GPT-56T 861 —
MUSE-SPK 837 +0.2%
GPT-56SC 790 -4.6%
GLM-5 781 -0.4%
CL-OP55X 780 -5.1%
GROK-46H 780 -5.1%
QWEN-38X 748 -9.2%
GPT-6A 743 -9.4%
KIMI-K3X 742 —
CL-FAB5H 698 -6.1%
CL-OP5H 675 -6.2%
GEM-38FH 672 -0.7%
CL-OP5X 670 -5.5%
CL-OP55H 668 —
CL-OP46H 657 -5.9%
CL-OP47H 648 -6.1%
GPT-56S 618 -0.6%
GEM-37FH 610 -7.2%
GEM-36FH 593 —
CL-OP48H 588 —
CL-OP47 581 -0.2%
GEM-35FH 580 —
GPT-55H 541 -7%
INKL 531 —
GEM-31P 512 -0.2%
CL-OP46 498 +0.4%
GEM-3P 498 -0.2%
CL-OP48 492 +0.4%
GPT-52 464 —
GPT-55 423 —
GPT-56T 861 —
MUSE-SPK 837 +0.2%
GPT-56SC 790 -4.6%
GLM-5 781 -0.4%
CL-OP55X 780 -5.1%
GROK-46H 780 -5.1%
QWEN-38X 748 -9.2%
GPT-6A 743 -9.4%
KIMI-K3X 742 —
CL-FAB5H 698 -6.1%
CL-OP5H 675 -6.2%
GEM-38FH 672 -0.7%
CL-OP5X 670 -5.5%
CL-OP55H 668 —
CL-OP46H 657 -5.9%
CL-OP47H 648 -6.1%
GPT-56S 618 -0.6%
GEM-37FH 610 -7.2%
GEM-36FH 593 —
CL-OP48H 588 —
CL-OP47 581 -0.2%
GEM-35FH 580 —
GPT-55H 541 -7%
INKL 531 —
GEM-31P 512 -0.2%
CL-OP46 498 +0.4%
GEM-3P 498 -0.2%
CL-OP48 492 +0.4%
GPT-52 464 —
GPT-55 423 —
← Back to feed

OpenRouter's Million-Request Study: Opus 4.7 Costs 27% More in the Agentic Sweet Spot

When Anthropic launched Claude Opus 4.7, they disclosed the new tokenizer would inflate token counts by 1.0–1.35x depending on content type. OpenRouter’s analysis of over one million real API requests gives the first empirical answer to what developers actually care about: what does that mean for the bill?

The answer depends almost entirely on prompt length and cache hit rate.

The Numbers

OpenRouter tracked a switcher cohort — users whose dominant model shifted from Opus 4.6 to Opus 4.7 — across more than one million requests. Net cost change per million normalised tokens:

Prompt SizeNet Cost Change
Under 2K tokens-1.6%
2K – 10K tokens+27.2%
10K – 25K tokens+25.2%
25K – 50K tokens+21.3%
50K – 128K tokens+11.9%
128K+ tokens+15.3%

The new tokenizer produces 32–45% more native tokens — that is the raw inflation figure. What varies is how much of that lands in the prompt cache.

Why the Middle Range Gets Hit Hardest

Prompt caching discounts repeated content at 90%. For very short prompts under 2K, the tokenizer adds tokens but completions shrank 62% — tighter queries, shorter responses. Net result: -1.6%.

For the 2K–10K range — the typical footprint of an agentic coding turn with tool calls, system instructions, and a few context exchanges — caching absorbs only 56% of the inflation. Completions also grew 4% longer. That combination produces the worst outcome in the dataset: +27.2%.

At 128K+, caching absorbs 93% of the inflation. Long-context retrieval workloads with heavy system prompts that repeat across calls benefit most from Opus 4.7’s cache characteristics. The 15.3% increase in that range is partly offset at extreme cache hit rates.

The full decomposition for each bucket:

Prompt SizeTokenizer InflationCache AbsorptionCompletion ChangeNet
Under 2K+45%—-62%-1.6%
2K – 10K+42%56%+4%+27.2%
10K – 25K+34%9%+30%+25.2%
25K – 50K+32%64%+13%+21.3%
50K – 128K+32%77%+19%+11.9%
128K++33%93%+26%+15.3%

Practical Implications

Anthropic’s nominal Opus 4.7 pricing is $5/M input, $25/M output — identical to Opus 4.6. The tokenizer shifts the effective rate without touching the published number.

Agentic coding pipelines in the 2K–10K range — a single Codex-style task, a code-review loop, or a multi-tool agent turn — will see real cost increases of 20–27% for identical work versus Opus 4.6. That is not a rounding error for teams running high-volume agent workflows.

Long-context retrieval agents (50K+) with aggressive caching and repeated system prompts land in the 12–15% range, which is less severe and partially predictable from cache configuration.

The headline 32–45% native token inflation figure was technically accurate. The empirical data shows that prompt caching and completion-length shifts compress the actual cost impact — but the compression is uneven, and it is worst at the prompt sizes that represent most agentic API spend.