Claude Opus 5 Halves the Price of Near-Frontier AI: Rerun Your Model Math

Anthropic released Claude Opus 5 this week at $5 per million input tokens and $25 per million output tokens, half the price of Claude Fable 5, and as of today it sits first on the Artificial Analysis Intelligence Index. If any part of your product or workflow pays for AI by the token, this is a billing event, not just tech news: the correct response is to rerun the assignment of every AI workload you pay for and move each one to the cheapest model that still passes your quality bar. Below are the current prices for the whole Claude line, worked cost math for a realistic workload, and a repeatable procedure you can reuse at every future launch.

What shipped, and what it costs

The launch announcement positions Opus 5 as coming "close to the frontier intelligence of Claude Fable 5 at half the price." The price itself did not move: $5 and $25 per million tokens is exactly what Opus 4.8 cost. What moved is the capability at that price. Anthropic's own launch claims include performing within 0.5% of Fable 5's peak score on CursorBench 3.2 at half the cost per task, and beating Fable 5 on OSWorld 2.0 at just over a third of the cost. Those are vendor-run numbers and should be read that way.

The third-party read is more interesting. On the Artificial Analysis Intelligence Index (version 4.1, which averages nine evaluations including Terminal-Bench, GPQA Diamond and Humanity's Last Exam), Opus 5 at maximum reasoning effort currently ranks first with a score of 61, one point ahead of Fable 5 and two ahead of GPT-5.6 Sol (checked July 25, 2026). One or two index points is within squinting distance of a tie. The price difference is not: it is a factor of two.

Current list prices on the Claude API pricing page (per million tokens, checked July 25, 2026):

Two modifiers matter for real bills. The Batch API halves both sides for anything that can wait for asynchronous processing. Prompt cache hits are billed at 10% of the input price, which changes everything for workloads that resend the same system prompt or document on every call.

Why a launch is a billing event

Model launches quietly reprice work you assigned months ago. The workload you put on Fable 5 in the spring because nothing cheaper was good enough may now clear the same bar on Opus 5 at half the cost, and the one you put on Opus 4.8 may clear it on Sonnet 5 at a fifth. Nobody sends you an invoice adjustment when that happens; the savings only exist if you re-test. This is the same discipline as the quarterly AI subscription audit, applied to API spend, and it compounds with the pricing pressure we covered when open-weight models became the budget tier.

The math for one real workload

The arithmetic that follows is ours, using the list prices above. Take a concrete workload: a document summarization feature that processes 10 million input tokens and produces 1 million output tokens per month. Monthly cost by model:

• Fable 5: (10 x $10) + (1 x $50) = $150
• Opus 5: (10 x $5) + (1 x $25) = $75
• Sonnet 5 at introductory pricing: (10 x $2) + (1 x $10) = $30 (rises to $45 on September 1, 2026)
• Haiku 4.5: (10 x $1) + (1 x $5) = $15

Now add the modifiers. If the summaries are not user-facing in real time, the Batch API takes Opus 5 to $37.50. If 8 of the 10 million input tokens are a repeated prompt served from cache, Opus 5 input drops to (2 x $5) + (8 x $0.50) = $14, and the monthly total to $39. A workload that "needs Opus" can cost anywhere from $75 to under $40 depending on how you call it, before you touch the model choice at all.

One caveat when comparing across model generations: Anthropic's pricing page notes that Claude 4.7 and later models (including Opus 5 and Fable 5) use a newer tokenizer that produces roughly 30% more tokens for the same text than Sonnet 4.6 and earlier. Comparing list prices across that boundary understates the newer models' effective cost per document, so benchmark cost on your actual payloads, not on the price table.

The downgrade ladder

Here is the procedure worth institutionalizing, even as a team of one:

1. Keep an eval set. Collect 25 to 50 real cases from your product (inputs plus what a correct output must contain). Twenty minutes of copy-paste from logs is enough to start.
2. Define the bar. A pass rate you will not go below, for example 46 of 50.
3. On launch day, run one tier down. Test your current model and the next cheaper one on the set. An open-source eval runner like promptfoo works, and so does a 40-line script.
4. Switch on evidence, not vibes. If the cheaper tier passes the bar, move the workload and note the date. If it fails, you have documented why you pay the premium.
5. Repeat per workload, not per app. Your classifier, your summarizer and your chat feature can and should sit on different models.

If you pay by subscription instead

The same launch moved the subscription math. Per Claude's plan pricing (checked July 25, 2026), Pro costs $17 per month billed annually or $20 monthly, and Max starts at $100 per month. The announcement makes Opus 5 the default model on Max and the strongest model on Pro. That last part is the quiet headline: a $20 plan now includes the model currently ranked first on a major independent index. If you bought Max primarily for frontier-model access rather than for its higher usage limits, that reasoning deserves a re-check this month. The rest of the audit logic from our piece on usage-based AI pricing applies unchanged: what you pay should track what you would actually lose by downgrading.

The decision rule to keep: every AI workload runs on the cheapest model that passes your eval set, re-tested on every launch day. Today that test takes an hour and will likely cut the bill of anything currently sitting on a top-tier model roughly in half. And set one calendar reminder for August 31: Sonnet 5's introductory pricing ends the next day, which is exactly the kind of silent 50% increase this habit exists to catch.