Anthropic released Claude Sonnet 5 on June 30 with a 1 million token context window and pricing of $2 per million input tokens — while claiming performance that approaches its flagship Opus 4.8 on reasoning and coding tasks. If both claims hold up on your workloads, the interesting question isn't "is this a good model." It's "how much of what I'm currently paying flagship prices for can move down a tier."
The spec that changes architectures, not just costs
A million tokens is several novels or a meaningful chunk of an enterprise codebase in one prompt. For contract analysis, multi-document legal discovery, whole-repo code review, or long customer support histories, that context length lets you skip a lot of the retrieval-augmented generation plumbing — the chunking, embedding, and reassembly pipelines — that existed mainly to work around small windows. RAG doesn't die, but for a real class of document-heavy workloads, "just put it all in the prompt" becomes a legitimate architecture again.
Availability is the other practical detail: Sonnet 5 launched day-one on Anthropic's own platform, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Azure. Whatever cloud you're already committed to, there's a path that doesn't add a new vendor relationship.
Why the pricing is aimed at agents
Sonnet 5 lands in the middle of the industry's shift from single-shot chatbots to agentic systems — agents that plan, call tools, evaluate results, and iterate. Those workflows are token-hungry by construction: one task can consume many multiples of what a single chat response would. A model priced for volume, with a window large enough to hold state across a long multi-step task, is built for exactly that load profile.
The pattern I'd expect to become standard: Sonnet 5 as the workhorse for the bulk of agent reasoning, with Opus 4.8 reserved for the specific sub-tasks that genuinely need frontier-tier depth. Routing by difficulty rather than defaulting everything to the flagship is where the real savings are.
Where to point it first
- Document-heavy workflows. Legal review, compliance analysis, technical documentation synthesis — the expanded window directly removes engineering overhead here.
- Support and internal help desk automation. The per-token price makes it economical to route far more interactions through an AI-assisted pipeline before a human sees them.
- Whole-repo code analysis. Dependency audits, security review, migration planning — tasks that previously required slicing a codebase into disconnected fragments.
- Multi-agent orchestration. The coordinating-agent-delegates-subtasks pattern, with Sonnet 5 handling the majority of subtasks.
What to verify before wide rollout
Cheap and capable doesn't waive the usual diligence, and a few items matter specifically here:
Benchmark on your own data. "Approaches Opus 4.8" is a directional vendor claim. The only evaluation that counts is your documents, your code, your tickets. Run a structured pilot with a fixed eval set before migrating anything.
Check data handling per platform. Regional infrastructure and retention policies differ across direct API, Bedrock, Vertex, and Azure. If you have residency requirements, the platform choice is a compliance decision, not just a billing one.
Set cost alerts before enabling agents. Attractive per-token pricing plus agents that make dozens of sequential calls per task can still add up fast. Usage monitoring and budget alerts go in before broad access, not after the first surprising invoice.
Keep humans on consequential decisions. Model capability doesn't change the rule: legal, financial, and safety-relevant outputs keep a human checkpoint, especially as agent autonomy grows.
The launch itself is the signal worth reading: the enterprise AI market is competing on the economics of running AI at production scale, not on capability announcements alone. If you've been running everything through a flagship model because the mid-tier wasn't good enough, that assumption is now stale — pick your two highest-volume workloads, run Sonnet 5 against them with a real eval set, and see how much of your bill it can absorb before you renew anything at flagship pricing.