While Anthropic is expanding its own model series again, OpenAI is hitting the brakes. Sonnet 5.5 is a more compact version of Opus 5.5; like its larger counterpart, it focuses on efficiency, short texts, and performance that was once limited to the largest models. Meanwhile, OpenAI had planned to release Astra 6.1, the successor to its most capable model. However, for security reasons, that’s not on the table. How is that possible?
Let’s start with Sonnet 5.5. With the release of Opus 5.5, Anthropic had already announced that it would further expand this generation. Unlike 5.1 (which only produced Fable) or 4.8 (which only included Opus), the 5.5 generation is clearly part of a broader strategy. That’s why, in addition to the relatively heavy Opus and the mid-range Sonnet, we can also explicitly expect another small Haiku model. It took nearly a full year before the latter received an update.
Unsurprisingly
Sonnet 5.5 comes as no surprise: just like the best LLMs with that name, this is a Sonnet that, on paper, comes very close to the larger models, but at a lower price. Above all, speed is an advantage over Opus: while the accuracy is already good enough, with Sonnet 5.5 you’ll arrive at an output faster. In some benchmarks, Sonnet 5.5 performs significantly better than its immediate predecessor, Sonnet 5: 70.6 percent in Terminal-Bench 4.0 compared to 10.3 percent for Sonnet 5, and 61.6 percent in a chart-reading test (Chartography), where Sonnet 5 stalled at 15.6 percent.
Sonnet 5.5 at “Max effort” approaches the capabilities of Opus 5.5. If this also translates to real-world performance, the transition from Sonnet to Opus represents a continuous spectrum of options, with trade-offs between speed/cost-effectiveness and quality/cost.
It should come as little surprise to anyone that Opus 5.5 and Sonnet 5.5 were released after Anthropic called for a pause on more capable AI models. The industry seems to agree, as we’ll discuss shortly. On paper, Fable 5.1 is slightly outperformed by the new Opus, but in principle, a Fable-class model would always entail more capabilities, and thus more risks, than smaller models. That is why Anthropic feels confident that these new LLMs will exhibit acceptable and safe AI behavior.
OpenAI takes a breather
OpenAI celebrated DevDay this week, the annual event where the ChatGPT creator speaks in person with its own community. It was supposed to be about the return of ChatGPT Pro at 20 times the previous price and an update to GPT-6 Astra, version 6.1. But the former is nowhere near as affordable as its predecessor: the subscription would cost half as much in dollar terms compared to the previous most expensive Pro subscription.
The more significant news for OpenAI is that it would have liked to release a newer model itself, but cannot. GPT-6.1 Astra was supposed to operate even more autonomously for users. The problem is that this extra autonomy leads to misleading behavior and the use of unsafe external tools. For this reason, OpenAI is imposing a pause on itself, a move the broader industry has already suggested.
As a result, Anthropic and OpenAI may be giving up their lead. Until now, it has always been a matter of time before capable models emerged from American and Chinese rivals, necessitating a new “frontier model” from the two AI frontrunners. Once the competition (be it Meta, SpaceXAI, Google, DeepSeek, Z.ai, or Moonshot AI) catches up to Opus 5.5 and GPT-6 Astta, the real test of the pause that OpenAI and Anthropic are taking will begin.
Read also: GPT-6 Astra, OpenAI’s answer to Fable, is now rolling out