Claude Sonnet 5 dropped yesterday at near-Opus quality for 60% less. Most people will just switch models and leave the real savings on the table.
The wins are in the config: effort control, model routing, and the intro-price window before it closes on August 31.
Set it up right and you get Opus-level results on most tasks while spending a fraction of what you did last week.
Here's the full setup ������
Before we dive in, I share daily notes on AI & vibe coding in my Telegram channel: https://t.me/zodchixquant[���](https://abs.twimg.com/emoji/v2/svg/1f9e0.svg)���
https://pbs.twimg.com/media/HMIMbdjXsAAxn6q.jpg

Why this launch matters for your bill

Sonnet 5 lands close to Opus 4.8 on agentic coding and actually edges it on knowledge work, while costing far less. Standard pricing is about 40% cheaper per token than Opus, and during the intro window it's roughly 60% cheaper.
That changes the math. Tasks you used to send to Opus for quality can now go to Sonnet 5 and come back nearly as good for a fraction of the cost. The trick is setting it up so the savings are automatic, not something you think about each time.
Three pieces do that: effort control, smart routing, and locking in the intro price.
Here's each.
https://pbs.twimg.com/media/HMINJQ8WQAAnNh6.jpg

1. Effort control: pay for thinking only when it helps

The biggest new lever is effort level. Sonnet 5 lets you dial how hard it thinks, and Anthropic raised the rate limits so you can push it when a task deserves it.
Low effort is cheap and fast for simple work. High effort costs more but rivals Opus on hard tasks.
The mistake is running everything at one level. Set a sensible default and let hard tasks opt up. In your settings.json:

{
  "model": "claude-sonnet-5",
  "effort": "medium"
}

Then, in CLAUDE.md, tell Claude when to spend more:

## Effort policy
- Default to medium effort for normal work.
- Use high effort only for: tricky debugging, multi-file
  refactors, architecture decisions.
- Use low effort for: formatting, renames, simple edits,
  boilerplate.

Match the effort to the task. Don't burn high effort on trivial work.

Now you pay for deep thinking on the 20% of tasks that need it, and stay cheap on the 80% that don't.
That single split is where most of the savings live.
https://pbs.twimg.com/media/HMIMiRjX0AAVP3q.jpg

2. Route the hard stuff to Opus, keep the rest on Sonnet 5

Sonnet 5 handles most work at near-Opus quality, but Opus 4.8 still wins on the very hardest reasoning.
The cost-optimal setup isn't "Sonnet 5 for everything," it's Sonnet 5 as the default with a clear rule for when to reach for Opus.
Put the routing rule in CLAUDE.md so it's automatic:

## Model routing
Default: Claude Sonnet 5. Use it for coding, tool use,
refactors, and day-to-day work.

Escalate to Opus 4.8 only when:
- Sonnet 5 has failed the same task twice, or
- the task needs the deepest reasoning (complex system
  design, subtle correctness proofs).

Start on Sonnet 5. Escalate on evidence, not by default.

This keeps 90% of your work on the cheaper model and spends Opus money only where it actually changes the outcome. Most people overpay by defaulting to the biggest model out of habit.
The rule fixes that without you thinking about it.

3. Lock in the intro price before August 31

Here's the time-sensitive part.
Sonnet 5 launched at $2 per million input tokens and $10 per million output, but that intro pricing ends August 31, 2026, after which it rises to $3 and $15. That's a 50% jump on both.
A simple move: if you have large one-time jobs on the roadmap, like a big migration, a codebase-wide refactor, or bulk generation, pull them into the intro window instead of pushing them past it. Same work, up to a third less cost, purely on timing.
https://pbs.twimg.com/media/HMIRuNCXQAA0b4S.png

The full config, ready to copy

Here's everything together. In settings.json:

{
  "model": "claude-sonnet-5",
  "effort": "medium"
}

In CLAUDE.md:

## Effort policy
- Medium by default.
- High only for: hard debugging, multi-file refactors,
  architecture calls.
- Low for: formatting, renames, boilerplate.

## Model routing
- Default to Sonnet 5 for everything.
- Escalate to Opus 4.8 only after two failed Sonnet attempts,
  or for the deepest reasoning tasks.

## Cost note
- Intro pricing ($2/$10) ends Aug 31, 2026. Run large batch
  jobs before then where possible.

Drop that in and your setup defaults to cheap, spends on quality only when the task earns it, and uses the intro window on purpose.

Where Sonnet 5 beats Opus, and where it doesn't

This is the part the announcement glosses over, and it's what actually decides your routing. Sonnet 5 isn't uniformly "almost Opus." It beats Opus on some work, trails it on others, and knowing which is which is the whole game.
Route by task type:

• Knowledge work → Sonnet 5, always. Research, analysis, summarizing messy sources, drafting docs. It actually edges out Opus here, at a fraction of the cost. No reason to pay Opus rates.
• Everyday coding → Sonnet 5. It's strongest on brownfield code, the race conditions and hidden tests nobody wants to touch, tracing bugs to the root instead of patching symptoms.
• Hardest multi-step coding → watch the gap. Sonnet 5 scores ~63% on agentic coding where Opus hits ~69%. Invisible on daily work, real on the toughest changes.
• Deep reasoning + anything cyber → Opus. Anthropic deliberately held Sonnet 5's cyber capability down, so security-sensitive and proof-heavy work still belongs on Opus.
  Route by task type, not by a vague sense of "important," and you'll spend correctly almost every time.

Two lifehacks the launch posts won't tell you

The high-effort trap. It's tempting to crank Sonnet 5 to its highest effort for max quality. Don't, without checking the math first.
At its maxed-out setting, Sonnet 5 can cost more to run than Opus 4.8 at a comparable level of reasoning. So "just set it to high" can quietly make you pay more than Opus, for slightly less quality.
The rule: if you're reaching for max effort on a truly hard task, that's your signal to check whether Opus at medium is both cheaper and better. High effort is for the middle band, not the ceiling.
The 10-minute self-test. Don't trust benchmarks for your own work. Test it.
Take five real tasks you'd normally give Opus. Run each on Sonnet 5 at medium, then on Opus, and compare side by side:

## My routing test
For each of 5 real tasks:
1. Run on Sonnet 5, medium effort. Save the result.
2. Run the same task on Opus 4.8.
3. Score both: did Sonnet 5 match? If yes, that task
   type routes to Sonnet 5 from now on.

After 5 tasks you'll know exactly which of your work
can drop to the cheaper model. Usually it's most of it.

Ten minutes of testing tells you more about your real savings than any benchmark chart, because it's measured on the work you actually do.

Common mistakes

Running everything at high effort. High effort is where Sonnet 5 rivals Opus, but also where it costs most. Reserve it for hard tasks, or you've just rebuilt an expensive model out of a cheap one.
Defaulting to Opus out of habit. Sonnet 5 covers most work at near-Opus quality now. Sending everything to Opus "to be safe" is how bills stay high for no gain. Default down, escalate on evidence.
Ignoring the intro window. The price rises 50% on September 1. Token-heavy work costs the same to run, so timing large jobs before the deadline is free savings you only get once.
Treating effort and routing as manual choices. If you decide per task, you'll forget and overspend. Put both rules in CLAUDE.md so the cheap path is the default path.

The 10-minute setup

2 minutes: set claude-sonnet-5 and medium effort in settings.json.
3 minutes: add the effort policy to CLAUDE.md so hard tasks opt up and easy ones stay cheap.
3 minutes: add the routing rule so Opus is used only on evidence, not by default.
2 minutes: check your roadmap for big batch jobs and pull them into the intro window.
The model didn't just get cheaper. You built a setup that spends like the cheap model and performs like the expensive one, which is the whole point of Sonnet 5.

Thanks for reading!
I share daily notes on AI, finance, and vibe coding in my Telegram channel: https://t.me/zodchixquant
https://pbs.twimg.com/media/HMIUB5TWwAAW7rR.jpg