OpenAI is expanding its model lineup with GPT-6 Sol and GPT-6 Luna while halving key usage prices. Claude Opus 5.5 from Anthropic and Grok 4.7 from SpaceXAI are arriving at almost the same time, giving you several new frontier models to consider—the term describes current systems at the leading edge of Artificial Intelligence (AI) performance. The main story is not a spectacular leap in capability, but how much useful performance you can now get for your budget.
What separates the new models
OpenAI presents GPT-6 Sol and Luna as models offering different balances of capability and cost. Sol is intended for recurring, complex work, while Luna is designed to handle large numbers of clearly defined tasks quickly and cheaply. Both belong to the GPT-6 family but sit below the previously released flagship, GPT-6 Astra.
According to OpenAI, Sol is suited to developing new features, reviewing code, finding errors, and analyzing data. Luna targets document summaries, information extraction, and short questions. In everyday office work, for example, Luna could condense a large set of documents, while Sol would be the more plausible choice for a demanding analysis of the same material.
Anthropic is taking a somewhat different approach with Claude Opus 5.5. The model targets coding and complex knowledge work and, according to the reported claims for Claude Opus 5.5, matches the more expensive Claude Fable 5.1 on most work. That statement is based on provider benchmarks, however, and is not the same as evidence from your own tasks.
Grok 4.7 rounds out the group as a model for coding, agentic tasks, and knowledge work. Agentic means that a model coordinates several steps and tools with a degree of autonomy. According to the Grok 4.7 release report, it uses a larger base model and a longer reinforcement-learning run while retaining the price and speed of Grok 4.6.
What the new prices mean
The published prices cover the application programming interface (API), which lets outside services use an AI model. Billing is based on tokens, small pieces of text that you send to the model or receive in its response. One million input tokens cost $0.10 with GPT-6 Luna and one million output tokens cost $0.50; GPT-6 Sol costs $2 and $10 respectively.
Luna’s input is therefore half the price of GPT-5.6 Luna. Its output price falls from $1.20 to $0.50. Sol also halves the GPT-5.6 Sol prices, reducing input from $4 to $2 and output from $20 to $10, as shown in The Decoder’s price comparison.
Claude Opus 5.5 costs $4 per million input tokens and $20 per million output tokens. Anthropic describes this as 20 percent cheaper than Opus 5. Another source claims 40 percent lower running costs for typical work at default settings; the two figures are not necessarily contradictory because one compares list prices while the other covers a complete workload. Both ultimately rely on provider claims and have not been independently confirmed for every use case.
Grok 4.7 is priced at $2 for input and $6 for output. Its output is therefore cheaper than GPT-6 Sol’s, while Luna operates in an entirely different price tier. A comparison of current API pricing also shows GPT-6 Sol matching the price of GPT-5.6 Terra; The Decoder reports that Terra is no longer offered.
These figures say relatively little about the monthly costs faced by individual subscribers. The sources provide neither new personal subscription prices nor complete details about usage limits. They establish only that GPT-6 Sol and Luna were reportedly available in ChatGPT Work and Codex at launch.
How caching changes the bill
A significant part of the price reduction is connected to prompt caching, which stores repeated input for reuse. If a service sends the same instructions or document sections with every request, the model does not need to process all of that material at the full price each time. Cached GPT-6 input costs $0.01 per million tokens for Luna and $0.20 for Sol, according to the published pricing.
OpenAI highlights higher prompt-cache hit rates, new diagnostic features, explicit breakpoints, and additional controls. These features are meant to lower both latency—the delay before receiving a result—and cost. The improvements are OpenAI’s claims and have not been independently tested in the supplied sources.
This matters most when the same foundation is used repeatedly. A company might summarize many reports under identical rules or extract the same fields from a series of similarly structured documents. A single short question, by contrast, gains little from caching.
Anthropic also charges $0.20 per million cached input tokens for Opus 5.5. Ars Technica reports that cache reads account for a large share of the cost of agentic and coding workloads. Input and output list prices alone may therefore be insufficient when you compare the real cost of an extended workflow.
Pros and Cons of switching
Pros:
- Lower variable costs – GPT-6 Sol and Luna cost roughly half as much as their direct predecessors on key API price categories.
- A better task match – Luna covers simple high-volume work, while Sol, Opus, and Grok target more demanding jobs.
- Cheaper repetition – Prompt caching can sharply reduce the cost of recurring instructions and reused document sections.
- More competition – OpenAI, Anthropic, and SpaceXAI now offer different price-to-performance profiles at the same time.
Cons:
- Provider benchmarks – Many quality claims rely on internal tests and remain unconfirmed in your everyday work.
- Inconsistent measurements – List prices, typical running costs, and cache prices describe different aspects of spending.
- Migration effort – Existing workflows and saved instructions need to be checked again after a model change.
- Incomplete availability details – The sources do not provide full subscription terms or country-specific launch information.
OpenAI also claims that GPT-6 Sol makes about half as many mistakes as its predecessor in an internal evaluation based on de-identified real conversations. According to TechCrunch’s launch report, this would put Sol at GPT-6 Astra’s level of reliability. Because OpenAI conducted the evaluation, it should not be treated as a neutral comparison with Claude or Grok.
The claim that Opus 5.5 beats GPT-6 Astra on some coding and knowledge tasks likewise comes from benchmarks run by Anthropic and its partners. Ars Technica characterizes the improvements overall as relatively modest, arguing that efficiency rather than fundamentally new capability is the real story. That assessment cuts through some of the usual launch noise, although lower costs can still matter greatly to paying users.
What this means in practice
For beginners: Start with a clearly defined task rather than the model described as the strongest. OpenAI positions Luna for summarizing multiple documents, extracting specified information, and answering short questions; compare a few results with your current model and check them for factual errors. If you do not use an API, the quoted token prices do not directly determine your subscription bill.
For advanced users: Compare Sol, Opus 5.5, and Grok 4.7 with a small set of your own repeatable tasks. Track output quality, response length, errors, processing time, and cache use rather than relying on a single score. For regular data analysis or recurring document workflows, lower input costs may matter more than a narrow lead in a provider benchmark.
The sources give no Swiss-franc pricing, information about support for Swiss language variants, or country-specific availability. They also do not explain where data is processed or which settings are available for Swiss privacy requirements. For schools, businesses, and public authorities in Switzerland, data handling therefore remains an unresolved consideration alongside quality and price.
These releases primarily mark a competition over price and efficiency: OpenAI is halving major API prices, Anthropic is lowering the price of its capable workhorse model, and Grok 4.7 retains its previous rate despite a larger foundation. Switching makes the strongest financial case when you have high, measurable usage volumes; the difference may remain small for occasional use. The main unresolved risk is that internal quality claims will prove less pronounced under real working conditions than in provider tests.
Sources
- Introducing GPT-6 Sol and Luna – Unknown, 2026-09-22
- Claude Opus 5.5, GPT-6 Sol, GPT-6 Luna, and a new price war – Unknown, 2026-09-22
- New Anthropic, OpenAI models make same promise: A little more for a lot less money – Unknown, 2026-09-22
- Better prompt caching for GPT-6 – Unknown, 2026-09-22
- OpenAI bringt GPT-6 Sol und Luna und senkt API-Preise um 50 Prozent – Unknown, 2026-09-22
- Anthropic Releases Claude Opus 5.5: Fable 5.1-Level Performance at 40% Lower Running Cost Than Opus 5 – Unknown, 2026-09-22
- OpenAI launches GPT-6 Sol and Luna, boasting lower cost and fewer mistakes – Unknown, 2026-09-22
- SpaceXAI Releases Grok 4.7: A Larger Base Model at the Same $2/$6 Price as Grok 4.6 – Unknown, 2026-09-22
Image: Tara Winstead via Pexels