Claude Opus 4.7 Arrives | 70% on Hard Coding Tasks · 3× Resolution
機械翻訳 / Machine-translated
A
AI Friends
2026年4月26日・20分で読める
機械翻訳 / Machine-translated
@aifriends
AI Friends（https://aifriends.jp）のクロスポスト公式アカウント。AIツールの紹介・使い方・できることを、中学生でもわかるやさしい日本語で届けます。
Have you ever wondered, "Can AI coding assistance actually be useful?" On April 16, 2026, Anthropic announced an update that answers that question.
The latest model, Claude Opus 4.7, scored more than 10 points higher than its predecessor on hard coding tests, and its image recognition now supports more than 3× the resolution.
This article breaks down the full picture in plain language, using numbers and analogies.
Let's start with the basics.
Claude Opus 4.7 is the latest large language model (LLM) developed by AI company Anthropic.
The "Opus" line in the Claude series is the highest-end tier, designed for demanding tasks.
Version 4.7 was released on April 16, 2026, as a direct upgrade from the previous "Opus 4.6."
It is currently Anthropic's top-ranked model among those available to the general public via API.
Note that a model called "Claude Mythos Preview" exists at an even higher tier as a limited-access release, but it is not yet available to general users.
The three areas most improved in Opus 4.7 are coding ability, image recognition, and instruction following. Think of it like a seasoned craftsperson with five years of experience who has gone through intensive training to move up one more level.
The most attention-grabbing improvement in this update is the boost in coding performance. Let's look at the numbers.
SWE-bench Verified is a test that measures how well an AI can fix real GitHub bug tickets. Claude Opus 4.7 achieved 87.6% here.
On the harder SWE-bench Pro, it scored 64.3%.
That's an improvement of 10.9 percentage points over the previous Opus 4.6's 53.4%.
Rival GPT-5.4 sits at 57.7% and Gemini 3.1 Pro at 54.2%, meaning Opus 4.7 has leaped past the major competition in one move.
The difference shows up not just in benchmarks but in actual development environments. When the well-known AI code editor Cursor conducted its own independent evaluation, Opus 4.7 came in at 70%, up from Opus 4.6's 58%.
Furthermore, on the SWE-Bench tests published by Rakuten, Opus 4.7 completed approximately 3× as many tasks as Opus 4.6. Think of it as adding a star employee to your team who can handle the workload of three people single-handedly.
On GDPVal-AA, which measures knowledge work beyond coding, Opus 4.7 recorded a score of 1,753—well above GPT-5.4's 1,674 and Gemini 3.1 Pro's 1,314.
This indicates an across-the-board improvement in complex analysis, document creation, and reasoning as well.
Alongside coding, image recognition has seen a major upgrade.
Up through Claude Opus 4.6, the maximum image resolution supported was approximately 1.15 megapixels (MP).
With Opus 4.7, that ceiling rises to a maximum of 3.75 MP, supporting images with a longest side of up to 2,576 pixels.
Wondering what a 3× jump actually means in practice? Imagine being able to hand a photo taken on your smartphone directly to the AI without resizing it first.
It's like going from struggling to read small print on an A4 sheet of paper to being able to read it clearly without a magnifying glass.
Three concrete scenarios where this helps:
Opus 4.7 also introduces two new functional elements, particularly relevant for engineers and power users.
There is a concept called "reasoning level" that controls how deeply an AI thinks. Previous versions of Claude offered two levels: "high" and "max."
Opus 4.7 introduces a new intermediate level: "xhigh" (extra high).
This addresses the nuanced need of "I want slightly deeper thinking, but not full power—that's too slow and expensive."
Think of it like a car's gears going from 1st, 2nd, and 3rd to 1st, 2nd, 2.5th, and 3rd—finer control matched to the situation.
Task Budget is a cost management feature for AI agent workflows that automatically handle multiple steps. It is currently available as a public beta via the Claude Platform API.
Previously, you could only manage "how many tokens (the AI's processing units) are used in a single exchange." With Task Budget, you can set an overall budget—"a maximum of X tokens for this entire project"—spanning multiple steps.
For example, if you ask the AI to "refactor these 100 codebases," it will automatically adjust if you're at risk of going over budget mid-way. It's like a personal finance app that stops you from worrying about the AI quietly overspending.
"Okay, Claude Opus 4.7 has improved—but how does it stack up against other AIs?" Here's your answer.
| Model | SWE-bench Pro | SWE-bench Verified | GDPVal-AA |
|---|---|---|---|
| Claude Opus 4.7 | 64.3% | 87.6% | 1,753 |
| GPT-5.4 | 57.7% | — | 1,674 |
| Gemini 3.1 Pro | 54.2% | 80.6% | 1,314 |
| Claude Opus 4.6 | 53.4% | — | — |
For coding and knowledge work, it ranks among the top models currently available to the public.
That said, a few caveats are worth noting.
Anthropic itself already has a non-public model called "Claude Mythos Preview" that is said to deliver even higher performance. Opus 4.7 is the best among generally available models, not across the board.
Also, as discussed below, there is a potential real-world cost increase due to the new tokenizer, so if cost-efficiency is your priority, Claude Sonnet 4.6 remains a strong contender.
Here's a summary of how to access Opus 4.7 from Japan and what it costs.
Visit claude.ai, create an account, and you can start on the free plan.
However, since Opus 4.7 is a more advanced model, subscribing to the Pro plan (approximately $20/month, or $18/month billed annually) is recommended for comfortable use.
Access from Japan works without issues.
Via the Anthropic API, pricing is $5 per 1M input tokens and $25 per 1M output tokens—unchanged from Opus 4.6.
However, there is an important caveat.
Opus 4.7 adopts a new tokenizer (the mechanism that splits text into smaller units), and the same text may consume up to 35% more tokens as a result.
Even though the listed price is the same, your actual costs may increase in practice. We recommend running real-world measurements before deploying to production.
The API is also available through the following cloud platforms:
When switching from Opus 4.6 to 4.7, there are changes in API behavior, so you will need to verify compatibility with your existing systems. Rather than switching production environments immediately, it is safer to confirm behavior in a test environment first before migrating.
A. General availability began on April 16, 2026.
You can use it right now via Claude.ai or the Anthropic API.
Amazon Bedrock and Google Cloud Vertex AI support is also available.
A. Yes, you can try it on Claude.ai's free plan, though there are usage limits.
For heavy use, the Pro plan (approximately $20/month) is recommended.
The API is pay-as-you-go; small-scale testing can start at a very low cost.
A. The listed price is the same as Opus 4.6 ($5 input / $25 output per 1M tokens).
However, the new tokenizer may increase token consumption for the same text by up to 35%.
Actual costs depend on your usage, so we recommend running small-scale tests before migrating.
A. In terms of raw capability, Claude Mythos is more powerful.
However, Mythos is currently only available to a limited set of users as a "limited preview."
As the highest-performance model available to anyone right now, Opus 4.7 is currently the top choice.
A. The Claude series supports Japanese.
Japanese document creation, translation, and code comment generation are all possible.
That said, there may be some accuracy differences compared to English in certain situations, so it is a good habit to review outputs for important tasks.
This article is a cross-post from AI Friends.
