Anthropic 发布 Claude Opus 5.5，OpenAI 以两款更便宜的 GPT-6 模型回应

UPDATED 18:58 EDT / SEPTEMBER 22 2026 AI Anthropic releases Claude Opus 5.5 and OpenAI counters with two cheaper GPT-6 models by Duncan Riley Despite rampant worries about runaway artificial intelligence, the two big AI model makers aren’t yet slowing down: Anthropic PBC released Claude Opus 5.5 today and cut its price 20%, and minutes later OpenAI Group PBC put out two new GPT-6 models, Sol and Luna, at half what their predecessors cost. Input on Opus 5.5 costs $4 per million tokens and output $20, which Anthropic said works out to 40% less on a typical workload than Opus 5. Cache reads took the sharpest cut, dropping 60%, to 20 cents. A faster serving mode carries a premium, at $8 per million input tokens and $40 per million output. OpenAI halved the price of both new models against the GPT-5.6 versions that carried the same names, bringing Sol to $2 per million input tokens and $10 per million output. Luna sits an order of magnitude below that, at 10 cents and 50 cents. Anthropic said Opus 5.5 matches its Fable 5.1 model on most work. Output generation is more than 30% faster than Opus 5, which shipped in late July. Terminal-Bench 4.0, an agentic coding test, came back at 66.4% against 55.8% for Fable 5.1. On AutomationBench the new model reported 40% task completion where Opus 5 managed 26.9%. The widest gap in the published figures sits in scientific research, where Terminal-Bench-Science 0.1 gave Opus 5 a score of 29% and the new model 58.7%. On GDPval-AA v2.1, a knowledge work test scored in Elo, Opus 5.5 rated 1846 against 1708 for Opus 5. Humanity’s Last Exam, run with tools, returned 67.7%. One tester finished a 680,000-line code migration in under a day, work Anthropic said would have taken an engineering team weeks. In testing notes Anthropic published with the release, Yashodha Bhavnani, vice president of AI products at Box Inc., said answers from the new model were 40% less verbose without losing accuracy. GitHub Inc. Chief Product Officer Mario Rodriguez said that in VS Code the model “solved more terminal tasks than Opus 5 in less than half the steps.” Anthropic said the model puts the most important information up front. Other customers quoted in the announcement described fewer retries and less rework. On safety, Anthropic called Opus 5.5 the strongest performer to date on the automated behavioral audit it uses for alignment testing, and said attempts to circumvent containment boundaries dropped 85% against Opus 5. Resistance to prompt injection improved enough to match Fable 5.1 on the Gray Swan benchmark. Restrictions on cybersecurity work carry over from Fable 5.1, and most such tasks are routed to the older Opus 4.8 model. Biology work judged high risk is fenced off as well, with broader access running through verification programs that now include a life sciences track. Frontier Design and METR both tested the model before release. OpenAI aimed its two releases lower down the cost curve. Sol and Luna were built on much the same training methods as GPT-6 Astra, which the company began rolling out Sept. 3, then tuned for cost. Luna is the cheaper of the pair, pointed at high-volume routine jobs such as summarization and extraction. Sol takes on recurring coding and agent work. Astra for Law, a configuration built for legal research, arrived Sept. 17. OpenAI reported Sol completing 33.2% of tasks on AutomationBench at 27 cents apiece, and put it at 68.8% on the DeepSWE v1.1 software engineering test, within 1.1 points of the earlier Claude Fable 5. Luna reached 66.6% on the same test. Sol makes about half as many mistakes as its predecessor, according to the company, and Luna at higher effort settings matches GPT-5.6 Sol at roughly a hundredth of the cost. OpenAI’s comparisons run against Fable 5.1 and Opus 5, so neither launch offers a direct head-to-head against the other. The company is discounting reads of cached input tokens by 90%. It said hit rates have improved, and that coding agents can now change reasoning effort and the set of tools available to them without breaking the cache. Answers from both models come back shorter, with less jargon and fewer low-value details, by the company’s account. Opus 5.5 is available on the Claude Developer Platform as claude-opus-5-5 and through Amazon Web Services Inc., Google Cloud and Microsoft Azure. Claude Sonnet 5.5 and Claude Haiku 5.5 are due in the coming weeks. The OpenAI models are live for developers as gpt-6-sol and gpt-6-luna, and in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu subscribers. Free and Go users get Luna in the desktop app, and neither model has reached the main ChatGPT chat surface yet. Image: SiliconANGLE/GPT Image 2.5 A message from John Furrier, co-founder of SiliconANGLE: Support our mission to keep content open and free by engaging with theCUBE community. Join theCUBE’s Alumni Trust Network, where technology leaders connect, share intelligence and create opportunities. 15M+ viewers of theCUBE videos, powering conversations across AI, cloud, cybersecurity and more 11.4k+ theCUBE alumni — Connect with more than 11,400 tech and business leaders shaping the future through a unique trusted-based network Are you an AWS customer?  Support SiliconANGLE financially by buying your AWS services from our Marketplace portal page and links: https://siliconangle.com/aws-marketplace/ About SiliconANGLE Media SiliconANGLE Media is a recognized leader in digital media innovation, uniting breakthrough technology, strategic insights and real-time audience engagement. As the parent company of SiliconANGLE, theCUBE Network, theCUBE Research, CUBE365, theCUBE AI and theCUBE SuperStudios — with flagship locations in Silicon Valley and the New York Stock Exchange — SiliconANGLE Media operates at the intersection of media, technology and AI. Founded by tech visionaries John Furrier and Dave Vellante, SiliconANGLE Media has built a dynamic ecosystem of industry-leading digital media brands that reach 15+ million elite tech professionals. Our new proprietary theCUBE AI Video Cloud is breaking ground in audience interaction, leveraging theCUBEai.com neural network to help technology companies make data-driven decisions and stay at the forefront of industry conversations. LATEST STORIES Web scraping startup Firecrawl closes $75M investment Cyera raises $400M more amid AI agent security push European neocloud Verda raises $189M to build the AI infrastructure of tomorrow Xiaomi introduces Mimo-V2.6 series open-source AI model family What to expect during Okta’s Oktane event: Join theCUBE Sept. 24 Nexstrom bags $12M to help chipmakers replace silicon with more efficient semiconductor wafers LATEST STORIES Web scraping startup Firecrawl closes $75M investment AI - BY MARIA DEUTSCHER . 48 MINUTES AGO Cyera raises $400M more amid AI agent security push SECURITY - BY MARIA DEUTSCHER . 3 HOURS AGO European neocloud Verda raises $189M to build the AI infrastructure of tomorrow INFRA - BY KYT DOTSON . 8 HOURS AGO Xiaomi introduces Mimo-V2.6 series open-source AI model family AI - BY KYT DOTSON . 9 HOURS AGO What to expect during Okta’s Oktane event: Join theCUBE Sept. 24 SECURITY - BY CHAD WILSON . 9 HOURS AGO Nexstrom bags $12M to help chipmakers replace silicon with more efficient semiconductor wafers INFRA - BY MIKE WHEATLEY . 10 HOURS AGO