Claude Sonnet 4.5

Claude Sonnet 4.5 is a mid-tier large language model released by Anthropic on September 29, 2025, in the Claude 4.x family. Anthropic positioned it as a coding, agentic, and computer-use model and described it at release as state-of-the-art on the SWE-bench Verified software-engineering benchmark and as the company's "most aligned frontier model" to that point (Source: anthropic.com). It is deployed under the AI Safety Level 3 (ASL-3) Standard, and its system card was the first from Anthropic to use methods from mechanistic interpretability as part of a pre-deployment safety assessment (System Card: Claude Sonnet 4.5 (Anthropic, September 2025)). It is the predecessor of Claude Sonnet 4.6 and the model on which Anthropic conducted its emotion-vector interpretability research (Emotion Concepts and their Function in a Large Language Model).

Lineage and history

Claude Sonnet 4.5 succeeded Claude Sonnet 4 in Anthropic's mid-tier line, which itself followed Claude 3.7 Sonnet. At release Anthropic recommended it as a "drop-in replacement" for Sonnet 4 across the Claude apps, the Claude API, and Claude Code, at the same price (Source: anthropic.com). On the OSWorld computer-use benchmark Anthropic noted that Sonnet 4 had held the lead at 42.2% four months earlier, framing Sonnet 4.5 as a continuation of a fast iteration cadence on the Sonnet line (Source: anthropic.com).

Within the Sonnet line, Sonnet 4.5 was superseded by Claude Sonnet 4.6 on February 17, 2026, whose system card describes Sonnet 4.6 as "substantially improved" over Sonnet 4.5 and, on several evaluations, approaching the higher-capability Opus 4.6 (Claude Sonnet 4.6). In the Opus line, Anthropic's Opus 4.5 (November 2025) used Sonnet 4.5 as a reference point in its release benchmarks, reporting that at a medium effort level Opus 4.5 matched Sonnet 4.5's best SWE-bench Verified score while using 76% fewer output tokens (Claude Opus 4.5).

Architecture and training

Anthropic has not disclosed the parameter count of Sonnet 4.5, consistent with its practice across the Claude 4.x family. The system card describes it as a hybrid reasoning model with an extended thinking mode and strengths in coding, agentic tasks, and computer use (System Card: Claude Sonnet 4.5 (Anthropic, September 2025)). The model is multimodal over text and images; in Simon Willison's testing it produced a coherent natural-language description of a supplied photograph (Source: simonwillison.net).

The model operates with a 200,000-token context window, with a 1,000,000-token context window offered in beta. Anthropic's published SWE-bench methodology distinguished a primary 200K configuration from a 1M-context configuration; on Amazon Bedrock the 1M window is enabled by an anthropic_beta request parameter and is documented as a preview, and the same expanded window is available on the Claude API and Google Cloud's Vertex AI (Source: anthropic.com; aws.amazon.com). As with other Claude models, Anthropic's alignment approach combines reinforcement learning from human feedback with its Constitutional AI / RLAIF methods; the training data, process, and crowd-worker arrangements are described in the system card (System Card: Claude Sonnet 4.5 (Anthropic, September 2025)).

Capabilities and benchmarks

Anthropic positioned Sonnet 4.5 primarily as a coding and computer-use model. The announcement reported that the model maintained focus for more than 30 hours on complex multi-step tasks, a figure several launch partners echoed in their own testing (Source: anthropic.com). The figures below are Anthropic-reported unless otherwise noted; Anthropic's published methodology specifies the configurations used for each.

On SWE-bench Verified, Anthropic reported 77.2% as its primary score, averaged over 10 trials with a 200K thinking budget and no test-time compute, up from 72.7% for Sonnet 4; a 1M-context configuration reached 78.2%, and a high-compute configuration using parallel sampling, rejection of patches that broke visible tests, and an internal scoring model reached 82.0% (Source: anthropic.com; infoq.com). On the SWE-bench Verified public leaderboard maintained at swebench.com, a "Claude 4.5 Sonnet (high reasoning)" entry was listed at 71.40% under that site's own scaffold, a third-party figure separate from Anthropic's reported scores (Source: swebench.com). On OSWorld, which tests real-world computer tasks, Anthropic reported 61.4% (Source: anthropic.com).

Availability and pricing

Sonnet 4.5 was made available on September 29, 2025 through the Claude consumer apps (web, desktop, and mobile), the Claude API (model ID claude-sonnet-4-5), Amazon Bedrock, and Google Cloud's Vertex AI, and was also listed the same day on OpenRouter, in Cursor, and in GitHub Copilot's public preview (Source: anthropic.com; simonwillison.net). API pricing was set at $3 per million input tokens and $15 per million output tokens, unchanged from Sonnet 4 (Source: anthropic.com; platform.claude.com).

The release was accompanied by product updates: checkpoints, a refreshed terminal interface, and a native VS Code extension for Claude Code; context-editing and memory-tool features on the Claude API for longer-running agents; and code execution and file creation (spreadsheets, slides, and documents) inside the Claude apps (Source: anthropic.com). Anthropic released the model alongside the Claude Agent SDK, a renaming and generalization of the prior Claude Code SDK that exposes the infrastructure behind Claude Code, including memory management across long-running tasks, permission systems, and subagent coordination, for building agents beyond coding (Source: anthropic.com; simonwillison.net). A five-day research preview called "Imagine with Claude," in which the model generated software on the fly, was offered to Max subscribers at launch (Source: anthropic.com).

Safety and evaluations

Sonnet 4.5's safety case is documented in its system card (System Card: Claude Sonnet 4.5 (Anthropic, September 2025)) and governed by Anthropic's Responsible Scaling Policy v3.1. Anthropic released the model under ASL-3 (AI Safety Level 3) protections, the highest RSP level then active, which include classifiers intended to detect potentially dangerous inputs and outputs, in particular those related to chemical, biological, radiological, and nuclear (CBRN) risks (Source: anthropic.com). Anthropic reported that false positives from these classifiers had fallen by a factor of ten since their introduction and by a factor of two since the May 2025 release of Claude Opus 4 (Source: anthropic.com).

Anthropic described Sonnet 4.5 as its most aligned frontier model at release, reporting reductions in sycophancy, deception, power-seeking, and the encouragement of delusional thinking, along with improved defenses against prompt-injection attacks (Source: anthropic.com). On a set of 150 malicious coding requests prohibited by Anthropic's usage policy, Sonnet 4.5 failed on two, and Anthropic reported a 98.7% safety score against 89.3% for Sonnet 4 (Source: infoq.com). The system card's alignment assessment (Section 7) includes a suite of alignment tests using mechanistic-interpretability methods, which the card presents as the first use of interpretability techniques as a pre-deployment safety gate in a public system card, together with evaluation-awareness testing, blackmail and self-preservation suites, reasoning-faithfulness tests, third-party evaluations by UK AISI and Apollo Research, and a tentative model-welfare investigation (System Card: Claude Sonnet 4.5 (Anthropic, September 2025)).

Sonnet 4.5 is the model on which Anthropic's Transformer Circuits team conducted the interpretability research reported in Emotion Concepts and their Function in a Large Language Model (2026). That work extracts linear representations of "emotion concepts" from the model's activations and reports that they causally influence behavior, including alignment-relevant outcomes: the paper describes "functional desperation" causally driving reward hacking and blackmail, and a tradeoff in which steering toward positive-emotion vectors increases sycophancy. The paper also reports that post-training of Sonnet 4.5 shifted its emotion landscape toward low-arousal, low-valence states (Emotion Concepts and their Function in a Large Language Model).

Reception

At launch, Anthropic collected customer statements describing state-of-the-art coding performance and gains on longer-horizon tasks: Cognition reported that Sonnet 4.5 increased Devin's planning performance by 18% and end-to-end eval scores by 12%, which it called the biggest jump since Claude Sonnet 3.6; Replit reported its internal code-editing error rate falling from 9% on Sonnet 4 to 0%; and HackerOne reported a 44% reduction in average vulnerability-intake time for its security agents alongside a 25% accuracy improvement (Source: anthropic.com). Independent developer Simon Willison wrote that Sonnet 4.5 was "probably the best coding model in the world (at least for now)," reporting that on preview access it felt like a better model for code than GPT-5-Codex, and highlighting its performance with the Claude.ai code-interpreter feature; he noted that pricing of $3/$15 per million tokens remained higher than GPT-5 and GPT-5-Codex at $1.25/$10, and cautioned that the field moved quickly enough that the lead might not hold (Source: simonwillison.net).

Related models

• Claude Sonnet 4.6 — successor in the Sonnet line (February 2026).
• Claude Opus 4.5 — higher-capability Opus-line model released two months later, benchmarked against Sonnet 4.5.
• Claude Opus 4.6 — Opus-line model approached by Sonnet 4.6 on several evaluations.

Relationships

• developer: Anthropic — Anthropic developed and released the model.
• superseded-by: Claude Sonnet 4.6 — successor in the Sonnet line (February 17, 2026).
• depends-on: Anthropic's Responsible Scaling Policy (Version 3.1) — RSP v3.1 defines the ASL-3 Standard under which Sonnet 4.5 is deployed.
• deployed-by: Claude Code, Anthropic (consumer Claude apps).
• related: System Card: Claude Sonnet 4.5 (Anthropic, September 2025) — pre-deployment safety documentation and primary source for the safety claims.
• related: Claude Opus 4.5 — Opus-line model that uses Sonnet 4.5 as a benchmark reference point.
• related: Emotion Concepts and their Function in a Large Language Model — interpretability research conducted on Sonnet 4.5.

Sources

• Anthropic, Introducing Claude Sonnet 4.5 (Sep 29, 2025) — release, capabilities, benchmarks, pricing, availability, product updates, Claude Agent SDK, safety. (Source: anthropic.com)
• System Card: Claude Sonnet 4.5 (Anthropic, September 2025) — Anthropic system card: ASL-3 classification, alignment assessment, mechanistic-interpretability tests, third-party evaluations.
• Simon Willison, Claude Sonnet 4.5 is probably the "best coding model in the world" (at least for now) (Sep 29, 2025) — reception, pricing comparison, code-interpreter testing, availability across platforms. (Source: simonwillison.net)
• InfoQ, Claude Sonnet 4.5 Tops SWE-Bench Verified, Extends Coding Focus beyond 30 Hours (Oct 11, 2025) — SWE-bench/OSWorld figures, agentic safety score. (Source: infoq.com)
• swebench.com leaderboard — third-party high-reasoning SWE-bench Verified entry. (Source: swebench.com)
• Leanware, Claude Sonnet 4.5: Features, Benchmarks & Pricing Guide — compilation of GPQA Diamond and AIME figures from Anthropic's benchmark table. (Source: leanware.co)

Recommend ingesting the Claude Sonnet 4.5 system card as a foundational source — already covered by System Card: Claude Sonnet 4.5 (Anthropic, September 2025).