For the past few months, getting top-tier LLM performance from Anthropic meant paying for Opus 4.8. As of this morning, you can get nearly identical quality for the price of Sonnet. The catch is you have to change one line of code.
This is not an incremental update. Anthropic’s new Sonnet 5 model effectively closes the performance gap with its flagship, making the Opus pricing tier a questionable expense for the majority of production use cases.
The End of the Opus Premium
The value proposition for Opus 4.8 was simple: you paid a premium for the best reasoning and instruction following. Sonnet 5 erodes that premium to almost nothing. The benchmarks show a model that matches or slightly exceeds Opus 4.8 on the metrics that matter for shipping products.
| Benchmark | Claude Opus 4.8 | Claude Sonnet 5 |
|---|---|---|
| Function Calling Accuracy | 90.7% | 91.2% |
| JSON Mode Adherence | 99.1% | 99.5% |
| Multi-step Instructions | 97.8% | 98.1% |
| MMLU (General Knowledge) | 95.4% | 95.1% |
| HumanEval (Coding) | 92.0% | 91.8% |
While Opus 4.8 holds a fractional lead on general knowledge and coding, Sonnet 5 is slightly better at structured data tasks common in agentic workflows. For most API-driven applications, the performance is a wash. The price is not.
| Model Tier | Cost per 1M Input Tokens | Cost per 1M Output Tokens |
|---|---|---|
| Opus Tier (Opus 4.8) | $15.00 | $75.00 |
| Sonnet Tier (Sonnet 5) | $3.00 | $15.00 |
Consider a typical workload: processing 50 million input tokens and generating 10 million output tokens per month.
- Opus 4.8 cost: (50 * $15) + (10 * $75) = $1,500 / month
- Sonnet 5 cost: (50 * $3) + (10 * $15) = $300 / month
That is an 80% cost reduction for what amounts to equivalent performance. You are paying a $1,200 monthly premium for a model that is now arguably worse at function calling.
One Line of Code: How to Switch
Migrating is simple, but it requires testing. The diff is trivial, but the potential impact on your application’s logic is not.
Here is the change in a Python API call:
# Before: Calling Opus 4.8
response = client.messages.create(
model="claude-opus-4-8",
max_tokens=1024,
messages=[
{"role": "user", "content": "Analyze this data and return a JSON object."}
]
)# After: Calling Sonnet 5
response = client.messages.create(
model="claude-sonnet-5",
max_tokens=1024,
messages=[
{"role": "user", "content": "Analyze this data and return a JSON object."}
]
)Before you ship this, run through a production migration checklist:
- Inventory: Find every service and endpoint in your stack that calls
claude-opus-4-8or older Sonnet models. - Update: Change the
modelparameter toclaude-sonnet-5. - Test: Run your regression tests. Pay special attention to structured output formats, prompt adherence, and any multi-turn conversation logic. Benchmarks do not capture everything. A model’s tone, verbosity, and reasoning style can shift in subtle ways that only surface in your specific context.
- Deploy & Monitor: Push the change to production and watch your cost and latency dashboards. You should see a significant drop in spend immediately.
The New Economics for Builders
This release changes the math on what is financially viable to build. It is a direct response to OpenAI’s aggressive price-performance positioning with GPT-4o, and it continues the commoditization of high-capability models. As Simon Willison often notes, the real story is always in the developer docs, not the marketing blog.
Take a solo developer building a legal document summarizer. A typical contract might be 100k tokens. Summarizing it with Opus 4.8 would cost around $1.88 per document. If you need to process 1,000 documents a month, your Anthropic bill is nearly $1,900. That is a difficult cost to pass on to early customers.
With Sonnet 5, the cost per document drops to $0.38. The monthly bill becomes $380. That is the difference between an interesting side project and a viable business. This “race to the middle” is the most important trend for small builders. Top-tier models are becoming cheap enough to be the default choice for new projects, eliminating the need to start with weaker, open-weight models just to manage costs.
The New Default
Claude Sonnet 5 is not a budget option. It is the new baseline. The Opus tier is now a niche product for a very narrow set of problems.
You might still justify the Opus 4.8 premium for workloads that require deep, sustained reasoning across massive contexts, like analyzing an entire codebase or synthesizing insights from a dozen financial reports at once. If your application’s core value depends on complex, multi-document reasoning chains where Opus demonstrably outperforms on your internal evals, then stick with it.
For everyone else, you are lighting money on fire. Sonnet 5 is the right default model choice for Anthropic users starting today.