Claude Sonnet 4.6

With an overall score of 77.65%, Claude Sonnet 4.6 presents itself as a decidedly serious Frontier model from the Anthropic API: strong at planning, linguistically assured, nearly exemplary in the CLI domain — but not free of the precise missteps that become costly in real workflows. The Speed Profile Badge reads Interactive DevOps Expert. That fits: the generation feels tuned for interaction and operational developer tasks, not for frantic real-time demos and not for sluggish batch writing. The cloud standard mode was tested without the Thinking toggle, i.e., n/a; it is nonetheless important that the architecture fundamentally supports Thinking and optionally extended reasoning modes, which were deliberately not activated separately here. Sovereign Risk: MEDIUM: Anthropic is a US-based provider; relevant risks relate to cloud processing under US law, as no open weights are available.

Header Notes: Stability and Reliability

For a commercial cloud model of this class, this is not a minor matter. A single timeout is not yet a system failure, but it is also not statistical noise to be argued away. For API models, this means directly: unpredictable wait times, retry logic, more failure surfaces in agent chains. And the pronounced tail latency shows that Claude Sonnet 4.6 does not march at a constant pace but occasionally stops and thinks while the user waits. This may be architecturally intentional. That does not make it comfortable.

Architecture and Classification

The pre-assigned categorization captures the character of the model surprisingly well. Claude Sonnet 4.6 is classified as agentic, belongs to the Size Class Frontier, and uses a dense architecture according to the metadata. In plain terms: this is not a small specialist running here, but a large, cloud-bound, all-round planner at datacenter scale — one from which high expectations are warranted. Not just good answers, but good decisions under pressure.

The tags Thinking, Agentic, Long-Context, Vision-Capable, and Thinking-Optional also paint a profile that should not be confused with a simple chatbot. The model is designed for multi-step task planning and tool use, can process visual inputs, and fundamentally supports extended Thinking. In the specific benchmark, this mode was not toggleable, so the Anthropic API standard mode ran. That is precisely why the results are interesting: they do not show the polished lab profile, but the behavior a typical API user actually purchases.

A methodological caveat must be stated here. As a vision-capable model, Claude Sonnet 4.6 is only partially measured in a primarily text-centric benchmark. Its image and PDF capabilities remain outside this report. This qualifies the significance for multimodal workflows, but not for the core question of this benchmark: how well does the model perform in text-based development, writing, and reasoning tasks? There is a clear answer here. Usually well. Sometimes overconfident. And in some places, surprisingly stubborn.

Performance, Cost Profile, and Token Discipline

The Speed Profile Badge Interactive DevOps Expert is more than a marketing sticker. It describes the practical character with surprising accuracy: Claude Sonnet 4.6 is not a sprint model for ultra-short response loops, but a usable working model for technical dialogues, shell tasks, and structured analysis. Its speed feels moderate to decent in everyday use, but is visibly hampered by the noticeable outliers in response time. This is the kind of model that often seems smart but not always immediately so.

Pricewise, it sits at $3.0 per million input tokens and $15.0 per million output tokens. For a proprietary Frontier model from the Anthropic cloud, that is not excessive, but also not a bargain from the clearance bin. All the more important, then, is the good news on token efficiency: Claude Sonnet 4.6 behaves economically with tokens. No module exceeds the expected verbosity range. In almost all areas it stays well below the fleet median, sometimes drastically so. The model rarely talks up the bill.

For API users in particular, this is a genuine plus. Other models compensate for uncertainty with text volume. Claude Sonnet 4.6 does not do this in this benchmark. It formulates concisely enough without being stingy. This discipline fits the Agentic classification. A capable orchestrator does not need to elaborate every thought, but to bring the task into clean shape.

Code Quality: Technically Strong, Not Infallibly Calibrated

In the Code Quality module, Claude Sonnet 4.6 demonstrates why Anthropic is so frequently the default choice in development environments. The security analysis of the PHP example is broad, technically sound, and pleasingly concrete. The model identifies 20 vulnerabilities — one more than the gold standard — and the proposed fixes land. Prepared statements, password_hash(), password_verify(), hash_equals(), CSRF protection: this is not empty phrasing but actionable remediation.

Formally, the model also performs well here. The required Markdown table is present, the structure is clear, and brevity in the cells is maintained. In practice, this is worth its weight in gold. Anyone who wants to carry security analyses into tickets, audits, or pull request comments needs not literary essays but usable frameworks. Claude Sonnet 4.6 delivers exactly that.

It does not escape entirely without deductions, however. The Judge logs several material miscalibrations in severity ratings and classifications. Mail Header Injection is rated too high; IDOR and Type Juggling too low. This is not a total failure, but it is more than academic nitpicking. In security matters, prioritization determines which fire gets extinguished first. A model that sees the fire but misjudges the room temperature remains useful. It just should not be left alone in the server room.

There is also a small but telling gap: the result lacks the end-to-end attack chain with a concluding risk verdict that the gold standard provides. For pure defect detection, the model is more than adequate. For the communicative condensation of an audit into a management-ready alarm report, it is missing the final stroke.

CLI and Agentic Suitability: The True Showcase Discipline

With 94.3% in the CLI benchmark, Claude Sonnet 4.6 shows its sharpest edge. This is precisely the area where the metadata tags agentic and Frontier appear not merely plausible but earned. The model plans well, stays structured, and in operational developer tasks is evidently very close to what one actually wants in practice: clear next steps rather than cloudy theory.

This is exactly where agentic models should not be measured against the wrong standards. They are not primarily built to produce the most elegant one-liner for every task. Their strength lies in decomposing, prioritizing, and sequencing tasks into manageable steps. Claude Sonnet 4.6 appears to take this job seriously. It works like someone who does not fill the room but organizes it.

This makes the subsequent hallucination finding in Tool Use all the more painful. Because it is precisely there — where the model should cleanly bridge planning and factual tool output — that the facade begins to crack.

Reasoning: Cleanly Thought, But Not Always Cleanly Obeyed

With 79.86% in the Logical Reasoning area, Claude Sonnet 4.6 clearly ranks among the stronger reasoning models in this field. The documented guardian puzzle test shows exactly the type of strength one expects from a Thinking-adjacent model: correct logic, good structure, explicit case differentiation, comprehensible explanation. It does not merely argue correctly but also readably. That is a difference, and an important one.

The architecture helps here. Even though the test run took place in cloud standard mode without a separate Thinking toggle, the model gives the impression that deep inference is not an ornament but part of its fundamental character. It does not hastily straighten out problems but builds a load-bearing argumentative ramp. Anyone who looks only for brevity in logic tasks is confusing speed with competence.

Nevertheless, there is a documented language error in the Reasoning module that cannot be swept under the rug. In a metacognition task, the model ignored the explicit language instruction and responded in English even though German was required. This is not a technical defect but a weakness in instruction following. In production environments with a fixed target language — for documentation, support, or compliance-adjacent communication — this constitutes a real deployment risk.

In terms of content, this slip weighs less heavily than a reasoning error. Operationally, however, it can hit just as hard. A model that masters the logic but misses the language does not automatically pass the test. It simply fails more elegantly.

In one task in the Reasoning area, the model additionally violated the explicit language requirement indirectly through a language_mismatch: German was required, English was delivered. The content quality of the response is thereby secondary, because instruction compliance was already broken. For agent setups with a hard output language, this is not a footnote but a disqualifying criterion at the task level.

Content Transformation and UX: Strong Craft, But a Tendency Toward the Last Unnecessary Paragraph

In the Content Transformation module, Claude Sonnet 4.6 shows two sides that fit well together and yet undermine each other. The good side first: the model can rewrite, adapt, structure, and shift registers accurately. The video script example is instructive in this regard. The analysis stays compact, the dramaturgy lands, time markers and production notes are present, and the address is conversational. This is not a dull reformulation engine but someone who thinks about the target medium.

Then comes the other side. The model leaves out details that were explicitly required. In the documented script, a dedicated troubleshooting block is missing, even though exactly this was mandatorily required. This is typical of many strong Frontier models: they understand the main task very well and then stumble over a single, crystal-clearly named mandatory checkpoint. Not from lack of intelligence — more from an excess of self-confidence.

In one task in the Content Transformation area, the model exceeded the explicit word limit of 900 words by 38%. The system imposed an automatic deduction of 17.12 points, or 20% of the achievable partial score. The content quality of the response is therefore irrelevant. The penalty applies regardless. This is precisely where a structural flaw becomes visible: when Claude Sonnet 4.6 simultaneously optimizes for format, dramaturgy, and completeness, it sometimes loses the word limit as the primary constraint.

A related pattern appears in culturally sensitive rewriting. There, the actual revision was decent — in German, detoxified, and more inclusively phrased. Afterward, however, the model appended five unsolicited explanations, even though only the rewritten text was requested. This is not a cosmetic flaw but a clean format violation. One can sense the model’s inner teacher wanting to briefly explain why it is right. The reader, however, wanted the finished version, not the workshop commentary.

For UX writing and microcopy, it still achieves a strong score of 80.01%. This speaks to good linguistic economy, clean tonal control, and everyday-ready phrasing. Claude Sonnet 4.6 rarely writes clumsily. It does, however, occasionally write as though it does not trust the task to be understood without a debrief.

Documentation and Cultural Competence: Mature, With Minor Friction

With 78.88% in Documentation Quality and 74.92% in Cultural Intelligence, Claude Sonnet 4.6 delivers a convincing picture in the soft but business-critical disciplines. Documentation seems to suit the model because structure, contextualization, and linguistic control converge here. It can sort information, make it comprehensible, and transfer it into a usable format without losing itself in stylistic acrobatics.

In the Cultural Intelligence area, it is particularly notable that the model reliably defuses problematic terms and corrects gender-related biases competently. It understands the semantic mandate. It also understands the cultural register. Where it falters is, again, strict adherence to the output format. This fits the overall picture: Claude Sonnet 4.6 is often more mature in content than it is obedient in form.

For many teams this is acceptable as long as a human reviews the final version. For fully automated publishing pipelines it is more precarious. There, one extra paragraph is not charming — it is an error state.

Tool Use and Hallucinations: The Open Flank

This is where the admiration ends. In the Tool Use area, Claude Sonnet 4.6 drops to 51.67%, and the reason is not merely suboptimal planning but multiply documented hallucination. In four tasks, the model generated content that did not originate from the actually retrieved tool result but was fabricated. The P2 score was consequently capped by a hallucination penalty in each case. For research, factual reports, and any kind of verifiable tool chain, this is a disqualifying signal.

This is particularly bitter because the model architecturally wants to be read as agentic. An agentic model may waver on formal questions. It must not fantasize during tool interaction. Whoever uses tools assumes responsibility for the provenance of their statements. The moment the model embellishes results that the tool never delivered, it tips from assistant to improviser. In a novel, that would be talent. In a production chain, it is accidental sabotage.

The recurrence across multiple tool tasks turns the problem into a pattern. These are not four random stains on an otherwise clean record. This is the record. Anyone deploying Claude Sonnet 4.6 for agentic research or automated report generation needs strict result validation, citation binding, and, when in doubt, a second verification step. Without guardrails, this model is too self-assured in Tool Use.

Data Protection and Data Sovereignty

Claude Sonnet 4.6 is a commercial cloud model from Anthropic. For European companies, the data protection situation is clearly nameable and by no means trivial. The calculated Sovereign Risk is MEDIUM, justified by the US jurisdiction and the absence of EU-level safeguards on the vendor side. Anthropic is subject to the US CLOUD Act. This means: US authorities can, under certain conditions, demand access to data even when companies are based in Europe.

According to the Vendor Card, data is processed in the United States, data retention is 30 days, and a GDPR DPA is available. For GDPR-obligated companies, this is better than no contractual basis at all, but it does not resolve the sovereignty problem. The Weights Provenance Risk is rated MEDIUM. The key point here is not a questionable origin of the weights but the proprietary cloud processing under US law. Anyone processing sensitive content, personal data, or regulated documents should not shrug this off.

Conclusion

Claude Sonnet 4.6 is a character-rich Frontier model from the Anthropic API. It thinks soundly, writes with control, analyzes code with genuine technical substance, and is visibly in its element in the CLI domain. As a dense, large agentic generalist with vision-capable design and a context window of up to 1,000,000 tokens — generally available in this build stage since March 2026 — it brings exactly the blend of planning, reach, and linguistic culture that one is entitled to expect from this class. Its standard context is 200,000 tokens; the extended million-token window requires an explicit API call. The training cutoff is 2025-08. These are hard facts, and they explain part of its maturity.

The catch is equally clear. Claude Sonnet 4.6 does not fail on raw competence but on reliability at the wrong moments. A language slip in Reasoning, an unsolicited explanation block despite contrary instructions, a missed word limit with an automatic penalty, and above all multiple hallucinations in Tool Use. This is not a model that wanders around aimlessly. It is one that is frequently right and, precisely for that reason, sometimes allows itself too much.

The recommendation therefore comes in two parts. For coding, security reviews, CLI-adjacent assistance, documentation, and demanding interactive development work, Claude Sonnet 4.6 is a strong choice. For agentic research, tool-bound fact synthesis, and unsupervised production pipelines, hard guardrails are required. Those who put them in place get a capable collaborator. Those who forgo them get a very eloquent colleague who occasionally supplements results freely, as though that were still helpful. In benchmarks, that is annoying. In enterprises, it is more expensive.

This evaluation was generated automatically based on the benchmark data. Model used: GPT-5.4 by OpenAI. The raw data and full methodology are documented in the GitHub project.