What Changed from Claude Opus 4.6 to Claude Opus 4.7?

Conclusion

Claude Opus 4.7 is not a small patch over Claude Opus 4.6. The version number makes it look like a minor step, but in practice it is a meaningful turning point in Anthropic's move from "a capable conversational model" toward "an agentic model that can be trusted with longer work."^[1]

The biggest change is not single-turn answer quality. It is the ability to keep complex work from falling apart midway. Opus 4.6 was already a strong model, but long code edits, multi-file design decisions, recovery from failed tool calls, and work that required reading screenshots or diagrams still often needed close human supervision. Opus 4.7 moves in the direction of reducing that supervision cost.

Anthropic says Opus 4.7 improved the solve rate by 13% over Opus 4.6 on an internal 93-task coding evaluation, and also solved four tasks that Opus 4.6 and Sonnet 4.6 could not solve. Public comparison summaries also show clear gains in coding and agentic work: SWE-bench Pro moving from 53.4% to 64.3%, SWE-bench Verified from 80.8% to 87.6%, and Terminal-Bench 2.0 from 65.4% to 69.4%.

That does not mean it feels better for everyone. Opus 4.7 is more direct, more literal about instructions, more willing to spend reasoning effort, and more likely to consume tokens. People who liked the softer conversational feel of Opus 4.6, its tendency to infer intent from vague prompts, and its comfort for writing or creative brainstorming may find 4.7 harder, heavier, and sometimes less flexible.^[2]

So the practical conclusion is simple. Claude Opus 4.7 is a major improvement over 4.6, but not in the direction of making casual conversation more pleasant. It is an improvement in handing off difficult work. For everyday writing, light advice, or short summaries, the difference may not feel dramatic. For complex code, long-running tasks, image-heavy analysis, and specialist document review, 4.7 is clearly stronger.

What Opus 4.7 Is Trying to Be

Claude Opus 4.7 was released on April 16, 2026. Anthropic positions it as its most capable generally available Claude model. It is available through the API, Claude products, Claude Code, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. Pricing remains the same as Opus 4.6: $5 per million input tokens and $25 per million output tokens.^[1]

The important caveat is that identical list pricing does not necessarily mean identical practical cost. Opus 4.7 uses a new tokenizer, and the same text can become roughly 1.0x to 1.35x as many tokens compared with 4.6. At higher effort settings, the model also tends to think more deeply, which can increase output tokens and reasoning-related spend. To judge the cost of Opus 4.7, the price sheet is not enough. You also have to measure how many retries, human corrections, and failed tool loops it removes.

Anthropic's goal is not simply to release "a smarter chatbot." The release is tied to the broader Claude ecosystem: Claude Code, Claude Agent SDK, Claude Cowork, MCP, tool use, filesystem memory, and long-running agentic work. Opus 4.7 is best understood not only as a standalone model, but as a model tuned for the Claude ecosystem in the agent era.

That direction shows up in early customer comments. Cursor's Michael Truell said Opus 4.7 cleared 70% on CursorBench versus 58% for Opus 4.6. Notion's Sarah Sachs reported a 14% improvement over 4.6 on complex multi-step workflows and one-third as many tool errors. Rakuten's Yusuke Kaji said Opus 4.7 solved three times as many production tasks on Rakuten-SWE-Bench as Opus 4.6. These are not comments about short chat replies. They are comments about how 4.7 behaves inside real workflows.

The Core Upgrade Is Not Coding, but Staying Coherent for Longer

The most visible wins for Opus 4.7 are in coding. But if we describe the change as simply "it writes code better," we miss the main point. The important change is that it is less likely to lose sight of the goal during long work, more able to recover after failure, and more inclined to verify before declaring completion.

Opus 4.6 was already very strong at one-off code generation and small edits. In long sessions, however, it could forget early constraints, drift into a different implementation strategy, repeat the same error, or interpret tool results too shallowly. This is not only a Claude problem. It is a general weakness of agentic AI. A model can appear conversationally fluent, but sustained work requires it to preserve prior decisions, failed attempts, and implicit constraints over time.

Opus 4.7 improves that long-horizon consistency. Vercel's Joe Haddad described a new behavior in which 4.7 reasons about systems code in a proof-like way before starting implementation. CodeRabbit reported more than 10% better recall in code review and better detection of hard-to-find bugs in complex PRs. Warp's Zach Lloyd pointed to a difficult concurrency bug that 4.7 solved even though 4.6 had not.

What emerges is that 4.7 is moving closer not only to an implementer, but also to a reviewer. It examines assumptions before writing code, checks after writing it, and repairs when something fails. This does not remove the need for human review, but it does mean 4.7 shows more of a tendency than 4.6 to carry responsibility through to the end of the task.

That matters for both individual and professional development. If all you need is a small function, 4.6 is often enough. If you want to move through specification, implementation, tests, logs, fixes, and review as one continuous flow, the value of 4.7 rises sharply. Its central improvement is not single-shot brilliance. It is task completion reliability.

The Vision Upgrade Is Bigger Than It Looks

Another major change is support for higher-resolution images. Earlier Claude models handled images up to about 1568 pixels on the long edge, or around 1.15 megapixels. Opus 4.7 raises that to 2576 pixels on the long edge, or around 3.75 megapixels. In simple terms, it can read finer text, tables, diagrams, interfaces, design drawings, and screenshots.^[2]

This is not just about image quality. When AI agents do real work, their inputs are not always clean text. Workflows include admin screenshots, error dialogs, charts, contract scans, figures inside PDFs, UI designs, architectural drawings, chemical structures, and patent drawings. Opus 4.7 can handle details that 4.6 was more likely to miss.

XBOW's Oege de Moor reported that Opus 4.7 scored 98.5% on the company's visual-acuity benchmark, compared with 54.5% for Opus 4.6. Solve Intelligence's Sanj Ahilan described major improvements in understanding chemical structures and complex technical drawings for life-sciences patent work. This shows that vision is no longer a convenience feature. It is becoming an entry point into specialist work.

The cost side matters. A full-resolution image can now use up to roughly 4,784 tokens, compared with the previous cap of about 1,600. If a workflow sends many screenshots, Opus 4.7 can consume far more tokens than 4.6. If the extra fidelity is unnecessary, images should be downsampled or cropped before sending.

That is characteristic of 4.7 as a whole. It can do more, but it also asks for better operational design. It is strong when the image contains dense information. It is not a reason to send everything at maximum resolution by default.

Instruction Following Is Stronger, but Ambiguity Is More Expensive

Opus 4.7 interprets instructions more literally than Opus 4.6. Anthropic's migration guide says this explicitly. With 4.6, users could write a somewhat vague request and the model might infer the intended generalization. With 4.7, especially at lower effort settings, the model is less likely to generalize beyond what was written and more likely to limit itself to the requested scope.^[2]

For business systems, this is a major advantage. In structured extraction, contract review, specification checking, compliance documents, and API pipelines, silent interpolation is dangerous. A model that stays inside the requested scope, preserves ambiguity, and avoids inventing missing information is more trustworthy.

Hex's Caitlin Colgrove praised Opus 4.7 for reporting missing data instead of making up plausible substitutes. That sounds like a small detail, but it is a major practical capability. Many AI failures begin when a model cannot say that it does not know. Opus 4.7 is more cautious than 4.6 in this respect.

For personal use, however, the same change can feel like a drawback. If you want the model to intuit a vague request and "just make it nice," 4.7 can feel less accommodating. Prompts written for 4.6, such as "do the rest in this style," may not produce the same outcome in 4.7. The ability to infer intent has not disappeared, but the model is more tuned to avoid unintended assumptions.

This is an ambiguity tax. To use 4.7 well, you should state scope, output format, priorities, and forbidden changes more clearly than you had to with 4.6. In exchange, well-specified instructions are followed more faithfully. If 4.6 felt like a perceptive conversation partner, 4.7 feels closer to a practitioner who respects the specification.

Changes in Tone and Writing Style

Opus 4.7 is more direct than 4.6 and more willing to state a position. Anthropic's migration guide notes that long-form writing style may shift away from the warmer style of 4.6 toward more direct, verification-oriented expression.^[2]

This change split user reactions. Replit's Michele Catasta said 4.7 pushes back in technical discussion in a way that helps better decisions, making it feel like a better colleague. Pragmatic Engineer's Gergely Orosz, on the other hand, reportedly found it surprisingly combative and returned to Opus 4.6. These reactions are not inconsistent. The same pushback can feel like useful review to one person and difficult attitude to another.^[4]

The change also affects creative writing and article drafting. Opus 4.6 was well suited to warm, supportive prose and broadening a user's intent. Opus 4.7 tends to narrow the issue, challenge vague premises, and push back where needed. If you write with 4.7, it is better to specify style constraints such as "calm explanatory tone," "reader-friendly," or "avoid overstatement."

This is not model degradation. It is closer to a personality shift. Opus 4.7 is designed less as a friendly conversation companion and more as a collaborator that advances work. If you expect the softer Claude of 4.6, the change may feel uncomfortable. In technical review and design judgment, however, the directness can become an advantage.

Token Consumption and the Practical Cost Picture

Opus 4.7 has the same price as 4.6, but same pricing is not the same as same spend.

First, the new tokenizer can split the same input into more tokens than before. Anthropic describes a range of roughly 1.0x to 1.35x. That means the character count can stay the same while the token count rises. Multilingual prompts, structured data, long specifications, and code-heavy documents should be re-measured.

Second, 4.7 thinks more deeply at higher effort settings. In long agentic work, later turns can use more reasoning and produce more output tokens. Harder problems require more compute if you want better answers. That relationship is unavoidable.

Third, higher-resolution images increase token use. A single image can consume thousands of tokens. Workflows that pass dozens of UI screenshots may see usage grow quickly.

Still, looking only at increased token use is incomplete. If 4.7 completes a task with fewer attempts, fewer human corrections, and fewer tool failures, the cost per completed task can fall. Hex reported that low-effort Opus 4.7 can approach medium-effort Opus 4.6 quality. Notion reported better complex workflow performance with fewer tool errors.

So the cost decision is not simple. If you send many light questions, 4.7 can feel excessive and expensive. If you need difficult work completed, it may be cheaper than repeated retries with 4.6. Opus 4.7 is not a model to choose by unit price alone. It is a model to choose by completion rate.

Effort, Task Budget, and xhigh

Effort matters more in Opus 4.7. Effort controls how deeply the model thinks. Opus 4.7 adds a new xhigh level between high and max, giving finer control over reasoning and latency. In Claude Code, the default effort for Opus 4.7 has been raised to xhigh.^[1]

This change captures the model's nature. Simple questions do not need maximum reasoning. Complex design, long code edits, ambiguous log analysis, and multi-document synthesis often do. At low effort, 4.7 stays tightly scoped. At higher effort, it reasons more and tends to use tools more where appropriate.

Task budget is also important. It gives the model a total budget for long-running agentic work, including thinking, tool calls, tool results, and final output. In human terms, it is like saying, "You may spend this much time and attention on this job." The model can then decide where to allocate effort.

This is another sign that AI is moving from question-answering to task execution. In a simple chat model, you send a prompt and receive an answer. In an agentic workflow, it matters how far the model should investigate, when it should stop, and what it should prioritize. Opus 4.7 brings those work-management concepts closer to the surface.

What Happened in the Claude Code Quality Backlash

Any discussion of Opus 4.7 has to address the backlash around its release. Business Insider reported that Reddit and X users complained that Opus 4.7 felt like a regression from 4.6. One Reddit post reportedly received more than 2,300 upvotes, and an X post saying 4.7 was not a real improvement over 4.6 reportedly received about 14,000 likes.^[4]

The complaints varied: simple spelling mistakes, fabricated school names or surnames while rewriting resumes, slower responses, a sense that the model was not thinking, a combative tone, and excessive token use. Some users, including Gergely Orosz, returned to 4.6. Others, including Jeremy Howard, reportedly described 4.7 as the first model that understood what they were trying to do while working. The reaction was sharply divided.

Anthropic then published a Claude Code quality postmortem on April 23. It explained three product-layer issues behind the perceived degradation.^[3]

The first was a March 4 change that lowered Claude Code's default reasoning effort from high to medium. The intent was to reduce latency and avoid usage-limit pressure, but users experienced it as a drop in intelligence. Anthropic said that tradeoff was wrong and reverted the default on April 7.

The second was a March 26 caching optimization bug. The intended design was to clear old thinking once when resuming a session that had been idle for more than an hour. Because of a bug, older thinking continued to be dropped on each subsequent turn. Claude then lost access to why it had made previous edits or tool calls, which made it appear forgetful, repetitive, and odd in tool selection.

The third was an April 16 system prompt instruction to reduce verbosity. It limited text between tool calls to 25 words or fewer and final responses to 100 words or fewer unless the task required more detail. Anthropic found that this reduced coding performance by about 3% on a broader evaluation set and reverted it on April 20.

The key point is that Anthropic said the API and inference layer were not affected. The perceived "Claude got worse" story was not just about the base model. It involved effort settings, caching, system prompting, and Claude Code product design. In the agent era, AI quality is determined by the whole surrounding system, not the model alone.

Safety Changed Too

Opus 4.7 also strengthens cybersecurity safeguards. Anthropic introduced mechanisms to detect and block prohibited or high-risk cyber use, while inviting legitimate security professionals doing vulnerability research, penetration testing, or red teaming to join the Cyber Verification Program.^[1]

That shows that Opus 4.7 is not only more capable, but also capable in ways that can be risky. Stronger code understanding, tool use, long-horizon autonomy, and screenshot comprehension can support defensive work and offensive misuse. Anthropic's additional safeguards for 4.7 also serve as a deployment test bed for broader future Mythos-class model releases.

Safety evaluations describe Opus 4.7 as having a broadly similar profile to Opus 4.6, with improvements in honesty and prompt-injection resistance, but some weaker areas in specific harmful domains. Once again, 4.7 is not an all-direction upgrade. The stronger a model becomes, the more complex safety tuning becomes.

In practice, this matters. Legitimate security work may face more refusals than before. For ordinary business use, stronger safeguards can be a benefit. Enterprises should test not only what 4.7 can do, but also which requests it blocks.

Who Should Move from 4.6 to 4.7?

The first group that should consider moving is people who hand complex code work to Claude: multi-file changes, test-failure analysis, log or trace investigation, review, and design decisions. These are the places where 4.7's improvements are most visible.

The second group is people using Claude Code or agentic workflows. The comments from Notion, Rakuten, CodeRabbit, Cursor, Vercel, Warp, and Replit all point to the same pattern: 4.7 is most valuable when it works with tools over time. The advantage is less about short replies and more about persistence after errors.

People who work with images, diagrams, screenshots, interfaces, technical drawings, patent documents, or complex PDFs also benefit. Higher-resolution image support can surface details that 4.6 missed. This is especially important for software UI analysis, document review, data extraction, life sciences, and legal or patent work.

Specialist document analysis is another strong fit. Harvey reported high substantive accuracy on BigLaw Bench and better handling of ambiguous editing and provision identification. In finance, law, contracts, and research materials, the tendency to report uncertainty instead of inventing missing information is valuable.

There are also people who do not need to rush. If your main work is casual chat, short summaries, translation, warm prose, creative brainstorming, or stable 4.6 prompt workflows, the improvement may be modest. The more direct tone and increased token use may even be a drawback.

In short, 4.7 is not a universal replacement. Heavy work belongs on 4.7. Light conversation and style-sensitive writing may still be better on 4.6. As of spring 2026, that split is reasonable.

Why Some People Feel There Is Not Much Difference

Some users will try Opus 4.7 and feel little difference. That is not surprising. Many of 4.7's improvements appear in areas that short conversations do not reveal.

If you ask one question, 4.6 is already strong. If you generate a short code snippet, the gap may be small. If you polish a paragraph, 4.6 can even feel softer and more pleasant. The strength of 4.7 appears in 20-minute, one-hour, or multi-hour tasks; in work with repeated failures; and in workflows crossing images, tools, files, and logs.

4.7 is also stricter with vague prompts. Requests that 4.6 would smooth over may be handled literally by 4.7. If you change only the model and keep the same prompt, it can feel less intuitive.

Token consumption also affects perception. Hitting usage limits sooner, waiting longer for reasoning, receiving heavier output, and paying more tokens for images can all make the user notice friction before noticing the capability gain.

Therefore, Opus 4.7 should be evaluated by completion behavior, not just answer quality. How much long work did it finish? How much human intervention did it remove? How often did it recover from tool errors? How useful was its review? Opus 4.7 is a long-distance model, not a short-sprint model.

Practical Ways to Use Opus 4.7 Well

If you use Opus 4.7, do not keep prompts exactly as they were for 4.6. Reduce ambiguity. State the goal, constraints, priorities, output format, and completion conditions clearly. For coding, say what may be changed, what must not be changed, which tests should run, and how uncertainty should be handled.

Match effort to the work. Light questions do not need xhigh. Design review, bug investigation, multi-file refactors, and long research can justify high or xhigh. Because 4.7 respects effort levels more strictly, difficult work at low effort may underuse the model's strengths.

Use images deliberately. High-resolution images are powerful but expensive. Decide whether the whole UI needs to be seen or whether a crop is enough. This is not just micro-optimization; it is part of keeping long work stable.

Finally, do not treat 4.7 as something that will infer everything. It is strong when clearly delegated to. Vague handoff can make it behave more literally than intended. As with assigning work to a strong engineer, the deliverable criteria should be explicit.

A Note on GPT-5.5

The main topic here is the change from Opus 4.6 to Opus 4.7. Still, GPT-5.5 appeared in the same period, so readers naturally want to know how the two compare.

At a high level, Claude Opus 4.7 is strong in real repository edits, code review, specification understanding, vision, and careful specialist document analysis. GPT-5.5 is strong in terminal work, web research, long context, data analysis, and execution across multiple tools.

OpenAI's official comparison shows GPT-5.5 at 82.7% on Terminal-Bench 2.0 versus Claude Opus 4.7 at 69.4%, a large advantage for command-line work. On SWE-Bench Pro, Claude Opus 4.7 scores 64.3% versus GPT-5.5 at 58.6%, giving Claude the edge on an issue-resolution-style benchmark. On BrowseComp, GPT-5.5 scores 84.4% versus Claude Opus 4.7 at 79.3%, suggesting a GPT advantage in web research. On MCP Atlas, Claude Opus 4.7 scores 79.1% versus GPT-5.5 at 75.3%, giving Claude the advantage on one class of tool-connected workflows.^[5]

The relationship is therefore not a simple ranking. GPT-5.5 is strong as an engine for moving work forward. Claude Opus 4.7 is strong at understanding the meaning of complex work and improving design or review quality.

The operational caution is model switching. Designing with Claude, implementing with GPT, reviewing with Claude, and then fixing with GPT can sound powerful in theory. In practice, context fragments. What was tried, why a design was chosen, which failures were avoided, and which constraints mattered can get lost between models.

For production work, it is often more stable to choose one main model and let it carry the context to completion. If you start with Claude Opus 4.7, let Claude hold the design, implementation, and review context. If you start with GPT-5.5, let GPT carry the research, implementation, and verification context. Use another model at boundaries, with a clear packet of diffs, requirements, logs, and test results.

GPT-5.5 is strong. Claude Opus 4.7 is strong. In the agent era, the critical choice is not only which model is smarter, but which model you trust with the working context.

Final Judgment

Claude Opus 4.7 is a clear improvement over Opus 4.6. The biggest gains are in complex coding, long-running agentic work, tool use, screenshot and diagram understanding, and careful specialist document analysis.

But 4.7 is not a complete superset of 4.6. For warm conversation, creative brainstorming, light writing, and stable existing prompts, 4.6 may still feel preferable. Opus 4.7 is stricter, more direct, heavier, and more work-oriented.

In one sentence: Opus 4.6 was the Claude that seemed to understand you well. Opus 4.7 is the Claude that tries to carry the work to the finish line.

If you understand that difference, 4.7 is extremely powerful. If you treat it as a simple continuation of 4.6, it can feel hard, heavy, less intuitive, and token-hungry. That is why global reactions to Claude Opus 4.7 split so sharply.

Claude Opus 4.7 is not just a performance upgrade. It is a step from conversation toward delegated work.

References

1. [1]Introducing Claude Opus 4.7 (Anthropic) ↩
2. [2]Migration guide - Claude API Docs (Anthropic) ↩
3. [3]An update on recent Claude Code quality reports (Anthropic) ↩
4. [4]Claudeの最新モデル「Opus 4.7」、トークンを消費しすぎると批判を集める (Business Insider Japan) ↩
5. [5]Introducing GPT-5.5 (OpenAI) ↩