Claude Opus 4.7 launches: what actually changed for developers

Another week, another AI model launch — but this one is worth reading past the headline. On April 16, 2026, Anthropic released Claude Opus 4.7, and the interesting part is not the marketing copy. It is the receipts. Same price as Opus 4.6. Real jump on SWE-bench Pro. One-third the tool-use errors. And — crucially — a migration that is basically a one-line model-ID change.

If you write code with Claude through the API, Claude Code, Cursor, or Bedrock, this is probably the cleanest upgrade path Anthropic has shipped in a while. But there are a few quiet behaviour changes that will bite if you ignore them. This guide walks through what Opus 4.7 actually changes for developers, what it costs in rupees, how to migrate from 4.6 without surprises, and where it still falls short. No made-up benchmarks, no "game-changer" language — just the parts that matter when you are shipping on a budget.

What Anthropic actually shipped on April 16

Opus 4.7 went generally available on April 16, 2026, roughly two months after Opus 4.6. The model ID for the API is claude-opus-4-7. It is live across the Claude API, Amazon Bedrock, Google Cloud Vertex AI, Microsoft Foundry, and GitHub Copilot's Claude integration. Claude Code also defaults to it on fresh installs.

The important pricing detail: nothing changed. Input stays at $5 per million tokens, output at $25 per million — the exact same rates as Opus 4.6. Prompt caching still offers up to 90% savings on repeat context, and batch processing saves another 50% on non-urgent work. The full 1 million token context window is included at standard pricing, not gated behind a premium tier.

That matters because a "new flagship" launch usually means a price bump to make the older model look cheap. Anthropic skipped that playbook. You can flip the model ID and your bill should actually drop slightly, because 4.7 uses fewer tokens per task than 4.6 on equivalent work.

For Indian developers paying the API in USD, the maths stays identical: about ₹417 per million input tokens and ₹2,085 per million output tokens at ~₹83 to the dollar. With aggressive prompt caching, input can drop to ~₹42 per million on cache hits.

The SWE-bench jump that actually matters

Headline number: Opus 4.7 scores 64.3% on SWE-bench Pro, up from Opus 4.6's 53.4%. It beats GPT-5.4 (57.7%) and Gemini 3.1 Pro (54.2%) on the same benchmark. On paper, that is a ten-point jump on one of the hardest agentic coding benchmarks in the industry.

But benchmark wins rarely translate one-to-one into real work. The line from Anthropic's own post that should interest developers more is this: Opus 4.7 produces roughly one-third the tool errors of 4.6 on multi-step workflows, and uses fewer tokens to finish the same task. Anthropic is quoting a 14% improvement on complex multi-step workflows while the model spends less time flailing.

If you have ever watched Claude Code retry a failing shell command four times before noticing the typo in its own invocation, this is the change that will actually show up in your day. Fewer tool retries mean shorter sessions, smaller output-token bills, and less babysitting.

Pro tip: Before you migrate, log your current Opus 4.6 sessions for tool-call retries and total output tokens on a typical task. After you switch to 4.7, measure the same thing for a week. The savings are real but they hide inside the usage dashboard, not the invoice.

For a broader view of how the major coding AIs compare head-to-head in real tasks (not just benchmarks), Claude vs ChatGPT for coding in 2026 walks through the practical differences Indian developers care about.

Pricing in rupees: what a real coding session costs

Abstract "per million tokens" numbers hide what Opus 4.7 actually costs you on a normal workday. Here is a concrete example.

A typical Claude Code session with a mid-sized repo: 60,000 tokens of context loaded (files, system prompt, conversation), 8,000 tokens of output (code, explanations, tool calls). At list price that is:

• Input: 60,000 × $5 / 1,000,000 = $0.30 ≈ ₹25
• Output: 8,000 × $25 / 1,000,000 = $0.20 ≈ ₹17
• Total per session: ~₹42

Run ten of these a day, five days a week, and you are at ~₹2,100 a week — roughly ₹8,400 a month. With prompt caching enabled on your system prompt and frequently-reused files, the input cost drops up to 90%, bringing the monthly bill closer to ₹3,500–₹4,500 for moderate use.

Compare that to ChatGPT Plus at $20/month (~₹1,660) for a single-user rate-limited subscription. If you are deciding whether a per-token API is worth it over a fixed subscription, Is ChatGPT Plus worth it in India breaks down the trade-offs. The short version: API usage wins when you are building something, not just chatting.

Migrating from Opus 4.6: what to actually change

The good news is that the minimum migration is a one-line change. In your API client, wherever you have model: "claude-opus-4-6", update it to model: "claude-opus-4-7". Claude Code users on the latest CLI will get it automatically. Bedrock and Vertex AI users need to update the model ID in their deployment config.

The less obvious part: a few behaviours changed, and over-engineered code around 4.6's weaknesses can now actively hurt you. Things to check:

• Tool-use retries: If your app retries failed tool calls three times as a safety net, Opus 4.7 rarely needs more than one. The retry loop is mostly dead code now — keep it, but monitor.
• Instruction tightness: 4.7 follows instructions more literally. Prompts that said "respond casually but be formal about numbers" gave 4.6 room to interpret. 4.7 will pick one and stick to it. Re-read your system prompts for genuine contradictions.
• Vision input: Resolution is noticeably sharper. Screenshots that 4.6 misread (small UI text, charts) are now readable. If you were pre-processing images to upscale, you can probably remove that.
• Cache invalidation: Every cached prompt's key is tied to the model ID. Switching models cold-starts your cache. Plan for one day of 10x input costs before caches re-warm.
• Tool schemas: No breaking changes in the tool-use API shape. Existing function definitions work unchanged.

If you are using Claude Code as your primary agent, the migration is free — just npm update -g @anthropic-ai/claude-code (or your package manager's equivalent) and restart. For a broader comparison of CLI-based coding agents, including how Claude Code stacks up against open-source alternatives, Claude Code vs Goose has the side-by-side.

Where Opus 4.7 still falls short

Worth being honest about the limits. Anthropic itself said Opus 4.7 is not as capable as its unreleased "Mythos" preview on several internal evaluations — so this is not the bleeding edge of what Anthropic has in the lab. It is the best they have publicly shipped.

A few things to know:

• Context window: 1M tokens is generous, but Gemini 3.1 Pro ships 2M. If your workflow involves loading entire large codebases in one shot (not chunked with caching), Gemini still has the edge on raw capacity.
• Cybersecurity content: Opus 4.7 ships with automatic blockers on prompts that look like offensive cyber work. This is aimed at malicious misuse, but legitimate penetration-testing and CTF-style educational use can also trip it. Enterprise tiers have better controls.
• No multimodal output: You can send images and text in, but Claude still only returns text. If you want image generation too, you are still reaching for a separate tool.

For a broader frame on where Claude sits among the major AI players right now, ChatGPT vs Gemini vs Claude vs Grok 2026 compares the four without the marketing filter.

Frequently Asked Questions

How is Claude Opus 4.7 different from Opus 4.6?

Opus 4.7 is mainly a quality-and-efficiency upgrade on the same architecture. It scores ~11 points higher on SWE-bench Pro (64.3% vs 53.4%), produces about one-third the tool-use errors on agentic workflows, and uses fewer tokens to complete the same task. Pricing, context window (1M), and the API shape are all identical. For most developers it is a drop-in replacement.

Does Claude Opus 4.7 cost more than Opus 4.6?

No. Opus 4.7 is priced at $5 per million input tokens and $25 per million output tokens — the same as Opus 4.6. Prompt caching (up to 90% savings) and batch processing (50% savings) both carry over. Because 4.7 uses fewer tokens per task, your actual bill for the same workload should be slightly lower, not higher.

How do I use Claude Opus 4.7 in Claude Code?

Update to the latest Claude Code CLI (npm update -g @anthropic-ai/claude-code or equivalent) and restart it. Opus 4.7 is the default model for new sessions. If you had a specific model pinned in your project settings, change claude-opus-4-6 to claude-opus-4-7. For faster responses on simpler tasks, the /fast command still routes to Opus 4.6, which is not a downgrade — it is intentionally faster.

Should you migrate?

For developers already paying for Opus 4.6 on API, Bedrock, Vertex AI, or Claude Code, the answer is yes — the upgrade is free, the migration is a model-ID change, and 4.7 is measurably better at the specific thing you are using it for (multi-step coding work). The only reason to delay is if you have complex prompt-caching infrastructure; in that case, schedule the switch for a quiet day and expect one cold-cache day of higher input costs.

For developers still comparing AI tools, this launch widens Claude's lead on coding benchmarks without raising the price. If you have been holding out on Claude because of cost concerns, the arithmetic just got friendlier. Start with a week of real work, measure tokens burned, and decide from there — not from the launch post.