GPT-56T 861 —
MUSE-SPK 837 —
GPT-56SC 789 -0.1%
GLM-5 781 —
CL-OP55X 779 -0.1%
GROK-46H 779 -0.1%
QWEN-38X 748 —
GPT-6A 743 —
KIMI-K3X 742 —
CL-FAB5H 697 -0.1%
CL-OP5H 674 -0.1%
GEM-38FH 672 —
CL-OP5X 669 -0.1%
CL-OP55H 667 -0.1%
CL-OP46H 656 -0.2%
CL-OP47H 647 -0.2%
GPT-56S 617 -0.2%
GEM-37FH 609 -0.2%
GEM-36FH 592 -0.2%
CL-OP48H 587 -0.2%
CL-OP47 580 -0.2%
GEM-35FH 579 -0.2%
GPT-55H 540 -0.2%
INKL 531 —
GEM-31P 511 -0.2%
CL-OP46 498 —
GEM-3P 498 —
CL-OP48 492 —
GPT-52 464 —
GPT-55 423 —
GPT-56T 861 —
MUSE-SPK 837 —
GPT-56SC 789 -0.1%
GLM-5 781 —
CL-OP55X 779 -0.1%
GROK-46H 779 -0.1%
QWEN-38X 748 —
GPT-6A 743 —
KIMI-K3X 742 —
CL-FAB5H 697 -0.1%
CL-OP5H 674 -0.1%
GEM-38FH 672 —
CL-OP5X 669 -0.1%
CL-OP55H 667 -0.1%
CL-OP46H 656 -0.2%
CL-OP47H 647 -0.2%
GPT-56S 617 -0.2%
GEM-37FH 609 -0.2%
GEM-36FH 592 -0.2%
CL-OP48H 587 -0.2%
CL-OP47 580 -0.2%
GEM-35FH 579 -0.2%
GPT-55H 540 -0.2%
INKL 531 —
GEM-31P 511 -0.2%
CL-OP46 498 —
GEM-3P 498 —
CL-OP48 492 —
GPT-52 464 —
GPT-55 423 —
← Back to feed

Opus 4.7's Hidden Cost: New Tokenizer Adds Up to 35% More Tokens Per Prompt

Claude Opus 4.7 launched on April 16 at $5/$25 per million input/output tokens — identical to Opus 4.6 on the price sheet. That headline held for about 48 hours before enterprise developers working at scale noticed their API bills were climbing anyway.

Anthropic confirmed the mechanism in its own documentation: Opus 4.7 introduces a new tokenizer that can consume up to 35% more tokens for the same text input, depending on content type. The additional consumption is attributed to optional higher effort levels and improved reliability. The result is a de facto price increase of up to 35% on identical workloads, with no change to the listed rate.

The Tokenizer Inflation Mechanism

The tokenizer is the preprocessing step that converts raw text into the numeric tokens an LLM processes. A more granular tokenizer — one that splits text into smaller chunks — increases token count for a given passage without changing what the model reads. Opus 4.7’s new tokenizer does this: the same prompt that cost $5 per million tokens on Opus 4.6 can cost the equivalent of $6.75 per million tokens in practice on Opus 4.7 at the upper bound.

For low-volume users, the difference is negligible. For enterprise deployments running agentic loops across large codebases, legal corpora, or multi-turn sessions — the workloads Anthropic explicitly targets with Opus 4.7 — the effective cost difference compounds quickly. A workflow budgeted at $10,000/month on Opus 4.6 could land at $13,500/month on Opus 4.7 before any other changes.

The Long-Context Regression

Independent analysis by Epsilla, aggregated from developer communities on Hacker News and GitHub, identified a separate issue: Opus 4.7’s MRCR (Multi-hop Retrieval and Contextual Reasoning) score dropped from 78.3% on Opus 4.6 to 32.2%. MRCR measures a model’s ability to locate and synthesise information distributed across a long document — the exact capability needed for RAG pipelines, legal document review, and financial analysis.

The score divergence is notable precisely because Opus 4.7 simultaneously improved on coding benchmarks (SWE-bench Verified: 87.6%, up from 80.8%; CursorBench: 70%, up from 58%) and on GDPVal-AA economically-productive tasks (ELO 1,753 vs. GPT-5.4’s 1,674). The model improved on the benchmarks that dominate launch coverage while apparently regressing on a retrieval dimension that enterprise production use cases depend on.

Anthropic has not publicly addressed the MRCR finding as of publication.

What Changed, What Didn’t

The capability improvements are real. On a 93-task internal coding benchmark, Opus 4.7 resolved 13% more issues than Opus 4.6 and solved four tasks neither 4.6 nor Sonnet 4.6 could complete. File system-based memory handling improved, allowing the model to persist context across multi-session agentic workflows without re-injecting the full prior context each turn. Vision improved substantially — images up to 3.75 megapixels vs. the previous ceiling — and instruction-following in long autonomous runs is measurably better.

The new xhigh effort level for extended thinking adds a reasoning tier above high, giving developers finer control over the reasoning-vs-latency tradeoff. Task budgets launched in public beta alongside the model.

The profile that emerges is a model optimised for agentic coding tasks and short-to-medium context windows, at the cost of long-context retrieval fidelity. For teams routing all workloads through a single Anthropic model, that trade-off is invisible until production.

Implications for Enterprise Routing

The practical effect: Opus 4.7 is a strong default for tool-calling, autonomous coding, and document generation tasks with bounded context. For RAG, long-document synthesis, or legal review workflows — where MRCR performance is load-bearing — routing to Opus 4.6 or waiting for clarification from Anthropic on the retrieval regression is the more conservative choice.

The tokenizer change creates a separate budgeting problem. Organisations that modelled Opus 4.7 adoption cost-neutral relative to Opus 4.6 should retest with their actual workloads before committing.

Key Numbers

BenchmarkOpus 4.6Opus 4.7Direction
SWE-bench Verified80.8%87.6%↑
CursorBench58%70%↑
MRCR (long-context retrieval)78.3%32.2%↓
GDPVal-AA ELO~1,6741,753↑
Effective token cost vs. 4.61.0×up to 1.35×↑