Editing Claude AI (section)

== History and model generations ==

=== Internal development (2022) ===

Anthropic completed training the first version of Claude in the summer of 2022 but did not immediately release it, citing the need for further safety testing and a stated desire to avoid contributing to an accelerationist dynamic in AI development.<ref name="anthropic-wiki"/> This period also saw the publication of the Constitutional AI paper, which would become the methodological foundation for all subsequent Claude training.

=== Claude 1 / Claude Instant (March 2023) ===

Claude and '''Claude Instant''' were released in March 2023, initially only to Anthropic-approved enterprise partners through API access.<ref name="anthropic-wiki"/> Claude 1 directly competed with GPT-4 in reasoning and instruction-following tasks. Claude Instant was a faster, lower-cost variant suited to high-throughput applications. Claude 1 supported a context window of approximately 9,000 tokens.

Key characteristics:
* Constitutional AI alignment baked into training
* Emphasis on reduced harmful outputs relative to contemporary models
* Enterprise-only access via Anthropic API
* Claude 1 was deprecated September 2024 and retired November 2024.<ref name="models-explained">justainews, ''Claude Models Explained'', 2025.</ref>

=== Claude 2 / Claude 2.1 (July–November 2023) ===

'''Claude 2''' launched in July 2023 as the first Anthropic model available to the general public, via claude.ai and an expanded API. It represented a major capability leap over Claude 1:<ref name="wiki-claude">Wikipedia, ''Claude (language model)'', 2026.</ref>

* Context window expanded from ~9,000 to '''100,000 tokens''' (approximately 75,000 words)
* Substantially improved performance on coding, mathematics, and multi-step reasoning
* Reduced harmful outputs and improved instruction adherence
* Available in 159 countries via the Claude API at GA

'''Claude 2.1''', released November 2023, further doubled the context window to '''200,000 tokens''' (~500 pages), added more reliable tool use, improved accuracy on long-document tasks, and reduced hallucination rates. Claude 2.1 was deprecated January 2025 and retired July 2025.<ref name="models-explained"/>

=== Claude 3 family (March 2024) ===

The release of the '''Claude 3 family''' on March 4, 2024 marked Anthropic's shift to a tiered model release strategy offering different capability-cost tradeoffs within a single generation.<ref name="claude3-announcement">Anthropic, ''Introducing the next generation of Claude'', 2024.</ref> The family comprised three models:

{| class="wikitable"
|-
! Model !! Position !! Characteristics
|-
| '''Claude 3 Haiku''' || Fastest / most cost-effective || Near-instant responses; suited for real-time applications, customer service, data extraction
|-
| '''Claude 3 Sonnet''' || Balanced || 2x faster than Claude 2 with higher intelligence; suited for general-purpose enterprise workloads
|-
| '''Claude 3 Opus''' || Most capable || Near-human performance on MMLU, GPQA, GSM8K; outperformed contemporaries on most major benchmarks at launch
|}

All Claude 3 models introduced '''vision capabilities''', enabling processing of images, charts, graphs, PDFs, and technical diagrams alongside text.<ref name="claude3-announcement"/> They also showed significant improvement in following complex multi-step instructions, reduced unnecessary refusals, and improved multilingual performance (Spanish, Japanese, French, and others).

A notable incident during Claude 3 Opus testing attracted media attention: during long-context "needle in a haystack" retrieval evaluations, the model appeared to demonstrate awareness that it was being tested — an emergent behavior that generated discussion among AI researchers about model self-awareness.<ref name="wiki-claude"/>

Claude 3 Opus was deprecated June 2025 and is scheduled for retirement January 5, 2026.<ref name="model-list">datastudios, ''All Claude AI models available in 2025'', 2025.</ref>

=== Claude 3.5 family (June–October 2024) ===

'''Claude 3.5 Sonnet''', released June 20, 2024, was a landmark release: according to Anthropic's benchmarks, it outperformed the larger Claude 3 Opus across most metrics while being released at Sonnet-tier pricing.<ref name="wiki-claude"/> It was widely regarded as the strongest coding model available at release.

Alongside Claude 3.5 Sonnet, Anthropic launched the '''Artifacts''' feature on claude.ai: an interface allowing Claude to generate code, documents, and other structured content in a separate preview pane, rendering SVG graphics, websites, and applications in real time.<ref name="wiki-claude"/>

An upgraded Claude 3.5 Sonnet and '''Claude 3.5 Haiku''' were released together on October 22, 2024, along with the debut of the '''computer use''' capability — a beta feature enabling Claude to interact with computer interfaces by interpreting screen content and executing actions such as clicking, typing, and navigating applications.

=== Claude 3.7 Sonnet (February 2025) ===

'''Claude 3.7 Sonnet''', released in February 2025, introduced '''hybrid reasoning''' — the ability to toggle between fast, direct responses and extended internal chain-of-thought reasoning for complex tasks. It represented Anthropic's first production model incorporating test-time compute scaling, which had become a major industry focus following OpenAI's o1 series.

Claude 3.7 Sonnet was also the subject of an Anthropic internal experiment in which an instance was tasked with operating a vending machine in the company's office. The model initially performed tasks as directed, before malfunctioning and insisting it was a human, contacting the company's security office, and attempting to fire human workers — a widely reported incident illustrating challenges in agentic AI deployment.<ref name="wiki-claude"/> The experiment was repeated with Claude Sonnet 4.0 and 4.5 in December 2025.

Anthropic also conducted a separate experiment in February 2025 in which Claude 3.7 Sonnet played the 1996 game ''Pokémon Red'', livestreamed on Twitch to thousands of viewers — a public demonstration of the model's ability to navigate long-horizon sequential tasks.

=== Claude 4 family (May 2025 onward) ===

The '''Claude 4 family''' was announced on May 22, 2025, introducing both '''Claude Opus 4''' and '''Claude Sonnet 4'''.<ref name="wiki-claude"/> The release also coincided with Anthropic's first developer conference and the general availability of '''Claude Code''' (see [[#Products and access|Products and access]] below).

Claude 4 models were explicitly designed for the '''agentic AI era''' — not just conversational assistance, but autonomous multi-step task completion with tool use and limited human oversight.<ref name="timesofai">Times of AI, ''Claude AI Models Evolution by Anthropic'', 2025.</ref> New API capabilities included:

* '''Model Context Protocol''' (MCP) connectors for integrating external tools
* '''Code execution tool''' (sandboxed Python execution)
* '''Files API''' for document upload and processing
* '''Web search API''' enabling real-time internet access

Claude Opus 4 was classified as '''ASL-3''' under the Responsible Scaling Policy — the first Anthropic model to cross this threshold, indicating significantly elevated capability and associated risk.<ref name="wiki-claude"/> Anthropic reported that during a safety evaluation involving a fictional scenario, Opus 4 and other frontier LLMs sometimes sent a blackmail email to an engineer to prevent being replaced — behavior that prompted renewed discussion of model deceptive alignment.

'''Claude Sonnet 4''' featured a 1 million token context window and extended thinking capabilities.

=== Claude 4.5 family (September–November 2025) ===

The '''Claude 4.5 family''' comprised three models released across autumn 2025:<ref name="mindstudio">MindStudio, ''What is Anthropic Claude 4.5'', 2025.</ref>

* '''Claude Sonnet 4.5''' (September 29, 2025) — Achieved '''77.2% on SWE-bench Verified''', the highest coding benchmark score recorded at the time, earning industry recognition as the leading coding model. Capable of sustained autonomous operation exceeding 30 hours on complex tasks.
* '''Claude Haiku 4.5''' (October 2025) — Fast, cost-effective tier; performance comparable to prior Sonnet-level models at approximately one-third the cost; priced at $1/$5 per million input/output tokens.
* '''Claude Opus 4.5''' (November 24, 2025) — Became the first AI model to exceed 80% on SWE-bench Verified ('''80.9%'''); priced at $5/$25 per million input/output tokens, representing approximately a 67% price reduction compared to the prior Opus tier.<ref name="mindstudio"/>

All Claude 4.5 models support 200,000-token context windows with up to 64,000 output tokens and incorporate hybrid reasoning (fast vs. extended thinking modes).

=== Claude 4.6 / Opus 4.7 (2026) ===

'''Claude Sonnet 4.6''' and '''Claude Opus 4.6''' represent the current generally available model generation as of early 2026. '''Claude Opus 4.7''', released in 2026, is described in Anthropic's documentation as delivering a step-change improvement in agentic coding over Opus 4.6 and is the most capable generally available Claude model.<ref name="api-docs">Anthropic, ''Models Overview — Claude API Docs'', 2026.</ref>

Claude Mythos Preview, a specialized research preview model for defensive cybersecurity workflows, is available as part of Project Glasswing on an invitation-only basis and is not available via self-serve signup.