What Is Claude Sonnet 3.5 | Performance of the Legacy Model and Differences from Sonnet 4.6

What Is Claude Sonnet 3.5 — Release Timeline and Positioning

Claude Sonnet 3.5 is the mid-tier model of the Claude 3.5 generation, announced by Anthropic on June 20, 2024. Its API ID is claude-3-5-sonnet-20240620. The lineup at that time consisted of "Opus 3 (top tier) · Sonnet 3.5 (mid tier) · Haiku 3 (lightweight)," and Sonnet 3.5 was positioned as the flagship model striking a balance between speed and intelligence.

Notably, Sonnet 3.5 outperformed the then-flagship Opus 3 on many benchmarks. Anthropic's official release reported improvements on GPQA (graduate-level reasoning), MMLU (knowledge), and HumanEval (coding), and at 1/5 the price of Opus 3 (input $3 / output $15 per MTok), it quickly gained support from developers who prioritized cost-performance.

On October 22, 2024, an improved version claude-3-5-sonnet-20241022 (also known as new Sonnet 3.5 / Sonnet 3.5 v2) was released, with the public beta of computer use (screen operation) launching at the same time.

Source: Anthropic Models overview / Model deprecations

Sonnet 3.5 Performance and Spec Details

The main specs for Sonnet 3.5 (common to both versions) are as follows.

The 200k token context window was standard at the time and handled long PDFs and multi-file code analysis. On the other hand, the model did not include "thinking before reasoning" features such as Extended Thinking and Adaptive Thinking that appeared in later models, making it an immediate-response model that prioritized response speed.

The pricing is on par with Sonnet 4.6 ($3/$15 per MTok), but intelligence, coding performance, and long-context support are significantly inferior.

Improvements in "New Sonnet 3.5" (v2) — The Arrival of Computer Use

The October 2024 v2 (claude-3-5-sonnet-20241022), referred to by Anthropic officially as "Claude 3.5 Sonnet (New)," announced the following improvements:

1. Recorded 49.0% on SWE-bench Verified (real GitHub issue resolution), setting the highest score at the time
2. Significant improvement on TAU-bench (agent-style tool use)
3. Public beta of computer use (viewing screenshots and operating mouse and keyboard)
4. Overall quality improvements in coding, instruction following, and reasoning

The computer use feature attracted attention as the first full-scale capability allowing AI to directly operate a PC's GUI, and the Anthropic Cookbook published a demo of operating a desktop inside a Docker container. This direction evolved into the agent capabilities of the later Claude 4 series, the subagent features of Claude Code, and Skills.

However, v2 also received a deprecation notice on 2025-08-13 and was retired on the same date as the initial version, 2025-10-28.

Can Sonnet 3.5 Still Be Used — Retirement Timeline

To put it plainly, Claude Sonnet 3.5 is no longer available as of May 2026. The history from Anthropic's official model deprecations page is as follows.

Sonnet 3.5 Deprecation History

1. June 20, 2024: claude-3-5-sonnet-20240620 released
2. October 22, 2024: claude-3-5-sonnet-20241022 (new Sonnet 3.5) released
3. August 13, 2025: Anthropic notifies deprecation of both versions (following the rule of 60+ days advance notice)
4. October 28, 2025: Both API IDs officially retired (calls fail)

After retirement, calls fail with 404 or invalid model errors in the Claude API, Claude Platform on AWS, and Microsoft Foundry. Amazon Bedrock and Vertex AI operate on their own schedules, so refer to the official AWS / Google documentation for availability periods on those platforms.

Anthropic's officially recommended migration target is claude-sonnet-4-6 (as of May 2026).

Source: Anthropic Model deprecations - 2025-08-13: Claude Sonnet 3.5 models

Differences Between Sonnet 3.5 and Sonnet 4.6 — Performance and Feature Comparison

Comparing Claude Sonnet 3.5 with the current Claude Sonnet 4.6 on key items, there is roughly 1.5 generations of difference as follows.

Notable is that while the price remains unchanged, the context window is 5x larger, the maximum output is approximately 8x larger, and reasoning features (Thinking series) have been significantly added. If you are paying the same $3/$15 per MTok, the performance gap makes the case for choosing Sonnet 4.6 clear.

The knowledge cutoff has also been extended to August 2025, and it includes recent ecosystem information such as Claude Code and MCP servers, making Sonnet 4.6's advantage particularly prominent for developer use cases.

Source: Anthropic Models overview

Steps for Migrating from Sonnet 3.5 to Sonnet 4.6

Code calling retired models will start returning errors, so please migrate using the following steps.

Step 1: Audit Your Usage

Export usage logs by API key in CSV format from the Usage page in the Claude Console, and check whether claude-3-5-sonnet-20240620 or claude-3-5-sonnet-20241022 appears. On the codebase side, use grep -r "claude-3-5-sonnet" to identify all references.

Step 2: Replace the Model ID

Rewrite the "model" field in the request payload to "claude-sonnet-4-6". In the Anthropic SDK (Python / TypeScript), the format is client.messages.create(model="claude-sonnet-4-6", ...). In many cases, simply replacing the ID is sufficient.

Step 3: Consider Leveraging New Features

With Sonnet 4.6, you can use Extended Thinking (thinking={"type": "enabled", "budget_tokens": ...}) and Adaptive Thinking (a 2026 new feature controlled with the thinking_effort parameter). Enabling these for complex reasoning tasks will significantly improve quality compared to Sonnet 3.5.

Step 4: Take Advantage of 1M Token Context

Long-document tasks that were cut off at 200k with Sonnet 3.5 can now be processed all at once up to 1M tokens with Sonnet 4.6. There may be cases where you can remove document-splitting logic, so review your codebase.

Step 5: Benchmarking and Regression Testing

After switching models, perform A/B comparisons with business-specific prompts and record output quality, token consumption, and latency. Anthropic officially recommends "running migration tests well before the retirement date," and this same discipline will be needed when migrating to future next-generation models.

Source: Migration guide

Common Pitfalls for Existing Apps That Used Sonnet 3.5

Common issues seen in apps calling the retired Sonnet 3.5 include the following:

• Hardcoded model ID: If not set as an environment variable, you will need to update every deployment. Going forward, using environment variables like ANTHROPIC_MODEL=claude-sonnet-4-6 is recommended
• Hardcoded max_tokens=8192: Since Sonnet 4.6 can output up to 64k, raising the limit for summaries and long-form generation will improve the experience
• Not using Extended Thinking: Not using the thinking parameter for reasoning tasks means you won't get the full value of Sonnet 4.6
• Via AWS Bedrock / Vertex AI: The retirement dates differ between Anthropic itself and AWS / Google, so always check the official documentation for each cloud you are using

Especially for SaaS integrations and customer-facing production apps, calls to retired models will lead to customer-facing errors, so switching early is the safe approach.

Summary — Sonnet 3.5 Is Fully Retired, Migrate to Sonnet 4.6

Claude Sonnet 3.5 emerged in 2024 as Anthropic's flagship model and attracted significant attention with benchmarks surpassing Opus 3 at the time and the announcement of the computer use beta. However, it was officially retired on 2025-10-28 and can no longer be called in the current API.

The successor Claude Sonnet 4.6 (claude-sonnet-4-6) is a powerful successor model at the same price range, featuring a 1M token context window, Extended Thinking, and Adaptive Thinking. If you have code referencing Sonnet 3.5, proceed with migration by combining an API ID swap with consideration of new feature adoption.

Note that Anthropic has committed to "long-term storage of past model weights and considering future re-release for researchers," so for research use cases that want to reproduce Sonnet 3.5 behavior, it is worth watching future Anthropic announcements.

Related Articles

• Differences Between Claude Opus, Sonnet, and Haiku | Usage Decision Chart
• Claude Opus 4.7 Release Explained
• Claude Opus 4.7 Migration Guide
• Getting Started with the Claude API