped at half its predecessor's price on day one without a temporary tag. If you're forecasting pipeline costs, treat $2/$10 and $0.10/$0.50 as the planning baseline.

2. 90% cache read discount + cache survives config changes. Cached input reads on Sol are $0.20 per million tokens; on Luna they're $0.01 per million tokens. The genuinely new piece is operational: you can now change reasoning effort or toggle tools on and off mid-conversation without invalidating the cached prefix. Previously, bumping effort from medium to high meant re-paying for the input. Now it doesn't. You also get explicit breakpoints so you control where the cache ends. OpenAI's prompt caching dashboard and diagnostics tool tell you what hit and what didn't.

For a 200-turn agent session where 80% of the input is the same context across turns, this is the difference between a viable product and one you quietly shut down.

3. Factuality halved (Sol) or matched-at-1/100-cost (Luna). Per OpenAI's internal evaluation — built from de-identified real conversations where users had previously flagged a factual error — GPT-6 Sol makes about half as many factual mistakes as GPT-5.6 Sol, approaching Astra-level reliability. GPT-6 Luna at higher effort levels reaches GPT-5.6 Sol factuality at roughly one-hundredth of the cost. OpenAI flags the evaluation as deliberately adversarial — these are prompts designed to surface errors, not typical traffic — but the directional move is the same. Work that required a mid-tier model three months ago can now run on the budget tier.

Five jobs I'm routing to Sol this week

Sol is the workhorse. It's the model you reach for when an agent loop needs to hold together across many turns, when a multi-step task spans more than one tool, or when the answer has to be right because it's going to a client.

1. Daily Reddit + forum monitors with citations. Sol handles the n8n-style pipeline that pulls 30+ posts, filters for relevance, summarizes each, and posts to Slack with citation links. The cached prefix (system prompt + filter rules + Slack formatting template) holds across the whole daily batch; each post is a fresh uncached chunk on top. At $10/M output, a 200-post daily sweep lands around $0.18 in output alone; with 80% cache hits it lands around $0.04. That's a job Astra at $50/M output makes uneconomical and Luna makes you wonder whether you trust the extraction.
2. Weekly 200-keyword SEO brief sweep. Drop a 200-row keyword table (volume, intent, current rank), the brand-voice rules, and the SERP template into one Responses call with cached prefix. Sol returns 200 briefs in two passes — extract SERP patterns per cluster, then write the brief. Two passes that previously needed GPT-5.6 Sol plus a human spot-check now fit in one session at ~$0.25–$0.40 per weekly batch.
3. CRM hygiene from Otter transcripts. Sol opens the transcript folder, matches each call to the right HubSpot account, updates the next-step field, and flags ambiguous matches. This is the same job I tested on Astra in the first-week Astra post, but Sol does 90%+ of what Astra did at one-fifth the cost. Astra is still the right pick for the 10% of CRM jobs where the click-through into HubSpot's UI matters; for the other 90% — read the transcript, write the summary, save the record — Sol is enough.
4. Meta Ad Library weekly digest → Notion. This is the competitor-monitoring pipeline scaled up: pull the last 7 days of competitor ads, classify each by hook angle and offer type, write the weekly digest. Sol at high effort gives you the same classification quality as Opus 5 at roughly one-fifth the per-task cost. The Computer Use piece (actually clicking through Meta's Ad Library UI) stays on Astra.
5. Brand-voice validation pass on AI-generated copy. Run the candidate copy through Sol at medium effort with a cached brand-voice rubric and a 12-dimension scorecard (the style-guide workflow I wrote about last year). Sol's improved factuality matters here: it catches the "I made up a stat that sounds plausible" failure mode that the cheaper models miss. Output is a pass/revise/rewrite label per block.

Three jobs I'm keeping elsewhere

Some work shouldn't move to Sol even though Sol could do it. Routing is about cost-per-task, not just per-token.

1. Computer Use work stays on Astra. Browser-driving, CMS clicking, dragging spreadsheet cells — anything where the value is the model pressing the buttons — is still Astra's job. Sol scored 60.5% on OSWorld 2.0 at xhigh, within 0.2 points of Claude Opus 5 at medium; Astra scored 72.6%. The 12-point gap is the gap between "the agent gets it done" and "the agent gets it done after three sessions of recovery." If the work needs the browser, pay for Astra.
2. High-volume cheap text stays on Luna or DeepSeek V4.1 Flash off-peak. Bulk meta description generation, RSA headline variations, UGC comment moderation, low-stakes extraction — anything where you want cents per thousand rows, not dollars. Luna at $0.10 input / $0.50 output handles 10K-row jobs that would otherwise go to V4.1 Flash off-peak. The quality bar matters: I trust Luna more than V4.1 Flash on extraction accuracy, and trust V4.1 Flash more than Luna on Chinese-language nuance. Pick per workload.
3. Brand-voice final-pass prose stays on Opus 5 or Astra. The job that turns a structured brief into a 1,200-word thought-leadership post is a creative-quality job, not an agent-quality job. The cost difference between Sol at $10/M output and Opus 5 at $25/M output is real, but the readability gap on a long narrative output is also real. Sol's improved factuality helps; Opus 5's prose rhythm helps more.

The routing principle

Let Sol carry the sustained agent work. Let Luna carry the high-volume background. Let Astra carry the surfaces.

That's the whole framework. Anywhere your workflow ends with "and then the agent held the context together across 20 turns and saved the record" — Sol. Anywhere it ends with "and then we processed 10,000 rows for $0.42" — Luna. Anywhere it ends with "and then it clicked through HubSpot and the result landed in the live CMS" — Astra.

The deeper principle, the one I'd write on the wall if I had one: the right answer in 2026 is almost never "use the flagship." The flagship exists to set the ceiling. Once that ceiling is established, the work moves down the stack to the cheapest tier that still clears your quality bar. Sol clears that bar for most agent work today. Luna clears it for most bulk work. Astra clears it for the surfaces.

What I'd watch this quarter

Three things will shift before the next post on this topic.

1. The Claude Opus 5.5 comparison is missing. OpenAI's launch numbers benchmark Sol against Claude Opus 5 and Fable 5 / 5.1. Anthropic shipped Opus 5.5 about 90 minutes before OpenAI shipped Sol. By Artificial Analysis's Intelligence Index v4.3.2, Opus 5.5 scored 58, Astra 53, Sol 48. Every "Sol beats Opus 5 at 1/11 the cost" line in OpenAI's chart is measured against a model superseded the same day. The head-to-head you'll actually care about — Sol vs Opus 5.5 — hasn't been published. Treat OpenAI's numbers as accurate for the comparison they ran; don't carry them into a procurement deck as the current state of play.

2. The max-effort latency is not what you expect. Third-party measurements (Artificial Analysis) put Sol at 102 seconds time-to-first-token at max reasoning effort, and Luna at 124 seconds. That's the most expensive setting to run, not the typical one — but if you're moving a synchronous endpoint from GPT-5.6 to Sol at high effort, the integration work is in your timeout and streaming layer, not in the prompt. Default HTTP client timeouts sit at 30 or 60 seconds. Measure on your own prompts before you trust either vendor's latency claims.

3. Caching is now the budget lever, not token price. Sol at list price with 80% cache hits costs less than GPT-5.6 Sol at promo with 20% cache hits. The new pricing model rewards you for building agents that reuse context across turns. If your current pipeline invalidates the cache on every request (different system prompt per call, different toolset, ad-hoc temperature), you won't see the 50% saving — you'll see the cache miss. The Prompt Caching Dashboard and diagnostics tool shipped the same day. Read them.

For my own stack this week: the daily Reddit monitor moves to Sol with cached prefix. The 200-keyword weekly SEO sweep moves to Sol. The CRM hygiene from transcripts moves to Sol, with the small subset that needs browser click-through staying on Astra. UGC comment moderation stays on Luna. Brand-voice final-pass prose stays on Opus 5. That's the rotation as of Wednesday.

The frontier is moving fast. The mid-tier is where most of the work belongs.