Claude Sonnet 4.5

Claude Sonnet 4.5 is Anthropic's mid-tier text and multimodal model in the Claude 4 generation, released on September 29, 2025. It processes text and images, supports hybrid extended-thinking inference, and is distributed through the Anthropic API, claude.ai, Amazon Bedrock, and Google Cloud Vertex AI. At its release, it held the top position on SWE-bench Verified across all evaluated frontier models, making it a reference model for agentic software engineering tasks as of late 2025.

At a glance

• Lab: Anthropic
• Released: September 29, 2025
• Modality: Text and multimodal (vision)
• Open weights: No (closed)
• Context window: 200,000 tokens
• Pricing: $3 per million input tokens, $15 per million output tokens; claude.ai Free, Pro, Max, Team, and Enterprise subscription tiers
• Distribution channels: Anthropic API, claude.ai (web and mobile), Amazon Bedrock, Google Cloud Vertex AI

Origins

The Claude family launched in 2023 when Anthropic released its first public model, Claude 1, built around Constitutional AI (CAI), the company's methodology for training models to follow a written set of principles through self-critique and revision. Claude 2 (2023) expanded context length and instruction following. Claude 3, released March 2024, introduced the tiered naming convention still in use: Opus for the largest variant, Sonnet for the mid-tier, and Haiku for the smallest and fastest. Claude 3.5 Sonnet, released June 2024, extended the family's coding performance substantially and remained a widely used production model through 2025.

The Claude 4 generation opened with Claude Opus 4 and Claude Sonnet 4 in May 2025, introducing expanded agentic capabilities including native tool use, computer use, and Model Context Protocol (MCP) integrations. Claude Sonnet 4 shipped alongside Claude Opus 4 as the more cost-efficient option, carrying the same May 14, 2025 release date, and was positioned as the practical default for most developers and enterprise workflows where per-token cost matters.

Claude Sonnet 4.5, released September 29, 2025, marked the first significant update to the Sonnet tier within the Claude 4 generation. It maintained the same API pricing as Sonnet 4 while delivering substantial performance gains in software engineering, computer use, and agentic operation. The release accompanied a set of product launches: checkpoints for rollback in Claude Code, a native VS Code extension, context editing, memory tools, the Claude Agent SDK, and native code execution and file creation in claude.ai applications. Anthropic described Sonnet 4.5 as "our most aligned frontier model yet" at the time of release, citing reduced rates of deceptive behavior and improved defenses against prompt injection attacks.

Anthropic has not published architecture details for Claude Sonnet 4.5. The parameter count and training data composition have not been disclosed.

Capabilities

Claude Sonnet 4.5's strongest documented capability is software engineering. At release, it scored 77.2% on SWE-bench Verified, the benchmark testing real-world repository bug-fixing on actual GitHub issues, and reached 82.0% in a parallel test-time compute configuration. These figures were the highest published on SWE-bench Verified across all frontier models at that point in time, ahead of GPT-5 and Gemini 2.5 Pro.

Computer use improved substantially relative to the prior Sonnet generation. Claude Sonnet 4.5 scored 61.4% on OSWorld, a benchmark for desktop control through screenshots and mouse/keyboard commands, compared to 42.2% for Claude Sonnet 4 four months earlier. Anthropic cited this as the highest computer use score among publicly evaluated frontier models at the time of the 4.5 launch.

Extended-thinking inference is supported. Claude Sonnet 4.5 can switch between standard fast inference and a deliberate-reasoning mode that generates a chain-of-thought trace before producing a final answer. In standard mode, output arrives at lower latency and cost; in extended-thinking mode, the model works through intermediate steps before finalizing, which improves performance on multi-step mathematical derivations, complex planning tasks, and problems that benefit from explicit verification. Extended thinking adds latency and per-token cost.

Agentic operation is a design focus for the Sonnet 4.5 generation. The model supports parallel tool execution, allowing it to issue multiple tool calls in a single step. Anthropic documented that Sonnet 4.5 maintains focus on complex, multi-step tasks for more than 30 hours, making it suitable for long-running agent workflows. New memory tools shipped alongside the model release allow agents to store and retrieve information outside the context window, extending effective working memory for tasks that exceed the 200,000-token limit.

Vision is supported natively. The model processes images and text together in the same context window, handling chart interpretation, document analysis, screenshot-based workflows, and visual question answering through the API and across claude.ai subscription tiers.

Domain-specific knowledge work in finance, law, medicine, and STEM was an explicit development target for the 4.5 release. Anthropic highlighted improved accuracy on domain-specific expert tasks in these areas without providing category-level benchmark scores.

Benchmarks and standing

As of late 2025, Claude Sonnet 4.5's strongest public benchmark position is on SWE-bench Verified, where it scored 77.2% at release, the leading figure across all publicly evaluated frontier models at that time.

On OSWorld, the desktop control benchmark, Claude Sonnet 4.5 scored 61.4%, described by Anthropic as significantly ahead of other evaluated models at the time of release.

On GPQA Diamond, the graduate-level scientific reasoning benchmark, Claude Sonnet 4.5 scored 83.4%.

Specific scores on LMArena ELO, the Artificial Analysis Intelligence Index, ARC-AGI Challenge, AIME 2025, and HumanEval+ were not prominently published by Anthropic or third-party evaluators in the verified sources consulted for this profile.

Benchmark leadership is point-in-time. Claude Sonnet 4.5 has been succeeded by Claude Sonnet 4.6 as Anthropic's current mid-tier model, and the benchmark landscape has shifted accordingly.

Access and pricing

Claude Sonnet 4.5 is available through the following channels.

The Anthropic API at https://www.anthropic.com/api provides programmatic access for text and vision tasks, tool use, extended thinking, and computer use. Pricing is $3 per million input tokens and $15 per million output tokens, matching the pricing of the original Claude Sonnet 4 it replaced. This positions Sonnet 4.5 at roughly 60% of the cost of Claude Opus 4.7 for input tokens and 60% for output tokens, making it the preferred default for cost-sensitive production deployments.

claude.ai is the consumer and professional product surface. The Free tier provides access to Claude Sonnet 4.5 with usage limits. The Pro tier ($20/month) removes most limits. The Max tier ($100/month or $200/month) provides the highest usage caps and priority access. The Team and Enterprise tiers add collaborative management, compliance features, and expanded context options.

Amazon Bedrock provides a managed deployment of Claude Sonnet 4.5 within the AWS ecosystem, using the model ID anthropic.claude-sonnet-4-5-20250929-v1:0. This is the primary channel for enterprises with existing AWS agreements or data residency requirements. Starting with the Sonnet 4.5 generation, Bedrock offers two endpoint types: global endpoints for dynamic routing with maximum availability and regional endpoints for guaranteed data routing through specific geographic regions.

Google Cloud Vertex AI provides an equivalent managed deployment within the Google Cloud ecosystem, using the model ID claude-sonnet-4-5@20250929. Vertex AI offers global, multi-region, and regional endpoint types from the Sonnet 4.5 generation onward.

Microsoft Foundry provides access through Anthropic's partnership with Microsoft. Availability details align with the broader API access terms.

Comparison

Direct competitors to Claude Sonnet 4.5 in the mid-tier text and multimodal category, as of late 2025:

• GPT-5 (OpenAI). OpenAI's base GPT-5 model competed with Sonnet 4.5 at launch across coding, reasoning, and multimodal tasks. Anthropic cited SWE-bench Verified and OSWorld scores ahead of GPT-5 at the time of the Sonnet 4.5 release. The competitive gap on coding tasks was described as the clearest differentiator for agentic software engineering use cases.
• Gemini 2.5 Pro (Google DeepMind). The main mid-tier competitor from Google at the time of Sonnet 4.5's launch. Anthropic described Sonnet 4.5's SWE-bench Verified and computer use scores as ahead of Gemini 2.5 Pro at release. Gemini 2.5 Pro's distribution advantage through Google Workspace and Android creates access paths that do not depend on API adoption.
• Claude Opus 4.7 (Anthropic). The flagship variant within the same Claude 4 family. Opus 4.7 carries a higher price ($5 input / $25 output per million tokens), a larger 1,000,000-token context window, and higher scores on most aggregate benchmarks. For engineering-intensive tasks requiring the full context window or the highest possible reasoning capability, Opus 4.7 is the appropriate choice. Sonnet 4.5 serves buyers for whom the 200,000-token window is sufficient and cost efficiency matters.
• Claude Haiku 4.5 (Anthropic). The smallest and fastest variant in the Claude 4.5 tier. Haiku 4.5 is priced at $1 input / $5 output per million tokens, runs at higher throughput, and is suited for high-volume, low-latency applications where Sonnet 4.5's capability ceiling is more than required. Haiku 4.5 does not support extended thinking.

Outlook

Open questions for Claude Sonnet 4.5 as of early 2026:

• Transition to Claude Sonnet 4.6. Claude Sonnet 4.6 is the current Anthropic mid-tier recommendation as of April 2026, with a 1,000,000-token context window and updated training data. Sonnet 4.5 remains available as a legacy model but is no longer the primary Sonnet-tier option. Anthropic plans to retire Claude Sonnet 4 (the original, unversioned model) on June 15, 2026; Claude Sonnet 4.5 does not have an announced retirement date as of April 2026.
• Agentic workflow maturity. Claude Sonnet 4.5 was the first Sonnet-tier model where Anthropic placed extended autonomous operation as a primary design goal. Whether the reliability, safety, and commercial uptake of these agentic use cases justify a Sonnet-tier model at this price point remains to be evaluated against the patterns that enterprise customers actually adopt.
• Computer use adoption. The OSWorld score of 61.4% at launch was the highest published figure for any frontier model at the time. Whether computer use as a product feature reaches production adoption at scale is an open question through 2026.
• Claude Sonnet 5. Anthropic has not disclosed timeline information for a Sonnet 5 release. If the Claude 4 generation follows the pattern of iterative point releases observed through late 2025 and early 2026, the next Sonnet-tier update after 4.6 would arrive before any major-version step.

Sources

• Anthropic: Introducing Claude Sonnet 4.5. Official launch announcement, September 29, 2025.
• Anthropic: Claude models overview. Model lineup, API identifiers, context windows, and pricing table; Claude Sonnet 4.5 listed as a legacy model as of April 2026.
• SWE-bench Verified leaderboard. Software engineering benchmark covering real repository bug-fixing tasks; Sonnet 4.5 scored 77.2% at release.
• OSWorld benchmark. Desktop control evaluation; Sonnet 4.5 scored 61.4% at release.
• AWS Blog: Introducing Claude Sonnet 4.5 in Amazon Bedrock. Amazon Bedrock availability and endpoint types.
• Anthropic API pricing. Per-token pricing for Claude Sonnet 4.5 and other Claude models.
• DataCamp: Claude Sonnet 4.5 overview. Third-party benchmark summary and capability overview.