Claude Opus 4.7: What Changed, What It Costs, and Whether You Should Switch

Quick Answer: Claude Opus 4.7, released April 16, 2026, scores 87.6% on SWE-bench Verified and 70% on CursorBench — both significant jumps over Opus 4.6. Pricing stays at $5/$25 per million input/output tokens, but a new tokenizer can produce up to 35% more tokens for the same text. It’s the best coding model you can use today, with a catch on effective cost.

Last updated: April 2026

What Opus 4.7 actually is

Opus 4.7 is Anthropic’s latest flagship model, replacing Opus 4.6 as the top-tier option across Claude.ai, the Anthropic API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. Model ID: claude-opus-4-7.

This is not a minor point release. Anthropic describes it as a notable improvement on Opus 4.6 in advanced software engineering, with particular gains on the most difficult tasks. Users report being able to hand off their hardest coding work — the kind that previously required constant supervision — and get usable results back.

The release came nine days after Anthropic disclosed Claude Mythos Preview, a more powerful model that remains gated and unavailable to the public. Opus 4.7 is positioned as the best generally available model, not the best model Anthropic has.

Benchmark numbers that matter

Here’s where Opus 4.7 lands against its predecessor and competitors:

Across 14 benchmarks reported by Anthropic, Opus 4.7 beats Opus 4.6 on 12. It loses on Terminal-Bench and BrowseComp.

The SWE-bench Pro result is particularly telling: 64.3% versus GPT-5.4’s 57.7% — a 6.6-point lead on the hardest coding evaluation available. On the Artificial Analysis Intelligence Index, Opus 4.7 scores 57, up from 53 for Opus 4.6.

The four changes that matter for daily work

1. Coding got meaningfully better

The 13% lift on coding benchmarks is not just a number. The practical difference shows up in agentic coding workflows: Anthropic claims 3x more production tasks resolved compared to Opus 4.6. Claude Code users report that multi-file refactors, complex debugging sessions, and architecture-level changes now complete with fewer intervention cycles.

2. Vision jumped from unusable to competitive

This is the first Claude model with high-resolution image support. Maximum image dimension went from 1,568px (about 1.15 megapixels) to 2,576px (about 3.75 megapixels) — a 3x+ pixel count increase. Document reasoning errors dropped 21%.

The XBOW visual acuity score went from 54.5% to 98.5%. That’s not an incremental improvement; it’s a capability that didn’t meaningfully exist before.

3. New effort level and task budgets

Opus 4.7 introduces an “xhigh” effort level between “high” and “max,” which is now the default for Claude Code. Anthropic recommends it for most coding and agentic use cases.

Task budgets are a new mechanism for controlling agentic loops — useful for teams running autonomous agents that need guardrails on how long they can run before checking in.

4. Adaptive thinking replaces manual budgets

Extended thinking budgets are gone. The model now uses adaptive thinking, automatically determining reasoning depth based on task complexity. This simplifies the API but removes a control lever that some power users relied on.

Pricing: same rate card, potentially bigger bill

On paper, nothing changed:

• Input: $5 per million tokens
• Output: $25 per million tokens
• Context window: 1 million tokens
• Max output: 128K tokens

This makes Opus 4.7 the first Opus upgrade that did not raise the price.

But there’s a catch. Opus 4.7 ships with a new tokenizer that can produce up to 35% more tokens for the same text, especially on code-heavy prompts. If your workload is code-intensive, your effective cost may increase by 10-35% even though the per-token price is unchanged.

For teams running high-volume agentic workflows, this is worth measuring before committing to a migration. Run your actual prompts through both tokenizers and compare token counts.

Breaking API changes

Three changes that will break existing integrations if you’re not prepared:

1. Extended thinking budgets removed — replaced by adaptive thinking
2. temperature, top_p, top_k parameters silently ignored — they’re accepted but have no effect
3. Custom sampling parameters removed

If your application relies on temperature tuning or explicit thinking budgets, test thoroughly before switching.

Where Opus 4.7 wins and loses

Wins:

• Coding benchmarks (SWE-bench, CursorBench) — clear leader
• Tool use (9.2-point lead on MCP-Atlas over GPT-5.4)
• Vision tasks (document analysis, screenshot understanding)
• Instruction following (stricter adherence)

Loses:

• Terminal-Bench (command-line workflow benchmarks)
• BrowseComp (web browsing tasks)
• Long-context multi-needle retrieval — actually regressed from Opus 4.6
• Effective cost on code-heavy workloads (tokenizer inflation)

Who should upgrade

Upgrade now if:

• You build or maintain agentic coding systems
• You process documents or screenshots as part of your workflow
• You’re on Opus 4.6 and want better instruction following

Wait if:

• Your workload is cost-sensitive and code-heavy (measure tokenizer impact first)
• You rely on temperature/sampling parameter tuning
• Your primary use case is long-context retrieval (regression risk)

The Mythos elephant in the room

Anthropic disclosed Claude Mythos Preview on April 7, nine days before Opus 4.7. Mythos outperforms Opus 4.7 on every benchmark. It discovered thousands of high-severity zero-day vulnerabilities. And it’s not available to anyone outside Anthropic.

Some frame Opus 4.7 as a “concession prize” — the best model Anthropic is willing to release, not the best model they have. Whether that matters depends on your perspective: Opus 4.7 is still the strongest coding model you can actually use today.

Sources: Anthropic official blog, Artificial Analysis, The Next Web, nxcode.io, Forbes, AWS Bedrock model card