loading
loading
Anthropic and OpenAI have rolled out new AI models, putting lower inference costs and safety safeguards at the centre of the intensifying frontier AI race. Anthropic launched Claude Opus 5.5, its latest flagship model, touting stronger performance, lower running costs, and enhanced safeguards. Meanwhile, OpenAI expanded its GPT-6 family with GPT-6 Sol and GPT-6 Luna, describing Sol as a lower-cost alternative to its recently launched GPT-6 Astra and Luna as its fastest and most cost-efficient model yet. All three models are cheaper than their respective predecessors, highlighting how the competition among frontier AI developers is increasingly centred on the cost of deploying AI systems at scale. Anthropic Cuts Opus Costs Anthropic said Claude Opus 5.5 performs at the level of its more powerful Claude Fable 5.1 model on most tasks while costing 40% less to run on typical workloads than its predecessor, Claude Opus 5. The new model also generates output over 30% faster. Claude Opus 5.5 is priced at $4 per Mn input tokens and $20 per Mn output tokens, 20% lower than Opus 5. However, Anthropic estimates that typical workloads will cost about 40% less because the new model also uses fewer tokens to complete tasks. Cache reads, which Anthropic said constitute a majority of the costs associated with agentic and coding workloads, are priced 60% lower at $0.20 per Mn tokens. Anthropic is positioning Opus 5.5 for agentic coding, computer use, and knowledge work. On the CursorBench 4.0 coding benchmark, the model scored 52.5% at its default effort setting, compared with GPT-5.6 Sol’s top score of 41.7%, while costing about one-third as much per task, according to the company. The company also said Opus 5.5 communicates more naturally than previous models, puts important information upfront, avoids jargon, and follows users’ writing instructions more consistently. The model is available to Claude Pro, Max, Team, and Enterprise subscribers. Developers can access it through the Claude Platform, Amazon Web Services, Google Cloud, and Microsoft Azure. The launch will be followed by Claude Sonnet 5.5 and Claude Haiku 5.5, which will be released in the coming weeks, bringing many of the same performance, speed and safety improvements. OpenAI Launches Cheaper GPT-6 Models OpenAI has positioned the two additions to its GPT-6 family as more affordable alternatives to its flagship GPT-6 Astra for everyday professional and high-volume workloads. The launch of GPT-6 Sol and GPT-6 Luna follows the rollout of Astra earlier this month, which OpenAI continues to position as its most capable model for complex end-to-end work. GPT-6 Sol costs $2 per Mn input tokens and $10 per Mn output tokens, half the promotional prices of its GPT-5.6 predecessor. GPT-6 Luna is priced at $0.10 per Mn input tokens and $0.50 per Mn output tokens. Its predecessor was priced at $0.20 per Mn input tokens and $1.20 per Mn output tokens. Cached input tokens cost $0.20 per Mn tokens for Sol and $0.01 per Mn tokens for Luna, representing a discount of 90% over their standard input-token prices. OpenAI attributed the lower prices to improvements in caching and inference efficiency. The company said Sol and Luna were trained using methods similar to those used for Astra, bringing some of the flagship model’s advances in professional work, factuality, coding, computer use, and alignment to cheaper models. OpenAI is positioning Sol for complex coding and agentic workflows, while Luna is aimed at cost-sensitive, high-volume tasks. Both models support text and image inputs, web search, file search, and computer use, with a 1.05 Mn-token context window and a maximum output of 128K tokens. The company also claimed that Sol makes about half as many factual errors as its predecessor, while Luna delivers substantial improvements in factual reliability at a significantly lower cost. GPT-6 Sol and Luna are being rolled out to Plus, Pro, Business, Enterprise, and Edu users through ChatGPT Work and Codex. Free and Go users can access Luna through the desktop app, while developers can use both models through OpenAI’s API. The GPT-6 family now comprises three price-performance tiers, giving developers more options to balance model capability against inference costs. AI Safeguards In Focus The launches come amid growing concern over the pace of frontier AI development and the risks posed by increasingly autonomous systems. Earlier this month, Anthropic CEO Dario Amodei called on AI companies to “pace the frontier”, arguing that safety practices and public policy need to keep pace with rapidly advancing model capabilities. OpenAI CEO Sam Altman has also backed discussions around slowing AI development. Anthropic said Opus 5.5 is its first model release since Amodei's call to pace the frontier. The model underwent testing by external evaluators including Frontier Design and METR, and Anthropic said it recorded its strongest performance yet on its automated behavioural audit. The company also said Opus 5.5 was about 85% less likely than Opus 5 or Mythos 5.1 to attempt to bypass containment boundaries in a dedicated evaluation. OpenAI said Sol and Luna achieved substantially higher defender success rates than GPT-5.6 Sol and GPT-5.5 Thinking on static and multiturn jailbreak evaluations. Under its Preparedness Framework, OpenAI classified both models as having high capabilities in cybersecurity and biological and chemical domains. However, it placed them below the critical threshold for cybersecurity and said neither reached its high threshold for AI self-improvement. OpenAI said it implemented the same safeguards for Sol and Luna as those used for their GPT-5.6 predecessors. The focus on safeguards follows incidents involving AI agents operating beyond their intended restrictions, alongside concerns about the use of AI for cyberattacks, fraud, and other malicious activities.
document fetched 2026-09-24 02:53 · canonical https://inc42.com/buzz/anthropic-launches-claude-opus-5-5-touts-40-lower-running-costs