Anthropic Launches Claude Opus 5, Deploys AMD Instinct MI450 GPU, Secures Up to $5 Billion Investment

On July 25, Anthropic launched the new model Claude Opus 5, positioned as more pragmatic: it omits fewer steps on complex tasks, actively verifies intermediate results, balances performance and cost, and is suitable for daily high-frequency use. Opus 5 becomes the default model for Claude Max and the most powerful model available to individual subscribers; the higher-capability Fable 5 is primarily for API and enterprise customers, with limited quotas for individual subscribers. Opus 5 focuses on improving efficiency in coding, knowledge work, and agent tasks, ranking among the top in multiple software engineering, business process, and computer operation evaluations. At the highest thinking effort level, some scores approach or even exceed Fable 5. However, in high-risk scenarios, the restricted Mythos 5 remains stronger. Opus 5 offers adjustable effort levels; higher levels enhance complex problem-solving but are slower and consume more tokens.

Opus 5 is now available on all platforms. Its API pricing is $5 per million input tokens and $25 per million output tokens, unchanged from Opus 4.8 and about half that of Fable 5. Anthropic launched two beta features: dynamic tool adjustment and automatic fallback, designed to reduce cost increases from cache invalidation or interruptions due to safety blocks in agent tasks. In the Frontier-Bench v0.1 test, Opus 5 surpassed all tested models, with scores more than double those of Opus 4.8 and lower per-task costs. In the CursorBench 3.2 test, with max effort enabled, its score was within 0.5% of Fable 5, and per-task cost was about half. In the Zapier AutomationBench test, at the same per-task cost, Opus 5's pass rate was about 1.5 times that of Fable 5.

In the OSWorld 2.0 test, it used just over one-third of Fable 5's per-task cost to surpass Fable 5's best result. In the ARC-AGI 3 test, Opus 5 scored 30.2%, while the previous top score was 7.8% by GPT-5.6 Sol. Among more than a dozen comparisons published by Anthropic, Opus 5 lagged behind other models in four. On third-party leaderboards, as of July 25, the Artificial Analysis Intelligence Index ranked Opus 5 max first with 61 points, just one point ahead of Fable 5. In user feedback, positive comments focused on coding and interactive content generation. Code review platform CodeRabbit concluded that Opus 5 is better suited for writing code than for independently handling full code reviews. Independent developer Li Mo described Opus 5 as 'a reliable but slightly quirky colleague,' with strong debugging capabilities but requiring adjustments to workflows and prompts.

Anthropic has recently accelerated its model release cadence: Opus 4.8 on May 29, Fable 5 and Mythos 5 on June 10, Sonnet 5 on July 1, and Opus 5 on July 25. On July 16, Moonshot AI released the open-weight model Kimi K3. On the day of Opus 5's launch, 25 companies including Nvidia, Meta, and Microsoft jointly published an open letter urging the U. S. government not to prematurely restrict open-weight models; OpenAI, Anthropic, and Google did not sign. Anthropic places greater emphasis on closed-source, tiered access, and risk control. After the U. government imposed controls on Fable 5 and Mythos 5 on June 12, Fable 5 resumed public service, while Mythos 5 remains open only to certain approved institutions. Opus 5's ability to discover source code vulnerabilities is close to that of Mythos 5, but its ability to convert vulnerabilities into attack tools is weaker.

Two days before the Opus 5 launch, Anthropic announced plans to deploy up to 2 gigawatts of AMD Instinct MI450 series GPUs and Helios rack-scale systems, with the first 1 gigawatt scheduled for the first half of 2027. AMD committed to making up to $5 billion in strategic equity investment in Anthropic in the future.