Claude Opus 5 Pricing: Why the Same $5/$25 Rate Can Still Raise Your Bill

Key takeaways

• Claude Opus 5 launched on 24 July 2026 at $5 per million input tokens and $25 per million output tokens, identical to Opus 4.8.
• Thinking now runs by default on every request, and it cannot be disabled at the xhigh or max effort levels, where the API returns a 400 error if you try.
• At an 80/20 input/output split the blended rate is $9.00 per million tokens; at 60/40 it is $13.00.
• Against a 4x H200 spot cluster at $318.17 a day, the $9.00 blend crosses over around 35M tokens a day; against an 8x H200 on-demand cluster at $919.46 a day, around 102M.
• The effort setting, not the thinking toggle, is now the primary cost lever, and a higher real output share pulls that crossover point earlier.

What actually changed in Opus 5 pricing?

Nothing on the rate card. The headline rate is the same $5/$25 per million that Opus 4.8 has carried since May. Fast Mode is unchanged at $10/M input and $50/M output. Batch is still a flat 50% off both directions, $2.50/M and $12.50/M. Cache-hit reads still bill at 0.1x base input, which is $0.50/M. The context window stays at 1M tokens with 128k max output.

Three behavioural defaults moved instead. Thinking runs unless you explicitly disable it. The effort ladder gained real teeth at xhigh and max. And the minimum cacheable prompt dropped from 1,024 tokens to 512, so shorter system prompts can now cache with no code changes.

Why does a default change your cost?

Because max_tokens is a hard ceiling on thinking tokens plus response text combined, and thinking tokens bill as output at $25/M. A request that used to spend its whole output budget on the answer now spends part of it reasoning first.

That is a shift in your input/output mix, not in your price. Run the arithmetic both ways:

• 80/20 split: (0.8 x $5) + (0.2 x $25) = $9.00 per million tokens
• 60/40 split: (0.6 x $5) + (0.4 x $25) = $13.00 per million tokens

If a workload that measured 80/20 on Opus 4.8 with thinking off now produces more output tokens on the same prompts, your effective blend drifts toward the $13.00 figure without a single number changing on the pricing page.

This is the observation most cost models miss. Blended rate is usually treated as a derived constant, calculated once from the price card and pasted into a spreadsheet. It is actually a measured quantity that a provider can move by changing a default. Track it as a metric, not an assumption.

Which lever should you pull now?

The effort setting. On Opus 5, thinking: {"type": "disabled"} is only accepted at high effort or below; set it alongside xhigh or max and the API returns a 400. Anthropic's own framing is that low and medium effort produce strong quality at a fraction of the tokens and latency of the higher settings.

So the practical sequence is: sweep low, medium, and high on your own eval set before assuming you need xhigh, and treat effort as the dial you tune for cost. Caching is the second lever. On a 50k input, 15k output session, caching 40,000 of those input tokens cuts the bill by roughly a third before you touch effort at all.

When does self-hosting actually win?

The crossover math is unchanged in structure because the sticker rate did not move. Using the $9.00/M blend against live H200 cluster pricing:

| Daily volume | Opus 5 API | 4x H200 spot | 8x H200 on-demand | | --- | --- | --- | --- | | 10M tokens/day | $90.00 | $318.17 | $919.46 | | 35M tokens/day | $315.00 | $318.17 | $919.46 | | 102M tokens/day | $918.00 | $318.17 | $919.46 |

That puts breakeven near 35M tokens a day on a 4x H200 spot cluster and near 102M on an 8x H200 on-demand production cluster. Open-weight candidates in that class include GLM-5.2 (744B total, roughly 40B active, MIT licence) and DeepSeek V4-Flash (284B total, 13B active).

The caveat matters more than the table: if your real blend sits above $9.00/M because output share climbed, the crossover arrives earlier than these rows suggest. Cluster costs are also flat only up to a throughput ceiling, so benchmark your own tokens per GPU-hour before committing capacity.

One line to take away: measure your blended rate rather than deriving it, because the rate you pay is set by your token mix, not by the price card.

You can re-run this comparison at your own token split at calcaas.com.