GPT-6 Sol, GPT-6 Luna and Claude Opus 5.5: What Just Launched and What It Means

September 2026 delivered one of the busiest model-launch days of the year. Within a few hours, OpenAI introduced two new GPT-6 models and Anthropic shipped a cheaper flagship. Here is a plain-English rundown of what launched, what each one actually costs, and what it means if you are choosing a model to build on. Where a number comes from a vendor claim rather than an independent test, we say so.

Claude Opus 5.5 (Anthropic)

Anthropic released Claude Opus 5.5, a new top-tier model that it says performs at the level of its strong Fable 5.1 model for most tasks while costing meaningfully less to run. Two numbers are worth separating here, because they are not the same thing:

• List price: roughly 4 dollars per million input tokens and 20 dollars per million output tokens. That is about 20 percent lower than Opus 5, which sat at 5 dollars and 25 dollars.
• Effective cost to run: Anthropic says Opus 5.5 is about 40 percent cheaper to run than Opus 5. The gap between the 20 percent list cut and the 40 percent claim is efficiency: fewer tokens to finish the same task, plus larger cache discounts. Independent analysts describe it as a 20 percent price cut paired with a bigger cache-hit discount.

The takeaway: Opus 5.5 keeps Anthropic at the top end of capability while narrowing the cost gap that used to push budget-conscious teams toward cheaper rivals. If you already run Opus 5, moving to 5.5 is close to a free upgrade on both quality and price. See our Claude review for the wider picture.

GPT-6 Sol (OpenAI)

OpenAI introduced GPT-6 Sol, a faster and more affordable member of the GPT-6 family. It builds on the advances behind GPT-6 Astra, the September flagship, but is tuned for work at scale rather than maximum capability. Pricing lands around 2 dollars per million input tokens and 10 dollars per million output tokens. OpenAI also said it improved caching and inference efficiency and raised usage limits, and framed the new models as roughly 50 percent lower priced than GPT-5.6 promotional rates.

Sol is aimed squarely at the sweet spot most production apps live in: strong enough for real work, cheap enough to run at volume. For a lot of chatbots, assistants and pipelines, this is the tier that actually decides the monthly bill.

GPT-6 Luna (OpenAI)

Alongside Sol, OpenAI launched GPT-6 Luna, the budget member of the family, priced around 0.10 dollars per million input tokens and 0.50 dollars per million output tokens. That is aggressive: at those rates Luna is one of the cheapest capable models available, undercutting most budget tiers on the market.

Luna is built for high-volume, lower-complexity work: classification, routing, extraction, bulk summarization, and the countless small calls that add up. For tasks that do not need a frontier brain, paying frontier prices is waste, and Luna is priced to make that obvious.

The real story: a cost-per-task price war

Put the three together and the pattern is clear. The frontier is not just getting smarter, it is getting cheaper, fast. But there is a trap in reading these launches purely off the per-token price, and it is the same trap on both sides:

• A lower list price does not always mean a lower bill. What you pay depends on how many tokens a model burns to finish your task, plus caching and batch discounts. The Anthropic 40 percent figure is a cost-to-run number, not a sticker number, precisely because efficiency matters as much as price.
• Half the price per token is not half the bill if the model reasons longer or writes more.

The number that actually matters is cost per finished task, on your workload. Two models with the same sticker price can bill very differently, and a pricier model that finishes in fewer tokens can be cheaper overall.

What to do about it

If you are running anything on these models, this is a good week to re-check your numbers rather than trust a headline:

1. Re-price your workload. Plug your real input and output tokens and monthly volume into a calculator and compare Opus 5.5, GPT-6 Sol and GPT-6 Luna side by side. You can do that in a minute with the Toolglade LLM API cost calculator, which we updated with these new models and prices.
2. Right-size, do not default. Send high-volume, simple calls to a budget model like Luna, and reserve Opus 5.5 or Astra for the genuinely hard tasks. Routing work usually saves more than picking a single best model.
3. Test on your own prompts. Vendor benchmarks are marketing until an independent test confirms them. Run five to ten of your real prompts through each, and log the cost and whether you would ship the output.

For the full method, see our guide on how to estimate your LLM API costs.

Bottom line

Claude Opus 5.5 makes top-tier quality cheaper to run. GPT-6 Sol targets the everyday production tier at a fair price. GPT-6 Luna pushes budget pricing to a new low. For buyers, the winner is not any single model. It is the fact that the same work now costs less across the board, provided you compare on cost per task and route work to the right tier.

Prices and vendor claims here come from launch announcements and third-party reporting, which move quickly. Confirm current rates on each provider pricing page before budgeting.