Frontier AI keeps racing despite calls to slow down

Despite a July open letter from AI company employees calling for frontier AI labs to slow their pace of development, a message echoed by Anthropic CEO Dario Amodei earlier this month, major model releases continue to appear at quite a clip. Anthropic and OpenAI each released new versions of their Claude and GPT models on Tuesday: Opus 5.5 and GPT-6 Sol and Luna. "We must slow the pace at which we improve the capabilities of AI models," said Amodei in an online post on September 12, 2026, reiterating the message from AI staff who put their names to "Pacing the Frontier." Ten days later, Anthropic celebrated the arrival of Opus 5.5, calling it "the strongest-performing model we've tested to date." The model arrives just 21 days after the release of Fable 5.1 and Mythos 5.1, which followed Opus 5 by 39 days. Overall, the cadence of Anthropic model releases has accelerated from roughly every quarter in 2025 to almost monthly in 2026. OpenAI's release cycle slowed somewhat after a flurry of releases early this year. The upstart delivered several model variants in February and March, followed by GPT-5.5 in April. 77 days later, on July 9, GPT-5.6 reached general availability. GPT-6 Astra debuted September 3, 2026, followed by GPT-6 Sol and Luna 19 days later. The competitive releases from the two rival companies arrive as both prepare for initial public offerings. For Anthropic, that could happen before the end of the year, while OpenAI isn't expected to go public until 2027. Opus 5.5 performs better than its predecessor on software tasks, according to Anthropic, which said the model allows speedy code migrations and can reduce software load times. Artificial Analysis puts Opus 5.5 on par with GPT-6 Astra for agentic work. The Claude model has taken the top spot in the Artificial Analysis Intelligence Index with a score of 58, based on performance on 10 different benchmark evaluations. With great power also comes great responsibility - Anthropic says Opus 5.5 is comparable to Mythos 5.1 in cybersecurity and biology, so it too can fall back to earlier models when safeguards kick in. The new Claude family member is also cheaper than its predecessor. "Our tests show that at default settings it will cost 40 percent less than Opus 5 on typical workloads," Anthropic said in its announcement. "Input and output tokens are $4 and $20 per million, 20 percent less than Opus 5. Cache reads are $0.20 per million tokens, 60 percent less than Opus 5. Opus 5.5 also generates output more than 30 percent faster than Opus 5." That's welcome news as cache reads make up the majority of agentic and coding work costs. OpenAI also addressed costs with GPT-6 Sol and Luna, by leaving high performance workloads to Astra. "Improvements in caching and inference let us serve these models at lower cost, and we're passing those savings directly on to users and customers by reducing API prices for Sol and Luna by 50 percent compared with their GPT5.6 promotional pricing," OpenAI said. Artificial Analysis found GPT-6 Sol (max) costs $1.06 per task compared to $1.99 per task for GPT-5.6 Sol, while GPT-6 Luna (max) costs $0.07 per task compared to $0.18 for GPT-5.6 Luna (max). OpenAI clearly believes it can win business from its rival through better pricing. "On AutomationBench, a test of business workflows across apps, GPT6 Sol at xhigh effort outperforms Claude Opus 5 at max effort at just 9 percent of Opus 5's cost per task," the GPT-maker said. Anthropic and OpenAI claim their models improve alignment and safety, though Anthropic's defenses are at least in part intended to protect its business from copycats. Opus 5.5 comes with "preserved thinking," a defense against model distillation - a method of model copying through interrogation - that was introduced with Fable 5.1. It prevents API customers from editing Claude's prior context in order to obtain the model's chain of reasoning. (R)