Which Claude, When

A working guide for knowledge entrepreneurs choosing between Haiku 4.5, Sonnet 4.6, Opus 4.7, and Opus 4.8 — built from what Anthropic actually published, not what the marketing says.

The landing, up front

The right Claude for the job is almost never the most expensive one. It’s almost never the newest one either.

It’s the cheapest model that crosses the quality bar for your task. Everything else is wallet leakage.

Here’s the one-page version. The rest of this guide is the case for why.

That’s the field manual. Now the case.

The honest caveat (read this once, then we move on)

Anthropic publishes one score per model. At max effort. That’s it.

There is no public table showing Opus 4.7 at low vs Opus 4.6 at xhigh. Anyone who tells you “Sonnet at medium beats Opus at low” is guessing — including me. The only per-effort number Anthropic has released is this: on their internal agentic coding eval, Opus 4.7 goes from ~71% at xhigh to ~74.5% at max. A 3-point lift for 2× the thinking tokens.

That’s the entire public dataset on the effort-vs-performance question.

So this guide tells you what we do know — model-to-model — and where the answer lives on your own keyboard.

The four models, side by side

A few things to notice before we move on.

SWE-bench Verified (coding): Haiku → Sonnet is a 6-point lift. Sonnet → Opus 4.7 is an 8-point lift. Opus 4.7 → Opus 4.8 is a 1-point lift at the same price.

GPQA Diamond (PhD-level reasoning): The Sonnet → Opus gap blows out from 8 points (on coding) to 20 points (74.1 vs 94.2). This is the single most important pattern in the chart — Sonnet looks fine on coding, then falls off a cliff on reasoning. If your task asks the model to think rather than transform, that 20-point gap is what you’re paying Opus for.

Dashes mean Anthropic didn’t publish that combination. Haiku has no published GPQA score. Sonnet has no published SWE-Pro. The gaps are the data.

Coding work — the sweet spot view

What the chart is saying:

• Each dashed line is the cost range that model covers from low → max effort, using Anthropic’s published 1× to 6× token ratio.
• The dot on the right end is the only score we know — published, max effort, SWE-bench Verified.
• Where the lines overlap horizontally is where the cost overlap is real. The score at those non-max points is the part you’d have to measure yourself.

The interesting overlaps:

Rule of thumb for coding: the generational gap between Opus versions is larger than the gap between effort levels within one version. So a newer Opus at lower effort beats an older Opus at higher effort, most of the time, at lower cost.

Creative, advisory, writing, sales — the sweet spot view

Different benchmark, different story. This is GDPval-AA — Anthropic’s eval for “economically valuable knowledge work” across finance, legal, advisory, writing, business analysis. The Elo scale: roughly 100 Elo gap = noticeably better; 200+ = different league.

Three things change versus the coding chart:

1. Haiku falls off a cliff. 902 Elo. That’s 500+ Elo below Sonnet. On SWE-V the Haiku → Sonnet gap was 6 points. Here it’s the gap between “useful junior” and “barely coherent.” Haiku is a transformation model — extract, classify, summarize, route. It is not a thought partner. For writing, advisory, brand voice — Haiku is the wrong default at any cost.

2. Sonnet gets more competitive. Sonnet → Opus 4.7 is only 125 Elo here, against 8 points on SWE-V. That’s roughly a 65/35 head-to-head, not a blowout. For bulk creative work — newsletter drafts, internal memos, advisory summaries — Sonnet at high is probably close enough for the cost saving.

3. The 4.7 → 4.8 upgrade actually matters. +96 Elo. That’s a meaningfully bigger jump than the +1 SWE-V point you got on coding. If you write or advise for a living, 4.8 is the move. Same price as 4.7, real lift.

The interesting overlaps:

The effort question (the part nobody can answer)

This is the part I have to be straight with you about.

Anthropic gave us one per-effort delta on one benchmark for one model. Opus 4.7 xhigh → max = +3 points at 2× the tokens.

That’s it. The rest of the curve is empty. Every chart in this guide shows a dashed line because the public data does not exist to fill it in.

What that means in practice:

• You cannot pick an effort level from a blog post. Including this one.
• The defaults are reasonable. Claude Code defaults Opus to xhigh. That’s based on Anthropic’s internal data, which they haven’t shared. Trust the default until you have a reason not to.
• Where it matters, measure it. If you’re shipping production traffic through these models, the right effort level for your task distribution is an empirical question, not a reading question.

The fastest way to answer it for your own work is below.

The 20-minute private eval (the part that actually answers your question)

Stop reading benchmarks. Run one. Here’s the minimum-viable version.

Setup:

1. Pick two tasks from your actual workload. Not toy tasks — real ones. e.g., “rewrite this article lead in my voice” and “summarize this client call into three insights.”
2. Grab 15 real inputs per task. Last month’s emails, last week’s drafts. Whatever you actually do.
3. Pick three (model, effort) cells worth testing. For creative work I’d start with: Sonnet 4.6 @ high, Opus 4.8 @ medium, Opus 4.8 @ xhigh.
4. Run each input through each cell. 15 × 2 × 3 = 90 calls.

Score:

5. Either grade them yourself with a 1-5 rubric (voice fidelity, insight depth, polish, total) — or hand the outputs to Sonnet 4.6 as a judge with the same rubric.
6. Average the scores per cell. Tally cost per cell.

Decide:

7. Pick the cheapest cell that hits your quality bar.

Roughly $30-50 of inference. Half a day of setup. You’d know more about your real model/effort sweet spot than every blog post on the internet combined.

This is the move. The benchmark for your work is your work.

The decision rules — printable version

Three rules. Keep them visible.

1. Default cheap. Escalate on failure. Start with the cheapest model that might work. Move up only when you see it fail. The model you reach for first is usually too expensive. The smart intern can handle 80% of the work and leave you the 20%.

2. The generational gap usually beats the effort gap. A newer model at lower effort almost always beats an older model at higher effort, at lower cost. Opus 4.7 → 4.8 is the same price — no reason to stay on 4.7. Same logic when Opus 4.9 ships next quarter.

3. The benchmark for your work is your work. Public benchmarks are model-to-model. Your task distribution is yours. If a routing decision matters enough to argue about, it matters enough to run a 90-call private eval before you commit.

What I’m watching next

A few things I’m keeping an eye on, in case you are too.

• Whether Anthropic ever publishes per-effort scores. They have the data. If they release it, this whole guide gets cleaner. Don’t hold your breath.
• Third-party effort sweeps. Artificial Analysis, EQ-Bench, Vellum — any of them could do this. As of June 2026, none have.
• The next Opus release. Opus 4.7 → 4.8 came in at the same price point with real gains on the creative side. If that pattern holds for the next release, the upgrade decision stays easy. If pricing moves, the whole sweet-spot math shifts. See rule 3.

That’s the guide. Three models, two charts, three rules, one private eval. Print it, share it, use it.

If you’re routing a lot of traffic through these models and want a hand setting up your own eval, that’s exactly the kind of thing we work through together in the group.

Let’s serve people, do good, have fun and make money — abundantly. Namaste.

Sources: Introducing Claude Opus 4.7 (Anthropic) · Claude Opus 4.8 benchmarks explained (Vellum) · Opus 4.7 xhigh mode explained (apiyi) · GDPval-AA Leaderboard (llm-stats) · Claude Benchmarks 2026 (morphllm)