The Claude 4 generation, in retrospect

Claude 4 launched in May 2025 and was retired in June 2026. An archive of what shipped, the benchmark numbers at launch, and the current models that replaced it.

Why this page still exists: Claude 4 is cited in thousands of blog posts, tutorials, procurement documents and Stack Overflow answers written between May 2025 and mid-2026. Version history is genuinely useful — you need to know what a claude-opus-4-20250514 string in an old repository meant, and what to replace it with.

Claude 4 was the model generation Anthropic released on 22 May 2025, consisting of two models: Claude Opus 4 and Claude Sonnet 4. It was the generation that made agentic coding Claude's defining product category, and it ran as Anthropic's frontier line for roughly four months before Sonnet 4.5 superseded part of it in September 2025. Both original Claude 4 models are now retired.

#What was Claude 4, and what shipped with it?

Claude 4 was a two-model release. Opus 4 was positioned as the frontier coding and reasoning model; Sonnet 4 was the balanced, high-volume model that replaced Claude 3.5 Sonnet in most production systems. Both shipped on the same day with the same 200,000-token context window and the same dated model-ID convention, claude-{name}-4-20250514.

Three things arrived alongside the models in May 2025, and all three outlived them:

• Hybrid reasoning became standard. The extended-thinking mode introduced with Claude 3.7 Sonnet three months earlier was carried into both Claude 4 models, with interleaved thinking available behind a beta header. That line of work eventually became the adaptive thinking and effort ladder used across the current lineup — the mechanics are explained on the thinking and reasoning page.
• Claude Code went generally available. The terminal agent that had launched as a research preview with 3.7 Sonnet became a shipping product. It is now the centre of gravity for Anthropic's developer story, covered on Claude for developers.
• Tool use during reasoning. Claude 4 models could call tools between thinking steps rather than only after them, which is the pattern every agent framework depends on today. See MCP and Claude agents for how that evolved into the current tool ecosystem.

Anthropic also activated its ASL-3 deployment and security standard for Opus 4 — the first time the company applied that tier of safeguard to a released model.

#Claude 4 launch benchmarks (May 2025 figures)

These are the numbers Anthropic published in its own launch material on 22 May 2025. They are historical. SWE-bench Verified is no longer reported by frontier vendors — Anthropic, OpenAI and Google have all moved on to newer agentic evaluations, so these figures are useful only as a baseline series showing how fast the category moved, not as a way to compare Claude 4 with anything current.

The most revealing row is the last one. By October 2025, Anthropic's cheapest model scored higher than its May 2025 flagship. That five-month gap between frontier and budget is the single clearest illustration of why running a retired model to save money stopped making sense.

Note also that Sonnet 4 edged out Opus 4 on SWE-bench Verified at launch, despite Opus costing five times as much. Opus 4's advantage showed up on longer, messier, multi-step work rather than on a single scored benchmark — a distinction that still applies when choosing between tiers today, and one worked through on the model comparison page.

#Where Claude 4 sits in the Claude timeline

Claude 4 was the fourth major generation and the first to be treated as a family that would receive point releases rather than a single monolithic launch.

The "4" branding stretched a long way. Opus 4.8, shipped in May 2026, still carried a 4-series name despite having almost nothing technically in common with the May 2025 original: a 1M-token context instead of 200k, adaptive thinking instead of a manual thinking budget, and a different tokenizer. Treat the version number as a marketing lineage, not a statement about architecture.

#What to migrate a Claude 4 model string to

If you have a codebase, notebook or workflow still carrying a Claude 4 model ID, it is broken today, not merely dated. Here is the mapping.

#Is there any reason to still want Claude 4?

Not really — but the honest reasons people ask are worth answering. Reproducibility is the main one: if you published research or a regression suite pinned to claude-opus-4-20250514, those results cannot be reproduced, because Anthropic does not keep retired weights callable. That is a genuine limitation of the whole industry's release model, and it argues for archiving your own outputs rather than trusting a model ID to stay alive.

Cost is not a reason. Opus 4 cost $15 per million input tokens and $75 per million output; Opus 5 costs $5 and $25. Sonnet 5 is $2/$10 against Sonnet 4's $3/$15, though the newer tokenizer emits roughly 30% more tokens for the same text, so the real saving is smaller than the sticker price implies. Current rates for every tier are on the pricing page.

Capability is not a reason either. The individual model records are preserved on the Claude Opus 4 archive and the Claude Sonnet 4 archive, and the wider lineage is set out on the Opus and Sonnet family pages. For a current-generation decision, start at the model index.

#Frequently asked questions