OpenAI and Anthropic both recently released new models aimed at lowering costs. Anthropic announced Opus 5.5, the latest version of its main mass-market workhorse model, used for tasks like coding and other complex knowledge work.
And OpenAI announced GPT-6 Sol and Luna, the latest versions of its middle-of-the-road or smaller models focused on efficiency and speed.
These new releases are not about groundbreaking new capabilities. Rather, they’re about efficiency. As both OpenAI and Anthropic target enterprise customers, they’re racing to compete with open-weight models as organizations have explored changing their practices and using model routers to use these pricey, frontier models less in favor of cheaper alternatives.
Anthropic and OpenAI argue that these new releases push the envelope at the frontier (albeit mostly in modest ways), while bringing costs substantially down.
Opus 5.5: A more affordable, more competitive flagship
Opus 5.5 is at the higher end of the models announced today, but the wider context here is that Anthropic is playing a bit of catch-up in its race with OpenAI. OpenAI earlier this month released GPT-6 Astra, which has sometimes been modestly beating Opus 5 in benchmarks and user sentiment. (By price and capability, Astra is competing with both Opus and Fable.)
Benchmarks by Anthropic and its partners now show Opus 5.5 performing better at coding and knowledge work than GPT-6 Astra in some cases, albeit modestly.
The real story here is cost. From Anthropic’s announcement: “Input and output tokens are $4 and $20 per million, 20% less than Opus 5. Cache reads (which make up the majority of agentic and coding work costs) are $0.20 per million tokens, 60% less than Opus 5. Opus 5.5 also generates output more than 30% faster than Opus 5.”
Anthropic further claims that the savings are closer to 40 percent compared to Opus 5 for typical workloads at default settings, because in addition to the cost of tokens going down, Opus 5.5 also uses fewer tokens when completing tasks.
Opus 5.5 is said to be notably capable in “high risk areas” like cybersecurity and biology, so the same protections that applied to Fable 5.1 will also apply here—your requests might be automatically and transparently routed to an older model if they get flagged as treading into protected territory.
GPT-6 Sol and Luna: A modest, cost-focused upgrade
The release of OpenAI’s GPT-6 Sol and Luna is an iterative step forward. Not long ago, the company introduced its Sol, Terra, Luna naming convention for its GPT-5.6 family of models. And more recently (just this month), it introduced GPT-6 Astra, which has been both its most advanced and most powerful model.
It may not be obvious what those names mean, so here’s the quick rundown. Astra is the most aggressively powerful (and pricey) model, meant for heavy-duty coding, research, and so on. Sol is a capable but more efficient and affordable alternative—meant to be the sort of daily driver for tasks like that. Terra is the balanced, general-use model. And Luna is the fast, cheap option.
You could reasonably and roughly position Astra against Anthropic’s Fable, Sol against Opus, Terra against Sonnet, and Luna against Haiku, but it would be an imperfect mapping, especially for the higher-end models.
OpenAI says GPT-6 Sol and Luna were trained with similar methods to those used to train GPT-6 Astra. Depending on the benchmark, they’re sometimes a few percentage points more capable than their predecessors at certain tasks, but they cost half as much to use.
GPT-6 Sol’s API pricing is $2 per 1 million input tokens and $10 per 1 million output tokens. For Luna, it’s $0.10 and $0.50, respectively.
Opinion: The devil is in the dollars
The discourse around frontier models is chaotic. You have people on social media declaring that it’s possible to one-shot complex 3D video games with models like GPT-6 Astra, but you also have news of security breaches and other alignment issues, calls for slowdowns and regulation, and so on.
Some of that is worthy of serious attention, while some of it is noise—and some of it is serious, but being spun or exaggerated for commercial positioning.
Further, there’s a growing recognition that the models—be they Anthropic’s, OpenAI’s, Alibaba’s, or any other big player’s—are not the only important engines of progress and results in recent months. The orchestration and harnesses for those models are at least as important.
The frontier is still moving forward, and today’s models are apparently more capable and aligned than those from a few months ago. But some developers and enterprises are a little less focused on the frontier and more focused on exactly how to operationalize all this and keep it cost-effective.
So while we’re seeing AI leaders calling for a slowdown ostensibly or partially for safety reasons, there’s also a practical and economic reality: The models we have now are good enough to do a lot of helpful things, but they require sophisticated contextualization and operationalization by human beings—whether in the runtime and harnesses or in organizational practices—to be put to use. As a result, some of these companies’ customers may be increasingly less focused on demanding better performance. They’re looking for predictable deployments and, most of all, reasonable costs.
That has the potential to be a natural slowdown of its own, and these models are both being positioned for that new, on-the-ground reality.
Samuel AxonSenior Editor
Samuel AxonSenior Editor
Samuel Axon is the editorial lead for tech and gaming coverage at Ars Technica. He covers physical and generative AI, large language models, software development, gaming, entertainment, and mixed reality. He has been writing about gaming and technology for nearly two decades at Engadget, PC World, Mashable, Vice, Polygon, Wired, and others. He previously ran a marketing and PR agency in the gaming industry, led editorial for the TV network CBS, and worked on social media marketing strategy for Samsung Mobile at the creative agency SPCSHP. He also is an independent software and game developer for iOS, Windows, and other platforms, and he is a graduate of DePaul University, where he studied interactive media and software development.
Source: arstechnica.com
