Anthropic just put its prices up 50%

🔭 Last week’s fallback is this week’s liability

In #146 I argued the defender’s problem is permission, not capability, and told you to vet a self-hostable model before you need one. I stand by it. What I missed is what you inherit when you take that advice.

Rent a model through an API and you rent the lab’s legal posture alongside its compute — their usage policy, their filters, their risk allocation. Download weights and none of that comes with them. Article 50’s transparency duties attach substantially to whoever puts the system in front of a person. Self-hosting doesn’t remove the obligation. It removes the co-signer.

Which is why the same seventy-two hours make sense together. On 28 July, 1,000-plus staff at five frontier labs signed Pacing the Frontier, asking Washington for the machinery to coordinate a slowdown. Then everybody cut prices. That isn’t hypocrisy — it’s timing. A hundred thousand deployments are harder to pace than a hundred.

Standing bet from #146: Daybreak or Glasswing drops the invite-only gate by end of Q4 2026. Still open — no movement either way as of Monday.

That’s the reporting. Here’s the new bet — open to disagreement: the first published Article 50 enforcement action naming a company will name a deployer, not a model developer, by the end of Q2 2027. Counts only if the decision is public and the company fined didn’t build the model. Wrong? Reply and tell me where.

Deep dive: the deployer inherits the bill (with the obligation map) →

🎯 PROMPT OF THE WEEK

The Migration Audit

You asked for this one — the migration question kept coming back in the replies. Run it while the old rate still applies. The audit is worth the same either way; the bill isn’t.

You are a cost architect for LLM workloads. I will describe a task

my team currently runs on a hosted model.

TASK: [paste the prompt or describe the workflow]

CURRENT MODEL: [name]

VOLUME: [calls per day or month]

Work through this in order.

1. CLASSIFY on three axes: reasoning depth (shallow/moderate/

deep), factual risk (low/medium/high), failure cost

(recoverable/expensive/unacceptable).

2. NAME the cheapest tier that reliably clears all three, and

say which single axis is binding. If nothing binds, say so.

3. IF I AM OVER-PROVISIONED, rewrite the prompt so a cheaper

model can do the job — add the scaffolding the expensive

model was supplying implicitly: explicit steps, an output

schema, one worked example.

4. GIVE ME a 3-line eval I can run against 20 real examples to

confirm the downgrade before I commit to it.

5. STATE the one failure mode that would justify staying on the

expensive model, in a sentence I could repeat to my CFO.

Why it works: most overspend is a provisioning problem, not a pricing one. Expensive models quietly compensate for vague prompts, so the cost of vagueness lands on the invoice rather than in the output. Step 3 makes that hidden work explicit. Step 5 stops you optimising into a disaster.

Where to be careful: models are optimistic about what cheaper models can do, and they can’t see your data. Step 4 is not optional. Do the arithmetic in tokens, not requests.

Works best on: Claude Opus 5, GPT-5.6 Sol. Use the expensive model — it’s the last job it should do for you.

🛠️ TOOL OF THE WEEK

‘It’s the price comparison site for models, and it doesn’t sell you one.’

Artificial Analysis ★★★★☆ 4 / 5

Where you check whether the Migration Audit’s shortlist survives contact with a benchmark. Free, independent, not owned by anyone selling inference.

— Price, latency and throughput per model, across hosting providers — the same model is not the same price everywhere.

— Intelligence indices across reasoning, coding and agentic benchmarks, updated on release day.

— Open-weight models sit in the same tables as closed ones, which is the comparison that matters now.

The missing star: benchmarks are not your workload. Terminal-Bench 82.7 says V4 Flash is strong at agent tasks in general. It says nothing about your log formats or your tolerance for a wrong answer. Shortlist three, then test them on your own data. Everyone does the first half.

artificialanalysis.ai →

💡 TIP OF THE WEEK

Your prompt didn’t change. Your token count did.

This won’t help if you measure prompts in words, which nearly everyone does. Sonnet 5 shipped with a new tokenizer that produces roughly 30–41% more tokens from identical English than Sonnet 4.6 did. Nothing about the text changed. The bill did — before the rate rise even lands.

1. Log the usage object. Every API response returns one. Store input and output token counts per call, tagged by model and provider.

2. Build a per-vendor multiplier. Push a thousand documents from your real workload through each candidate tokenizer and record tokens per thousand characters. One number per vendor.

3. Apply it before you compare prices. $0.14 against $0.20 per million means nothing until both are multiplied by how many tokens that vendor makes from your text.

Compression ratios vary by vendor, language and content type — code, JSON and non-English text all fragment differently, and two models with identical headline prices can differ by 40% on the same corpus. An hour’s work, and the only way to tell whether a price cut was actually a price cut.