Opus 4.7 wins on almost every benchmark. But the new tokenizer can raise your effective costs by up to 35%. Here's exactly when to upgrade — and when to stay.
You do agentic coding (+10.9 SWE-bench points), need vision accuracy (54.5% → 98.5%), require 1M+ context, or are starting something new. The performance gap is real.
You're cost-sensitive and the new tokenizer would meaningfully increase your spend on simple or routine tasks.
| Metric | Opus 4.7 | Opus 4.6 | Note |
|---|---|---|---|
| SWE-bench Pro | 64.3% (#1)▲ | 53.4% | +10.9 points, beats GPT-5.4 (57.7%) |
| SWE-bench Verified | 87.6%▲ | 80.8% | +6.8 points |
| CursorBench (coding agents) | 70%▲ | 58% | +12 points |
| GPQA Diamond (reasoning) | 94.2%▲ | ~90% | Near-tied with GPT-5.4 Pro (94.4%) |
| Vision accuracy | 98.5%▲ | 54.5% | Near-perfect — the biggest leap |
| Vision resolution | 2,576px (~3.75MP)▲ | ~800px | 3× higher resolution |
| Context window | 1,000,000 tokens▲ | 200,000 tokens | 5× larger |
| Input pricing (per 1M) | $5.00 | $5.00 | Same nominal rate |
| Output pricing (per 1M) | $25.00 | $25.00 | Same nominal rate |
| Effective cost (tokenizer) | Up to 1.35× higher | Baseline▲ | New tokenizer encodes more tokens |
| Effort levels | low / medium / high / xhigh / max▲ | low / medium / high / max | New xhigh level added |
| Speed | Moderate (similar to 4.6) | Moderate | Not meaningfully different |
53.4% → 64.3%. Biggest single-version coding jump in the Opus family. Now ahead of GPT-5.4 (57.7%) and Gemini 3.1 Pro (54.2%).
54.5% → 98.5%. Near-perfect accuracy on diagrams, charts, and documents. Combined with 3× higher resolution (2,576px).
200K → 1M tokens. Fit entire codebases, legal docs, or multi-hour transcripts. A category change, not just an increment.
Opus 4.7 uses a new tokenizer that encodes the same text into up to 1.35× more tokens than Opus 4.6. The per-token price ($5/$25 per million) is identical — but your token count per request is higher.
| Content type | Tokenizer impact | Migration risk |
|---|---|---|
| English prose (short) | ~1.0–1.05× | Low — negligible difference |
| Long documents / books | ~1.1–1.2× | Medium — test cost impact |
| Code / technical content | ~1.15–1.35× | High — benchmark before migrating |
| Cached prompts | Cache boundaries may shift | Medium — re-verify cache hit rate |
5 steps to safely move a production app from Opus 4.6 to 4.7.
Run your 10 most common prompts against claude-opus-4-7-20260416 and compare token counts vs Opus 4.6. Anything over 1.1× warrants cost modelling before you migrate.
Opus 4.7 follows instructions more literally. Any prompt that relied on Opus 4.6 inferring intent or being lenient about format may produce different output. Diff a sample of 50–100 responses.
Replace claude-opus-4-6-20251101 (or whichever version you're pinned to) with claude-opus-4-7-20260416 in your API calls.
If you use extended thinking, test whether xhigh gives you better results than high at a lower latency cost than max. This is one of Opus 4.7's most useful new controls.
Because the tokenizer changed, your cache break points may shift. Re-verify that your cache prefixes are still hitting the boundaries you expect.
Yes, across almost every metric. SWE-bench Pro improved from 53.4% to 64.3%, vision accuracy jumped from 54.5% to 98.5%, and the context window grew from 200K to 1M tokens. For new projects, especially agentic coding or vision-heavy work, Opus 4.7 is the clear choice.
The per-token price is identical: $5/$25 per million tokens. However, Opus 4.7 uses a new tokenizer that can encode the same text into up to 1.35× more tokens. So the same prompt that costs $1 on Opus 4.6 may cost up to $1.35 on Opus 4.7. Always benchmark your specific prompts before migrating high-volume workloads.
For greenfield projects: yes, always start on 4.7. For existing production apps: test first. The new tokenizer changes how text is encoded, which can affect cost and occasionally output length. The model also interprets instructions more literally, which can break prompts that relied on Opus 4.6's more lenient behavior.
Opus 4.7 uses an updated tokenizer that maps the same text to up to 1.35× more tokens depending on content type. Code and technical text tends to see the largest increases. The per-token price is unchanged at $5/$25 per million, but your token count per request will be higher.
The full model ID is claude-opus-4-7-20260416. Use this in your API requests.
Probably not. For routine tasks like summarization, simple Q&A, or short writing tasks, Opus 4.6 is sufficient and will be cheaper due to the tokenizer difference. Save Opus 4.7 for complex coding, long-context work, or vision tasks where the improvements are meaningful.
