For developers shipping products with LLMs, inference cost isn’t a footnote. It’s a line item that determines which features ship and which die in planning. This week, Anthropic launched Claude Opus 5 at $5 per million input tokens and $25 per million output tokens. That’s the same price as its predecessor, Opus 4.8, and exactly half the cost of their frontier Fable 5 model. Early reports from developers describe performance that approaches Fable-level quality. This isn’t a minor update; it fundamentally changes the build-vs-shelve calculation for any serious agentic workflow.
The New Economics of Agentic Workflows
The price drop from Fable 5 to Opus 5 moves entire product categories from economically unviable to deployable. Consider an automated code review agent that analyzes pull requests, identifies potential bugs, and suggests fixes. This is a multi-turn, high-token task that was a perfect candidate for Fable 5’s reasoning, but the cost was prohibitive for many teams.
Let’s run the numbers for a small team running this agent on 10 million input tokens and 20 million output tokens per month.
-
Fable 5 Cost:
- Input: 10M tokens * ($10 / 1M tokens) = $100
- Output: 20M tokens * ($50 / 1M tokens) = $1,000
- Total: $1,100 / month
-
Opus 5 Cost:
- Input: 10M tokens * ($5 / 1M tokens) = $50
- Output: 20M tokens * ($25 / 1M tokens) = $500
- Total: $550 / month
Saving $550 a month might seem small, but at ten times that scale, it’s a $66,000 annual difference. As one VentureBeat analysis put it, inference costs have become a “board-level line item.” A 50% cost reduction for near-equivalent performance makes the decision to ship simple. For existing Opus 4.8 users, this is a free performance upgrade at the same price point.
Core Capabilities and Unlocked Verticals
Opus 5 isn’t just cheaper; its capabilities are tuned for complex, multi-step tasks. Scott Wu, CEO of Cognition, noted its “particular strength in debugging and root-cause analysis,” directly supporting its use in engineering automation pipelines like the one modeled above.
More importantly, Anthropic has relaxed certain safeguards that previously blocked entire application categories. Biology-related requests that would have been blocked on Fable 5, for example, are now routed to Opus 5. This is a significant change. It unblocks development for biotech and pharma research tools that depend on analyzing and generating scientific text. These verticals were previously non-starters on the platform.
The model ships with a 1,000,000 token context window and, according to OpenRouter, a max output of 128,000 tokens. This massive context is critical for tasks requiring deep analysis of large codebases, legal documents, or research papers.
How to Deploy Opus 5 Today
Getting Opus 5 into production is straightforward.
The model is available immediately through Anthropic’s official API and on platforms like OpenRouter. Availability on AWS Bedrock and Google Cloud Vertex AI has not been confirmed at launch, so check their documentation for the latest status.
- API Model Name:
claude-opus-5 - Paid Plan Users: If you’re a subscriber to the Claude Max plan, Opus 5 is now the default model in the chat interface. No changes are needed.
For businesses with strict data residency requirements, Anthropic offers a US-only inference option. This ensures your data is processed exclusively on servers located in the United States, which is necessary for certain compliance standards like HIPAA or FedRAMP. This option comes at a 1.1x price multiplier. It is a compliance feature, not a performance tier.
Related
- Claude Sonnet 5 Isn’t a Mid-Tier Model. It’s the New Default for AI Agents.
- Claude Sonnet 5 Makes Opus 4.8 Obsolete for Most Workloads
The launch of Opus 5 is a pragmatic move by Anthropic. They’ve made frontier-level intelligence accessible enough for builders to actually use it in production without destroying their margins. For most agentic workloads, the marginal reasoning benefit of Fable 5 no longer justifies double the cost. Your move, OpenAI.