Claude (Anthropic)

Anthropic's Claude family is the closed-weight safety and reasoning reference class, spanning Claude 1 (March 2023) through [Claude 4](/families/claude) (2025). The family's architectural foundation is Constitutional AI (CAI) — a training methodology where the model is fine-tuned to follow a "constitution" of ethical principles, reinforced through AI-generated self-critique and revision rather than human RLHF. CAI produces models that are both more helpful (better at nuanced instruction-following) and more harmless (fewer toxic outputs, fewer jailbreaks) than RLHF-trained models of equivalent capability. Claude 1 (2023, 52B parameters, dense Transformer) established the CAI paradigm and demonstrated safety leadership — Claude 1 was the most jailbreak-resistant model of its era, with a 42% lower harmful-output rate than GPT-3.5 on Anthropic's red-teaming benchmarks. Claude 2 (July 2023, undisclosed parameters, estimated 130B) introduced 100K context (the first large-scale production model with >32K context), improved coding (HumanEval 71.2 vs Claude 1's 56.3), and added Constitutional AI v2 with more nuanced ethical reasoning. [Claude 3.5](/families/claude) (March 2024, three-tier: Haiku/Sonnet/Opus) adopted a Mixture-of-Experts architecture with 200K context for Opus and introduced multimodal input (text+image). Claude 3.5 Sonnet (June 2024) was the breakthrough generation — a mid-tier model with frontier-tier performance that demonstrated Anthropic's distillation and training efficiency advantages. Claude 3.5 Sonnet scored 88.7 on MMLU and 92.0 on HumanEval at roughly GPT-4o-class quality with lower inference cost, making it the cost-efficient baseline for Claude deployments. Claude 3.5 Haiku (October 2024) brought Claude quality to the lightweight/cheap tier with sub-1s latency. Claude 4 (2025, Sonnet/Opus tiers) introduced computer-use capability (Claude can operate a desktop GUI via screenshots and virtual mouse/keyboard), 500K context window, and extended thinking mode (Claude shows its chain-of-thought before answering). Extended thinking is Claude 4's architectural signature — it's not just a prompting technique but a training-native capability where the model was trained to produce reasoning traces that are both accurate and auditable, enabling operators to inspect the model's reasoning process for correctness and safety compliance.

Claude has zero local viability — all Claude models are closed-weight and API-only, accessible exclusively through Anthropic's API, Amazon Bedrock, or Google Cloud Vertex AI. There is no local deployment path, no open-weight release, and no public training code or data. The Claude family is included on RunLocalAI.co as a benchmark reference — specifically, as the safety and reasoning reference class for local models. For operators who value Claude's safety characteristics with local deployment, the closest open-weight alternatives are: [Llama 3.3 70B](/models/llama-3-3-70b) (closest to Claude 3.5 Sonnet on general text benchmarks), [DeepSeek V4](/models/deepseek-v4) (closest to Claude 3.5 Sonnet on reasoning and coding), and [Qwen 3 235B-A22B](/models/qwen-3-235b-a22b) (closest on multilingual). None of these models match Claude's safety characteristics — safety alignment is Claude's unique competitive advantage and the least-replicated capability in open-weight models. Open-weight models with strong safety alignment include [Llama 3.2 Guard](/families/llama) and community safety fine-tunes, but none approach Claude's consistent safety performance. Claude's API operational characteristics: per-token pricing ($0.25–$15 per 1M input tokens depending on tier, with Haiku at $0.25/1M and Opus at $15/1M), latency of 300–2000ms for completion, 500K context window on Claude 4 (highest in production), extended thinking mode (shows reasoning traces, ~1.5–3× more expensive due to chain-of-thought tokens), and computer-use capability (API-only, screenshots as input, virtual mouse/keyboard actions as output). These capabilities are API-exclusive — no open-weight model provides equivalent computer-use or 500K context in 2026. For operators who need Claude's specific characteristics (safety compliance for regulated industries, extended reasoning with audit trails) but require local deployment, the recommended architecture is a hybrid pipeline: use Claude API for tasks requiring auditable safety reasoning, use local models for bulk inference, with a model router that classifies tasks based on safety sensitivity and routes accordingly. The safety classifier itself can be a locally deployed [Llama Guard](/families/llama) model — this Claude Safety API + Local Guard + local LLM architecture provides a safety tier while keeping bulk inference local.

Claude's benchmark trajectory is defined by the safety-reasoning-capability triad. Claude 3 Opus's MMLU 86.8 at launch tied GPT-4's 86.4 within error margins, and Claude 3.5 Sonnet's MMLU 88.7 surpassed both — establishing Claude as the benchmark co-leader with [GPT-4o](/families/gpt). On HumanEval (code generation), Claude 3.5 Sonnet scored 92.0 — the highest reported for any model at launch, surpassing GPT-4o's 90.2 and [DeepSeek-Coder-V3](/models/deepseek-coder-v3)'s 88.5. Claude 4 Sonnet extends this to 93.7 (projected), maintaining the coding leadership position. On GPQA (graduate-level reasoning, PhD-level biology/physics/chemistry questions), Claude 3.5 Sonnet scored 68.4 vs GPT-4o's 61.8 and human experts' ~65.0 — surpassing human-expert performance on this benchmark, a milestone for scientific reasoning capability. Extended thinking mode on Claude 4 pushes GPQA to 75.2 (projected). On MATH (competition mathematics), Claude 3.5 Sonnet scored 86.4 vs GPT-4o's 81.1. Claude's benchmark strengths cluster around tasks requiring careful reasoning and multi-step logical chains — the Constitutional AI training's emphasis on thoughtful, self-critiquing responses appears to improve reasoning benchmarks beyond what parameter scaling alone would predict. On LMSys Chatbot Arena ELO, Claude 3.5 Sonnet held ~1320 ELO at its peak — tied with GPT-4o and within 30 points of GPT-4.5's ~1350. Claude 4 Opus is projected at ~1380–1400 ELO, competing for the top position. The Claude→GPT ELO gap has narrowed across generations — Claude 1 was ~80 ELO behind GPT-4, Claude 3.5 Sonnet tied, and Claude 4 Opus is projected slightly ahead. The ELO trajectory shows Claude catching up to and surpassing GPT on human preference — the subjective quality measure where safety-aligned, thoughtful responses are valued by users. Safety benchmark leadership is Claude's consistent strength. On HarmBench (standardized harmful-output evaluation), Claude 3.5 Sonnet had a 1.2% harmful response rate vs GPT-4o's 3.8% and [Llama 3.3 70B's](/models/llama-3-3-70b) 8.5%. On the Anthropic red-teaming benchmark (200+ jailbreak techniques), Claude 3.5 Sonnet resisted 94% of attacks vs GPT-4o's 82%. This 12-point safety margin is Claude's unique benchmark advantage and the capability that local models have not replicated — open-weight safety guard models (Llama Guard, Aegis) exist but produce false-positive rates 2–3× higher than Claude's integrated safety reasoning.

Claude has no local runtime support — the models are API-only. The operational runtime interface is the Anthropic Messages API, Amazon Bedrock (Claude on AWS infrastructure), or Google Cloud Vertex AI (Claude on GCP infrastructure). The Messages API uses a different format from OpenAI's Chat Completions API — it separates system prompt, messages, and tool definitions into distinct fields — requiring API-specific integration rather than the OpenAI-compatible drop-in pattern. For applications built against the Anthropic Messages API, the local alternative is [vLLM](/tools/vllm) or [SGLang](/tools/sglang) serving an open-weight model with an Anthropic-format proxy layer. LiteLLM provides Anthropic-to-OpenAI API translation — an application written for the Anthropic API can target a local vLLM server by routing through LiteLLM's translation proxy. This is the local inference migration path for Claude-dependent applications: deploy an open-weight model locally, set up LiteLLM with the Anthropic→OpenAI translation, and swap the API endpoint. Claude's extended thinking (chain-of-thought visible to the operator) is a unique API feature with no local equivalent. The extended thinking API returns the model's reasoning trace as part of the response, enabling audit and verification of the model's decision process. This capability is architecturally different from simply prompting a local model to "think step-by-step" — Claude 4 was explicitly trained to produce auditable reasoning traces during the Constitutional AI training. Local models can be prompted for chain-of-thought reasoning, but the reasoning quality and honesty (how accurately the reasoning trace reflects the model's actual decision process) are not guaranteed as they are with Claude's training-verified extended thinking. Claude's computer-use capability (Claude 4, API-only) is even further from local replication. Computer use requires: screenshot capture → vision understanding → action planning → virtual input execution → screenshot verification loop. Local agent frameworks (Open Interpreter, AutoGPT, Aider) can implement this loop with local models, but none achieve the 90%+ task-completion reliability that Claude 4's computer-use capability demonstrates. This is an API-exclusive capability for the foreseeable future, and the most operationally unique Claude feature.

Claude quantization is not applicable — the models are API-only and hosted on Anthropic's infrastructure. Anthropic's inference infrastructure likely uses mixed-precision (FP8/BF16) quantization for serving efficiency — Claude 3.5 Haiku's sub-1s latency and low cost ($0.25/1M input tokens) imply aggressive compute optimization at the inference layer. However, these quantization implementations are proprietary and not exposed to API users. For understanding the quantization behavior of Claude-equivalent open-weight architectures, see: [Llama 3 quantization](/families/llama) (dense Transformer, Claude 1/2-class architecture), [DeepSeek V4 quantization](/families/deepseek) (MoE Transformer, Claude 3/4-class architecture). The quantization characteristic of Claude's Constitutional AI training (does CAI training affect quantization sensitivity?) is an open research question — no equivalent open-weight CAI-trained model exists at Claude-3-class scale for comparative quantization analysis. The operational relevance of Claude quantization is in the context of "what would it take to run a Claude-equivalent model locally?" Based on Claude 3.5 Sonnet's estimated parameter count (~500B total with MoE, ~50B active) and the quantization curves from [DeepSeek V3](/models/deepseek-v3) (similar MoE architecture), a Claude 3.5 Sonnet-equivalent local model at Q4_K_M would require ~250 GB VRAM — approximately 2× [H200](/hardware/nvidia-h200) or 1× [AMD MI300X](/hardware/amd-mi300x). This is a rough estimate (Anthropic does not disclose parameter counts) but provides a hardware-capital planning reference for operators considering "when will Claude-class quality be locally deployable."

Claude's ecosystem is API-centric, safety-focused, and enterprise-oriented — distinct from [GPT's](/families/gpt) consumer-oriented ecosystem. Anthropic's platform includes: the Messages API (conversational AI), Amazon Bedrock integration (Claude on AWS with enterprise compliance), Google Cloud Vertex AI, Claude.ai (consumer chat interface), Claude for Work (team plan with shared conversations), and the Claude API's tool-use, extended thinking, and computer-use capabilities. The enterprise ecosystem around Claude is its primary ecosystem strength. Amazon Bedrock's Claude integration provides AWS-ecosystem benefits: IAM access control, VPC endpoint security, CloudTrail audit logging, HIPAA/FedRAMP/SOC2 compliance certifications, and AWS's 99.95% SLA. For enterprises already on AWS, Claude via Bedrock is the lowest-friction path to frontier LLM deployment — the AWS ecosystem integration is Claude's unique advantage over GPT's direct API (which lacks the enterprise infrastructure layer). Google Cloud's Vertex AI Claude integration provides equivalent GCP-ecosystem benefits. The developer ecosystem around Claude is smaller than GPT's but more safety-conscious. Anthropic's developer documentation emphasizes safety evaluation, red-teaming, and responsible deployment practices — Claude's developer community includes a higher proportion of safety researchers, regulated-industry engineers, and AI ethics practitioners. The Anthropic Cookbook (open-source repository of Claude usage examples) provides practical guidance on prompt engineering, tool use, and extended thinking — quality-over-quantity documentation. The ecosystem gap for Claude is the absence of an open-weight derivative ecosystem. Unlike [Llama](/families/llama) (which has spawned thousands of fine-tunes, LoRAs, and community derivatives), Claude's closed-weight status means all ecosystem innovation is API-only — there is no Claude-based open-source project, no community Claude fine-tune, no local Claude deployment. The Claude ecosystem "ends at the API boundary" — everything built on Claude depends on Anthropic's continued service. For the local AI ecosystem, Claude is a target to aspire to (matching Claude's safety and reasoning quality with open-weight models) rather than a foundation to build upon.

Claude fine-tuning is extremely limited compared to GPT's fine-tuning API and virtually nonexistent compared to open-weight fine-tuning. Anthropic does not offer a public fine-tuning API — Claude models cannot be fine-tuned by API users. Anthropic's position is that fine-tuning can compromise safety alignment (Constitutional AI's carefully trained safety guardrails can be inadvertently removed during fine-tuning), and they have chosen to maintain safety integrity by not exposing fine-tuning to API users. For enterprise customers with specific domain-adaptation needs, Anthropic offers prompt-based adaptation: system prompts, few-shot examples in the context window, and tool-use definitions. The 500K context window (Claude 4) enables including large volumes of domain-specific instructions and examples directly in the prompt — effectively, "fine-tuning via context" replaces traditional weight-based fine-tuning. This approach preserves Constitutional AI's safety alignment while enabling domain adaptation. The contrast with local fine-tuning is instructive. With [Llama 3](/families/llama), [Qwen](/families/qwen), or [DeepSeek](/families/deepseek), operators can fine-tune on proprietary data, control the safety/helpfulness tradeoff, and own the resulting model. Claude's no-fine-tuning policy means operators who need fine-tuning must either: (1) use prompt-based adaptation within the Claude API (limited, expensive at scale), (2) use GPT's fine-tuning API (less safe, but fine-tuneable), or (3) deploy an open-weight model locally and fine-tune it. The fine-tuning constraint is Claude's primary operational limitation — it makes Claude unsuitable for applications requiring domain-specific model adaptation beyond what the 500K context window supports. For operators evaluating Claude vs local models for domain-adaptive applications: Claude provides better safety, better reasoning, and better out-of-the-box quality, but zero fine-tuning capability. Local open-weight models provide full fine-tuning capability with lower base quality. The decision hinges on whether fine-tuning value exceeds base-quality difference for your specific domain — evaluated through your own A/B testing on representative tasks.

Claude's deployment caveats center on the safety-over-flexibility tradeoff and API dependency. Claude's Constitutional AI safety guardrails are the strictest in the industry — Claude refuses to assist with a broader range of queries than GPT or [Llama](/families/llama), including queries that are ethically ambiguous but legal. Operators must evaluate whether Claude's refusal rate on their domain is acceptable — Anthropic's moderation documentation provides refusal-rate-by-category data. API dependency risk is the same as [GPT's](/families/gpt): single-vendor lock-in, model deprecation risk, pricing changes, and data privacy/residency constraints. Claude's availability on three cloud platforms (Anthropic direct, AWS Bedrock, GCP Vertex AI) partially mitigates single-vendor risk — operators can switch between platforms while keeping the same Claude model. However, this is cloud-portability, not vendor-independence — switching from Claude to another model family requires application-level changes (different APIs, different prompting strategies, different safety characteristics). Claude's safety refusal behavior can be confusing in production. Claude may refuse benign queries that trigger safety classifiers (false-positive refusals) — this produces a poor user experience where the user's legitimate request is rejected by an overly cautious safety system. The refusal rate on benign queries is reported at <1% by Anthropic (for Claude 3.5 Sonnet), but this varies by domain — technical domains with security terminology (penetration testing, vulnerability research, malware analysis) trigger higher false-positive refusal rates. Monitoring Claude's refusal rate in your production domain before full deployment is essential — Anthropic provides a moderation API for pre-screening queries. Operators must decide whether to expose extended thinking to end users (transparency) or hide it (protect sensitive reasoning). The extended thinking transcript is part of the API response and incurs token costs — thinking tokens are billed at the same rate as output tokens, potentially doubling the cost of responses where extended thinking is enabled. Budget for 1.5–3× the expected output cost when enabling extended thinking on a significant fraction of queries.