Ten days after co-signing the Pacing the Frontier appeal with over 1,100 experts asking frontier labs to slow down, Dario Amodei shipped Claude Opus 5.5 while OpenAI put GPT-6 Sol and Luna on sale. The race continues, accelerating exactly as its own protagonists contest it.
Release cadence tells the story
Anthropic calls Opus 5.5 "the strongest-performing model we've tested to date," claiming faster code migrations and reduced load times. Artificial Analysis rates it on par with GPT-6 Astra for agentic work and places it at the top of its Intelligence Index with a score of 58. The company confirms Opus 5.5 matches Mythos 5.1 in cybersecurity and biology, so it falls back to earlier models when safeguards trigger. According to 1thegioi.vn, Anthropic launched Claude Opus 5.5 on September 22 as the newest version in its Claude 5.5 line, with performance equivalent to Fable 5.1 but at a lower operating cost.
The release calendar is the most telling data point: Opus 5.5 arrives just 21 days after Fable 5.1 and Mythos 5.1, which themselves followed Opus 5 by 39 days. In 2025 releases came roughly quarterly; in 2026 the cadence has become almost monthly. OpenAI slowed after a flurry of variants in February and March: GPT-5.6 in July, then Astra on September 3, with Sol and Luna nineteen days later.
The price war heats up
Commercially, both companies are cutting costs aggressively. Anthropic says Opus 5.5 costs 40 percent less than Opus 5 at default settings: $4 per million input tokens, $20 output, $0.20 cache reads (60 percent cheaper) and 30 percent faster generation. OpenAI halved Sol's and Luna's API prices versus the GPT-5.6 promo: per Zerohedge, Sol runs $2 in and $10 out per million tokens; Luna a dime in, fifty cents out. Artificial Analysis measured GPT-6 Sol (max) at $1.06 per task versus $1.99 for GPT-5.6 Sol; Luna $0.07 versus $0.18.
OpenAI aims squarely at Anthropic's customers: on AutomationBench, GPT-6 Sol at xhigh effort beats Claude Opus 5 at max effort for just 9 percent of the per-task cost. Xiaomi's MiMo-V2.6-Pro just climbed to 46 points on the Artificial Analysis index as the smartest open-weights model in the world, showing price pressure now comes from open weights too.
Safety and antitrust, the two shadows of the race
Amodei reiterated on September 12 that the industry must slow down, demanding independent evaluators with internal access and shared standards among democratic countries. Sam Altman and Elon Musk amplified the letter on social media, but these two weeks of action tell a different story. Anthropic explicitly equates Opus 5.5 with Mythos 5.1 in sensitive capabilities; OpenAI justified pausing Astra months ago citing alignment issues, as documented by AlexTech.ai. According to 1thegioi.vn, OpenAI trained GPT-6 Sol and Luna using methods similar to those used for Astra.
The stakes are financial too: Anthropic could go public before year-end, OpenAI not until 2027. Allegations of collusion to slow rivals are already in court, and the US government has rejected any slowdown to avoid ceding ground to China. In Singapore, over twenty-one heads of government signed a joint statement for stronger AI safeguards.
The paradox is simple: the labs called upon to brake are the same ones accelerating toward IPOs.
No comments yet. Be the first!