Stop Paying the Opus Tax

Originally published in Temrel, a weekly newsletter helping developers become better agentic engineers.

Picture this: last week you used Opus 4.8 and tasked it with a rename-and-wire-up job. It did this job perfectly, because of course it did. However, you paid roughly 5x more than you needed to. Your default model is overpowered compared to what's available now. You're taking the Ferrari to pick up the groceries.

On June 30th, Anthropic released Claude Sonnet 5, which is almost as good as Opus 4.8, but much cheaper. They also made it the default model in Claude Code for Free and Pro plans (if you're on those, this issue won't mean much to you; save yourself a few minutes and move on).

If you're on the paid API, and to a lesser extent the Max plans, then this is of interest to you, and you might have missed it.

You need a system to select the model your prompt invokes.

The workhorse tier grew up while you weren't looking

Sonnet 5 is, by Anthropic's own measurements, almost as good in every way as Opus 4.8:

Benchmark	Sonnet 5	Opus 4.8
Agentic coding (SWE-bench Pro)	63.2%	69.2%
Agentic coding (Terminal-Bench 2.1)	80.4%	82.7%
Multidisciplinary reasoning (Humanity's Last Exam)	43.2% (no tools) / 57.4% (with tools)	49.8% / 57.9%

Source: Anthropic

Except in one important metric: price. Sonnet 5 is $2/$10 per Mtok vs $5/$25 on Opus.

In other words, the gap you're paying extra for just got thin. There's not much air now between Sonnet 5 and Opus 4.8.

"Just use the best model" stopped being rational

Thus far you've probably decided that benchmarking models every two days is counter-productive and that you can move a lot faster by just choosing one model and focusing on your output. It's not a dumb approach.

However, at scale this approach will cost you millions. Every task you send to the frontier model out of habit is now paying the Opus Tax.

It's time to start being a little more selective. Opus 4.8 still wins at higher-effort tasks (it may even be more cost-effective there) according to HN consensus. Model selection is no longer somewhat pedantic. It's resource allocation. You need to size it like a server.

Route tasks like you size infrastructure

We use a four-axis heuristic for model selection, with each axis scoring 1-3 points:

Axis	Criteria (1 / 2 / 3)
Scope	Single file / Multi-file / Cross-cutting
Novelty	Pattern already in repo / Familiar domain / Unfamiliar or complex domain
Risk	Throwaway / User-facing but stable / Destructive data changes or hard to roll back
Iteration	One-shot / A few cycles / Long-horizon agentic session

Scoring based on this leads to the following partition of tasks (in the context of Claude Code models):

Score	Model tier
4-6	Haiku
7-9	Sonnet
10-12	Opus

In addition, any risk score of 3 bumps you up one tier automatically.

Here's a worked example: this week at Temrel we were tasked with adding a new admin-only form to the Maitris backend. Relatively simple stuff. Here was our decision:

Axis	Score	Rationale
Scope	2	No cross-cutting
Novelty	1	Patterns already well established in repo
Risk	2	New tables, but additive and easily rolled back
Iteration	2	A few cycles of review expected
Total	7	Model tier: Sonnet

Note the bump rule never fired here. If that same form had needed a destructive migration on production data, Risk goes to 3 and the whole task bumps up a tier, regardless of the total.

The over-thinking complaint is a context lesson in disguise

Sonnet 5 over-thinks smaller tasks: thinking longer, using more tokens and generally doing more work than is asked for, or necessary. In the aggregate, this will eat into your productivity and cost effectiveness significantly.

The lesson is that we need to be systematically more selective with our models. Scoping tasks ahead of prompt execution is part of cost control, and therefore part of context engineering.

Effort is now a parameter, set with words.

temrel-agentic-toolkit: audit yourself

This week's free Temrel tool is the model-router: the four-axis heuristic above as a Claude Code skill, plus a CLI that parses local Claude Code transcripts and reports actual cost-per-task by model.

The audit function flags where you paid Opus prices for work Sonnet's profile covers and estimates the total overspend. Run it on your own machine and check the last 30 days.

Caveat: the exact dollar figures depend on a pricing file you must verify, and Claude Code's 30-day rolling transcript window caps how far back you can go.

Do this today

1. Score your next three tasks on the four axes before you pick a model.
2. Send anything scoring 7-9 to Sonnet 5 and judge the output blind.
3. Install model-router and run model-router audit --since 30. Note the overspend number.
4. Add one sentence of task scoping ("this is a small, single-file change") to your next prompt and watch the token count.

Why this matters

Model routing is about to become a default, must-have skill, just like instance sizing. As agents multiply, per-task allocation compounds quickly into both speed and money.

As an agentic engineer, your job description just grew by one line: resource allocation is part of the craft now.

Enjoyed this? Subscribe to Temrel for a new issue every week.