GPT-6 Sol and Luna vs Claude Opus 5.5: Which New Model Actually Saves You Money? (September 2026)
Anthropic released Claude Opus 5.5 on September 22, 2026, priced at $4 per million input tokens and $20 per million output tokens. About 90 minutes later, OpenAI answered with two new models: GPT-6 Sol at $2 in / $10 out per million tokens, and GPT-6 Luna at just $0.10 in / $0.50 out. None of these prices carry an expiration date. Luna is the cheapest of the three by a wide margin, Sol undercuts Opus 5.5 on paper, and Opus 5.5 claims performance close to Anthropic’s own flagship Fable 5.1 at 40% less cost to run. Which one actually saves you money depends entirely on what you’re building — not just which sticker price looks smaller.
Jump to a Section
- What Exactly Did OpenAI and Anthropic Each Ship on September 22, 2026?
- How Do the Actual Prices Stack Up Against Each Other?
- Which Model Is Actually Cheaper for a Real Workload, Not Just on Paper?
- Is Claude Opus 5.5 Really as Good as Fable 5.1 for Less Money?
- What’s the Catch With OpenAI’s “No Expiration Date” Pricing Claim?
- Which One Should You Actually Switch To?
What Exactly Did OpenAI and Anthropic Each Ship on September 22, 2026?
Anthropic shipped Claude Opus 5.5 on September 22, 2026 as the first release in its new 5.5 family, priced 20% below Claude Opus 5. When I compared the release timestamps directly, OpenAI’s GPT-6 Sol and Luna followed roughly 90 minutes later — close enough that neither company could have been reacting to the other’s actual pricing sheet, just to the fact that a launch was coming.
Sol and Luna join OpenAI’s existing GPT-6 Astra, which has been the top-tier model since its own release earlier this month. That gives OpenAI a three-tier lineup now: Astra for the hardest reasoning work, Sol for general-purpose tasks, and Luna for high-volume, low-stakes jobs like classification or short summarization.
Two labs, one afternoon, three new price points. That’s the whole story in one line.
How Do the Actual Prices Stack Up Against Each Other?
Here’s every model in play right now, with official list pricing as of September 22, 2026.
| Model | Maker | Input ($/M tokens) | Output ($/M tokens) | Cached input ($/M) |
|---|---|---|---|---|
| GPT-6 Luna | OpenAI | $0.10 | $0.50 | Not published |
| GPT-6 Sol | OpenAI | $2.00 | $10.00 | Not published |
| Claude Opus 5.5 | Anthropic | $4.00 | $20.00 | $0.20 |
| Claude Opus 5 (previous) | Anthropic | ~$5.00* | ~$25.00* | $0.50 |
| GPT-6 Astra | OpenAI | $10.00 | $50.00 | $1.00 |
*Opus 5’s baseline price isn’t restated in Anthropic’s Opus 5.5 announcement — I worked it out from the company’s own “20% below Opus 5” claim, since $4 and $20 are exactly 80% of $5 and $25.
Luna is the outlier here. At $0.10 in and $0.50 out, it undercuts every other model on this list by at least 20x on input cost. Sol sits almost exactly half of Opus 5.5’s price on both input and output. Astra remains the most expensive of the five, unchanged since its own launch.
Which Model Is Actually Cheaper for a Real Workload, Not Just on Paper?
A million output tokens is roughly 750,000 words — far more than almost any single task produces, so the per-million rate mostly matters at scale, not for one-off questions. If you’re running high volume, short-response work (tagging, moderation, simple extraction), Luna’s $0.10/$0.50 pricing is not close — it’s in a different category of cheap, and Sol or Opus 5.5 would be significant overkill.
For work that needs actual reasoning — multi-step coding, long document analysis, anything where a wrong answer costs you a retry — the comparison that matters is Sol against Opus 5.5, since Luna and Astra sit at opposite extremes most teams won’t use for this kind of task. Sol’s list price is half of Opus 5.5’s on both input and output. That sounds like an easy win for OpenAI, until you factor in output length and correction rate.
Here’s where it gets less obvious: a cheaper model that needs a second pass to get a complex task right can end up costing more than a pricier model that gets it done in one. Neither company has published third-party benchmarks putting Sol head-to-head against Opus 5.5 on real coding or reasoning tasks yet — anyone telling you definitively which one “wins” this week is guessing ahead of the data.
Is Claude Opus 5.5 Really as Good as Fable 5.1 for Less Money?
Anthropic’s own claim is that Opus 5.5 performs at the level of Claude Fable 5.1 on most work, while costing about 40% less to run than Opus 5 on typical workloads. That’s Anthropic’s internal testing, not an independent benchmark — worth remembering before treating it as settled.
What is independently checkable: the spec sheet. Opus 5.5 ships with a 1,000,000-token context window and a 128,000-token max output, matched with adaptive thinking that’s always on and can’t be switched off — you control its depth through an effort parameter (low, medium, high, and above) instead. That’s a real architecture change, not just a price adjustment dressed up as one.
I found that detail buried well past the pricing announcement in Anthropic’s own materials, and it matters more than the discount headline: an always-on reasoning layer changes how the model behaves on simple prompts, not just complex ones. Expect slightly slower responses on easy questions in exchange for better ones on hard ones.
What’s the Catch With OpenAI’s “No Expiration Date” Pricing Claim?
OpenAI told VentureBeat the new rates for Sol and Luna carry no expiration date, which is a notable thing to say out loud given how often API pricing has shifted across the industry this year. That doesn’t mean the price can never change — it means OpenAI isn’t framing this as an introductory or limited-time rate the way some past launches were.
The other catch sits in what’s missing from the announcement: neither company has published cached-input pricing for Sol or Luna the way Astra and Opus 5.5 both have. Caching is where the real savings usually live on high-repeat workloads — Astra’s cached rate cuts its input cost by 90%, for example. Until Sol and Luna publish their own cache pricing, you’re comparing uncached rates only, and the eventual gap could look different once caching numbers land.
Which One Should You Actually Switch To?
If your workload is high-volume and low-stakes — classification, tagging, short extraction, simple chat replies — Luna is the clear move. Nothing else on this list gets close to its price for that kind of job.
If you need real reasoning and you’re already on Claude, Opus 5.5’s 40% efficiency claim over Opus 5 makes it worth testing before you assume you need to switch providers at all. If you’re on GPT-6 Astra and the cost has been the sticking point, Sol is worth a side-by-side test on your actual prompts before committing — half the price only matters if the output quality holds up on your specific task.
Don’t switch your whole pipeline on launch-day numbers alone. Run the same 20–30 real prompts through your current model and the challenger, compare output quality and retry rate, then decide.
Before You Migrate a Workload to a New Model
- ✅ Test on your own prompts, not published benchmarks — model “wins” vary heavily by task type
- ✅ Factor in retry rate, not just per-token price, for anything involving multi-step reasoning
- ✅ Check whether cached-input pricing exists yet for the model you’re considering — it can change the real cost by 10x on repeat-heavy workloads
- ✅ Match the model tier to the task tier: Luna-class pricing for simple jobs, Astra/Opus-class for anything a wrong answer would be expensive to fix
One Number Worth Double-Checking Before You Commit
List price and delivered price are not the same thing once retries, context length, and cache-miss rates enter the picture. A $2-per-million model that needs three attempts to get a task right is not actually cheaper than a $4-per-million model that gets it in one.
Also worth noting: none of the benchmark comparisons circulating this week are independently verified yet. Both companies’ own performance claims — Anthropic’s “Fable 5.1-level” line included — come from internal testing published alongside the price cut, not from a neutral third party.
🔗 What this costs you
Benchmarks decide which model you pay for. Here is what happens when you demand the same rigour from an AI product’s accuracy claim.
FAQ
Is GPT-6 Sol or Claude Opus 5.5 cheaper?
On list price, GPT-6 Sol is cheaper at $2 input / $10 output per million tokens versus Opus 5.5’s $4 input / $20 output. Neither company has published independent benchmarks yet showing which one produces more usable output per dollar on real tasks.
What is GPT-6 Luna actually meant for?
Luna is OpenAI’s lightest and cheapest GPT-6 tier, priced at $0.10 input / $0.50 output per million tokens as of September 22, 2026, aimed at high-volume, low-complexity tasks like tagging, short classification, and simple extraction rather than multi-step reasoning.
Did Claude Opus 5 get more expensive when Opus 5.5 launched?
No — Opus 5 wasn’t repriced. Opus 5.5 launched as a new, separate model priced 20% below Opus 5’s own rate, giving Anthropic customers a cheaper option rather than a price change on the model they were already using.
Sources
- TechCrunch — Anthropic releases Opus 5.5 with lower prices and Fable-level performance
- The New Stack — OpenAI releases GPT-6 Sol and Luna, cuts token prices in half
- Unite.AI — Anthropic Releases Claude Opus 5.5 With Lower Pricing and New Safeguards
📌 Hub guide: For every fix, buying decision, and work-from-anywhere setup in one place — see the Tech & Digital Hub.
🔗 Related guides
Fact-checked based on public sources as of September 24, 2026. Pricing reflects official vendor announcements at time of publication and may change — always confirm current rates on the provider’s own pricing page before budgeting.
