Anthropic just made two small-number decisions that will have an outsized operational effect: Sonnet 5 is the default on Claude Free and Pro, and the $2 per million input / $10 per million output API price that launched as an intro rate is now permanent. Put bluntly—Sonnet 5 just became the low-cost, high-volume default for anything you run on Claude.ai or the Claude API.
This is not just a marketing pivot. Anthropic has positioned Sonnet 5 as the lower-cost option across product tiers, with a higher-tier Opus-family model available at materially higher rates. Use the published per-million input/output rates to model costs: at $2 input + $10 output per 1M tokens, one million input tokens and one million output tokens cost $12. A higher-tier model at roughly $5/$25 per 1M would cost about $30 for that same token split. That gap changes architectural trade-offs for retrieval-augmented generation (RAG), agentic orchestration, and high-frequency chat applications.
Why the $2/$10 split matters
Anthropic’s split pricing (input vs output) rewards designs that push more of the work into inputs (better context selection, semantic compression) and penalizes long generated outputs. For typical RAG workloads where outputs are short summaries and inputs contain long context windows, Sonnet 5 becomes very attractive. Conversely, systems that produce long, multi-paragraph outputs or large code blobs will still see higher-tier models rack up cost quickly.
Two operational consequences follow immediately:
- You’ll rethink prompt engineering. Teams will be incentivized to compress signals into inputs and keep outputs terse. That’s a useful discipline, but it also encourages token-shifting that can obscure compute and storage costs.
- Caching and cache-read economics gain leverage. Anthropic and other vendors have been evolving how they price cached reads and runtime tokens; cheap runtime tokens plus aggressive caching is a clear optimization path.
Where Sonnet 5 fits technically
Anthropic positions Sonnet 5 as the everyday model across product tiers. Expect it to be the default for customer-facing chatbots, internal assistants, and agent runners that perform frequent, short interactions. Opus-family models slot as higher-cost, higher-quality options—useful for longer-form synthesis, creative tasks that tolerate higher generation cost, or requests where hallucination risk must be further minimized.
Model identifiers and exact rates can change; confirm the precise API model IDs and current rates in Anthropic’s docs. For cost projections, switch from a "calls per month" metric to explicit token budgeting: expected_input_tokens × input_rate_per_1M + expected_output_tokens × output_rate_per_1M.
A blunt opinion
This was the right call from Anthropic. Making Sonnet 5 the cheap default pushes the market toward clearer cost tiers and forces engineering teams to be explicit about where they need quality vs. where they just need throughput. Vendors who continue to hide token economics behind opaque defaults will lose to pricing clarity.
That said, this does open a new operational wrinkle for platform teams: auditability. Default models with dramatically different cost profiles make it easy for unvetted features to blow up your bill. If your usage governance still treats models as interchangeable, you will get burned.
What to test first
Benchmark Sonnet 5 versus the higher-tier Opus-family model on the exact prompts you use in production: latency, determinism, hallucination rate, and token footprint. Run a token-cost projection (inputs × $2/1M + outputs × $10/1M using the current published rates) over your real traffic shape. Add a cheap-cache experiment for repeated queries.
Final thought
Cheap defaults shift where value accrues: from per-request quality to prompt engineering, indexing, and caching. If your team hasn't instrumented token-level cost and model usage by feature, this change will expose that gap fast. Expect other vendors to respond with clearer low-cost defaults or bundled cache pricing; in the short term, Sonnet 5 is the new floor for cost-conscious LLM architectures.