Anthropic Completes AI Mannequin Upgrades With Claude Opus 4.5—And Slashes Costs – Decrypt

Anthropic launched Claude Opus 4.5 on Monday, finishing its three-model household and marking the corporate’s third main launch in simply two months. The brand new flagship mannequin claims the highest spot in coding benchmarks whereas reducing costs dramatically.

The discharge caps a rapid-fire rollout that started with Claude Sonnet 4.5 in late September and continued with Claude Haiku 4.5 in October. Now with Opus becoming a member of its siblings, Anthropic affords builders an entire toolkit: Opus for advanced manufacturing work, Sonnet for on a regular basis duties, and Haiku for velocity and efficiency-related duties that require easy logic.

Claude Opus 4.5 scored 80.9% on SWE-bench Verified, a benchmark testing real-world software program engineering duties. That edges out OpenAI’s GPT-5.1-Codex-Max at 77.9% and Google’s Gemini 3 Professional at 76.2%. Anthropic says Opus outperformed each human candidate on its inner efficiency engineering examination—a two-hour evaluation designed to guage judgment underneath strain.

There was a race between AI giants to finish the 12 months within the high of the leaderboards. Google launched Gemini 3 Professional on November 18, positioning it as a breakthrough in multimodal reasoning. OpenAI countered the subsequent day with GPT-5.1-Codex-Max.

Anthropic’s response with Opus got here just some days later, nevertheless it arrived with a hook: pricing at $5 per million enter tokens and $25 per million output tokens, which represents a 67% minimize from the earlier Opus mannequin.

Alibaba’s Qwen fashions add one other dimension to the race. The corporate launched Qwen2.5-Max in late January with over 20 trillion coaching tokens, claiming it outperforms DeepSeek-V3 on key benchmarks. Qwen3-Max, launched in September with greater than 1 trillion parameters, ranks third globally on LMArena and excels at completely different duties like deep analysis, multimodal reasoning, or workflows in jap languages. Whereas Qwen fashions stay comparatively obscure in Western markets, they characterize China’s push for AI self-reliance amid U.S. chip export restrictions

That pricing sits between the OpenAI’s latest GPT-5.1 ($1.25/$10) and Anthropic’s older Opus 4.1 ($15/$75), although it is nonetheless pricier than Gemini 3 Professional’s $2/$12. The discount indicators market strain as main AI labs compete not simply on functionality, however on making frontier intelligence economically viable for scaled deployment.

Claude’s newest providing continues to be pricier than many Asian opponents, however can be a bit extra succesful. So customers now have the flexibility to decide on between cost-efficiency or pure technical functionality.

Sonnet 4.5, launched September 30, introduced state-of-the-art coding and agent capabilities at reasonable price and was already higher than Opus 4.1 at particular duties. The less complicated Haiku 4.5 was  unveiled October 15. Opus 4.5 now sits on the high, dealing with the toughest reasoning and longest-running duties.

Much like Sonnet and GPT-5, Claude Opus 4.5 makes use of what Anthropic calls a “hybrid reasoning” structure—a single mannequin skilled for each direct inference and chain-of-thought processing. It helps a 200,000 token context window and may output as much as 64,000 tokens. The mannequin’s data cutoff is March 2025, barely forward of Sonnet’s January date.

Developer Simon Willison examined Opus 4.5 extensively over the weekend, utilizing it to refactor one among his initiatives. The mannequin dealt with 20 commits throughout 39 information, including 2,022 traces and eradicating 1,173 others. “It is clearly a superb new mannequin,” Willison wrote, although he famous that reverting to Sonnet 4.5 afterward did not dramatically cut back his productiveness.

“I’m not saying the brand new mannequin isn’t an enchancment on Sonnet 4.5—however I can’t say with confidence that the challenges I posed [to] it have been capable of establish a significant distinction in capabilities between the 2,” he wrote.

Theo Browne, a developer, YouTuber, and CEO of AI platform T3 Chat referred to as Claude Opus 4.5 “insane,” including in a video assessment that it’s “positively the perfect coding mannequin ever made.”

The aggressive panorama has turn out to be more and more crowded. Google’s Gemini 3 Professional dominated headlines final week, scoring 1501 on LMArena and incomes reward from Salesforce CEO Marc Benioff, who mentioned he is ditching ChatGPT for Google’s mannequin. That announcement despatched Alphabet’s refill greater than 6% and reportedly rattled OpenAI CEO Sam Altman, who advised colleagues Gemini would create “non permanent financial headwinds.”

Microsoft and Nvidia introduced multi-billion-dollar investments in Anthropic final week, boosting the startup’s valuation to roughly $350 billion. The offers embody expanded Azure integration and Nvidia-powered infrastructure for coaching and deploying Claude fashions. 

Opus 4.5 is obtainable instantly through Anthropic’s API, AWS Bedrock, Google Vertex AI, and the Claude net and desktop apps.