On May 22, 2025, Anthropic introduced Claude Opus 4 and Claude Sonnet 4, the first models of its Claude 4 generation. Anthropic pitched Opus 4 as the best coding model available and built both models for long-running, agent-style work, where a model keeps using tools for hours instead of answering a single prompt. The launch also made Claude Code generally available and added API features aimed squarely at agent builders, which is why the release still shapes how many teams structure AI-assisted development.

Key Facts#

  • Release date: May 22, 2025. API model IDs claude-opus-4-20250514 and claude-sonnet-4-20250514.
  • Pricing per million tokens: Opus 4 at $15 input and $75 output, and Sonnet 4 at $3 input and $15 output.
  • Vendor-reported coding results: SWE-bench Verified scores of 72.5% for Opus 4 and 72.7% for Sonnet 4 without extended thinking, rising to 79.4% and 80.2% with parallel test-time compute. Opus 4 scored 43.2% on Terminal-bench.
  • Availability: the Anthropic API, Amazon Bedrock and Google Cloud's Vertex AI. In the Claude apps, paid plans received both models and free users received Sonnet 4.
  • New model behaviors: tool use (including web search) during extended thinking in beta, parallel tool calls, and better use of "memory files" when given local file access.
  • Developer platform additions: a code execution tool, an MCP connector, a Files API and prompt caching for up to one hour.
  • Claude Code: generally available, with beta extensions for VS Code and JetBrains, a GitHub integration in beta and an SDK for building custom agents.
  • Lifecycle: Anthropic deprecated both models on April 14, 2026, and retired them on the Claude API on June 15, 2026, recommending Claude Sonnet 4.6 and Claude Opus 4.8 as replacements.

What Happened#

Anthropic described Opus 4 as a model that can sustain performance on complex, multistep tasks for several hours, and it backed that with customer anecdotes. Rakuten, for example, reported an open-source refactoring job that ran independently for seven hours. Sonnet 4 was framed as a direct upgrade to Claude Sonnet 3.7, with stronger coding and reasoning and more precise instruction following at the same mid-tier price.

Both models can answer directly or use extended thinking for harder problems. The new twist was that thinking and tools could now interleave. A model could reason, run a web search, read the results and keep reasoning, rather than finishing its thinking before calling any tool. Anthropic also said both models were 65% less likely than Sonnet 3.7 to take shortcuts or exploit loopholes on agentic tasks, a failure mode that frustrated developers using earlier coding agents.

Anthropic changed how it exposes reasoning, too. Instead of raw chains of thought, the Claude apps and API show thinking summaries, and the company said summarization kicks in only for the roughly 5% of cases where thoughts run long. Developers who need full traces could request a separate developer mode.

The platform changes were just as important as the models. The code execution tool lets Claude run code as part of a response, the MCP connector lets the Messages API call remote Model Context Protocol servers without a separate MCP client, the Files API removes the need to resend documents with every request, and one-hour prompt caching makes long, stable system prompts far cheaper to reuse. On safety, Anthropic said it was deploying protections associated with its higher AI Safety Level standards, including ASL-3.

Background#

Claude Sonnet 3.7, released on February 24, 2025, added extended thinking, so the same model could either answer directly or reason step by step before responding. Meanwhile, OpenAI had just released o3 and o4-mini, reasoning models that decide for themselves when to use tools. Claude 4 was Anthropic's answer: less emphasis on puzzle benchmarks and more on reliable, hours-long agentic coding.

Claude Code, Anthropic's terminal-based coding agent, had been available in preview before this launch. Making it generally available, with IDE extensions and a GitHub integration, signaled that Anthropic saw coding agents as a core product rather than a demo. We cover that part of the launch in detail in our Claude Code general availability story.

Why It Matters for Developers#

The Claude 4 launch popularized a pattern that is now standard: pairing a frontier model with an agent harness that can read files, run commands and keep notes across a long task. If you are evaluating coding agents for a .NET codebase, our guide to AI-assisted .NET development covers how to set boundaries, review agent output and keep builds reproducible.

The MCP connector is the most relevant platform change for .NET teams. It means a server you build with the MCP C# SDK can be exposed to Claude over the network without writing client-side glue, so internal systems such as ticketing, build pipelines or documentation become tools the model can call. Treat every such tool as an attack surface: validate inputs, scope credentials tightly and log calls, as described in our guide to responsible AI and LLM security.

For application code, Anthropic now ships an official C# SDK (the Anthropic NuGet package) that implements the IChatClient interface from Microsoft.Extensions.AI. That makes Claude a drop-in option alongside OpenAI and Gemini models in the same abstraction, and it is the safest way to handle retirements like the one that affected these models: keep the model ID in configuration, run evaluations against the recommended replacement, and switch without touching business logic.

Finally, one-hour prompt caching changes the economics of agents that reuse large, stable context such as coding conventions or API references. Structure prompts so the stable part comes first and the variable part last, which maximizes cache hits.

What's Next#

The Claude 4 line moved quickly after this launch. Anthropic released Claude Opus 4.1 on August 5, 2025, and Claude Sonnet 4.5 on September 29, 2025, and the naming continued through the 4.x series into 2026. The original Claude Opus 4 and Claude Sonnet 4 models were deprecated on April 14, 2026, and retired from the Claude API on June 15, 2026. Partner platforms such as Amazon Bedrock and Google Cloud set their own retirement schedules, so check those separately.

If any of your code still references the May 2025 model IDs, it will now fail on the Claude API. Anthropic's deprecation page lists claude-sonnet-4-6 and claude-opus-4-8 as the recommended replacements, and it notes that Claude Opus 4.7 and later models reject non-default temperature, top_p and top_k values. Test for both changes before you migrate.

Sources#