Claude (Anthropic)

From Systems analysis Wiki
Jump to navigation Jump to search

Claude is a family of multimodal large language models (LLMs) developed by the research company Anthropic.

Built on a transformer architecture, Claude models are positioned as AI assistants focused on safety, usefulness, and honesty. A key feature of their development is the Constitutional AI methodology, aimed at creating steerable and ethically aligned systems.

History and Philosophy

Founding and Mission of Anthropic

Anthropic was founded in 2021 by former senior employees of OpenAI, including siblings Dario[1] and Daniela Amodei[2]. Their departure was prompted by disagreements with OpenAI's leadership regarding its direction, particularly concerns that the partnership with Microsoft and increasing commercialization could compromise its commitment to AI safety.

Anthropic's mission is "to develop and maintain advanced AI for the long-term benefit of humanity." The company is registered as a Public Benefit Corporation (PBC) in the United States, which legally obligates it to balance financial interests with public benefit. This approach is reinforced by a unique governance structure featuring the Long-Term Benefit Trust (LTBT), an independent body with the authority to influence the composition of the board of directors to ensure adherence to the safety mission.

Philosophy: HHH and Constitutional AI

The behavior of Claude models is based on the HHH formula: Helpful, Honest, and Harmless. To implement these principles, Anthropic developed its own training methodology—Constitutional AI (CAI).

Unlike traditional RLHF (Reinforcement Learning from Human Feedback), where human annotators directly rate the model's responses, CAI uses a "constitution"—a set of explicit ethical principles—based on which the model learns to self-evaluate and correct its own answers. This makes the process more scalable, transparent, and controllable.

Architecture and Key Technologies

Transformer-based Foundation

Like other modern LLMs, Claude uses a decoder-only transformer architecture that autoregressively generates text token by token. However, Anthropic has introduced significant improvements aimed at enhancing performance, safety, and steerability.

Constitutional AI (CAI)

The CAI training process consists of two stages:

  1. Supervised Learning Phase: The model generates responses to prompts, and then another critic model, guided by the "constitution," evaluates them and suggests revisions. The original model is then fine-tuned based on these revisions.
  2. Reinforcement Learning from AI Feedback (RLAIF) Phase: The model generates pairs of responses, and the critic model selects the better one based on the "constitution." This data is used to train a preference model (reward model), which then serves as a signal to fine-tune the main model using reinforcement learning algorithms.

Long Context and Multimodality

One of Claude's main advantages is its very large context window. Starting from 100k tokens in Claude 2, it was increased to 200k in Claude 3 and up to 1-2 million tokens in versions 3.5 and 4. This allows the models to analyze entire books, codebases, or multi-hour transcripts within a single prompt.

Starting with the Claude 3 family, the models became multimodal, gaining the ability to process images alongside text.

Hybrid Reasoning and Agentic Capabilities

Starting with versions Claude 3.7 and 4, a hybrid reasoning architecture was introduced. It allows the models to switch between two modes:

  • Fast Answers: The standard mode for simple tasks.
  • Extended Thinking: For complex tasks, the model pauses to "think," performing internal reasoning steps, calling tools (web search, code execution), and formulating a more well-founded response. This makes the process more transparent and reliable.

Evolution of Claude Models

Claude 1 and 2 (2023)

  • Claude 1 (March 2023): The first publicly available version, released through the Anthropic API. It introduced the Claude Instant model for fast, lightweight tasks and a flagship version that soon offered a 100,000 token context window—among the largest available at the time.[3][4]
  • Claude 1.3 (2023): A later first-generation flagship revision that improved performance and safety and became the principal Claude 1 model immediately preceding Claude 2. Anthropic used Claude 1.3 as the comparison baseline when announcing Claude 2.[5]
  • Claude 2 (July 2023): An improved version that became publicly available via a web interface. It showed significant improvements in coding (71% on Codex HumanEval) and mathematics. In November 2023, Claude 2.1 was released with a 200,000 token context window and reduced hallucination rates.[4][3]
  • Claude Instant 1.2 (August 2023): The final major update to the lightweight Instant line. It incorporated improvements from Claude 2 while retaining lower latency and cost, with gains in mathematics, coding, reasoning, multilingual performance, structured responses, and safety.[6]

Claude 3 (March 2024)

A family of models that surpassed GPT-4 on several benchmarks for the first time.[7]

  • Versions: Haiku (fastest), Sonnet (balanced), and Opus (most powerful)—establishing the three-tier structure that became the basis of Anthropic's mainstream Claude lineup and was later supplemented by the higher-capability Mythos tier.
  • Key Innovations: Introduction of multimodality (image analysis), significant improvements in reasoning and coding, and a reduction in unwarranted refusals. Opus achieved 86.8% on MMLU.[7][8]

Claude 3.5 and 3.6 (2024)

An intermediate generation focused on increasing intelligence, coding ability, tool use, and speed.

  • Claude 3.5 Sonnet (June 2024): Surpassed Claude 3 Opus in performance while being twice as fast.[9] It introduced the Artifacts feature[10]—an interactive panel for working with generated code or documents.
  • Claude Sonnet 3.6 (October 2024): Released at the time as an upgraded Claude 3.5 Sonnet and later referred to by Anthropic as Claude Sonnet 3.6. It delivered major improvements in coding and tool use and became the first frontier Claude model with Computer Use in public beta, allowing it to inspect a screen, move a cursor, click, and type.[11][12]
  • Claude 3.5 Haiku (October 2024): The second model in the 3.5 family and the successor to Claude 3 Haiku. It retained Haiku-class speed while substantially improving coding and general intelligence, matching or exceeding Claude 3 Opus on a number of evaluations.[11]

Claude 3.7 (February 2025)

A transitional release centered on controllable reasoning.

  • Claude 3.7 Sonnet: Introduced hybrid reasoning, allowing the model to combine fast answers with deep, step-by-step reasoning within a single model. It was Anthropic's first model with an explicit extended-thinking mode, and it launched alongside the Claude Code command-line agent, laying the groundwork for the agentic capabilities of the next generation.[3]

Claude 4 (2025–2026)

A flagship generation focused on autonomous AI agents and complex, long-horizon reasoning.

  • Opus 4 and Sonnet 4 (May 2025): The first Claude 4 models, built around autonomous AI agents. The models are capable of performing multi-step tasks, working with the file system (via Computer Use), calling tools, and maintaining long-running work sessions (up to several hours) without performance degradation. Opus 4 achieved 72.5% on the complex coding benchmark SWE-bench Verified.[13][3]
  • Point releases: The generation was refined through a rapid series of incremental upgrades—Opus 4.1 (August 2025, 74.5% SWE-bench Verified)[14], Sonnet 4.5 (September 2025, 77.2%)[15] and Haiku 4.5 (October 2025)[16], followed by Opus 4.5 (November 2025), which raised the SWE-bench Verified score to 80.9%.[17][18]
  • Claude Opus 4.6 and Sonnet 4.6 (February 2026): Extended the generation with stronger coding, computer use, long-context reasoning, agent planning, cybersecurity, knowledge work, and design. Opus 4.6 introduced a 1-million-token context window in beta, which was also adopted by Sonnet 4.6.[19][20]
  • Claude Opus 4.7 (April 2026): Further improved coding, visual understanding, self-verification, and long-running professional and agentic work; it also increased the resolution supported for image inputs and expanded the available reasoning-effort range.[21]
  • Claude Opus 4.8 (May 2026): The final major Opus release of the fourth generation, emphasizing reliability, judgment, autonomous long-horizon coding, and professional work. It introduced a cheaper Fast mode running at roughly 2.5 times normal speed and launched alongside dynamic workflows for large-scale agent orchestration.[22] By this stage, the leading Claude 4 models supported a 1-million-token context window and adjustable reasoning effort.[23]

Mythos tier (2026)

A new capability tier positioned above the Opus class, rather than simply another point release within the mainstream Haiku/Sonnet/Opus hierarchy.

  • Claude Mythos Preview (April 2026): The first publicly disclosed Mythos-class model, initially released on an invitation-only basis through Project Glasswing. Although a general-purpose language model, it was particularly notable for frontier-level cybersecurity capabilities and was used by Anthropic to explore highly capable defensive-security agents under controlled access.[24]
  • Claude Fable 5 and Claude Mythos 5 (June 2026): The first production models in the Mythos class. Fable 5 became generally available on June 9 as the safety-tuned version intended for broad use, while Mythos 5, the successor to Mythos Preview, offered the same underlying capabilities with different safeguards and remained limited to vetted participants in Project Glasswing. Both models provided a 1-million-token context window and up to 128,000 output tokens.[25][26] Access to the models was temporarily suspended in June under a U.S. government export-control directive and subsequently restored after the restrictions were lifted.[27]
  • Claude Fable 5.1 and Claude Mythos 5.1 (September 2026): A major update to the Mythos-class family. The two versions use the same underlying model but differ in their safeguard and access regimes: Fable 5.1 is broadly available, while Mythos 5.1 remains available through trusted-access programs, including Project Glasswing. They retain the 1-million-token context window and 128,000-token maximum output, use always-on adaptive thinking, and substantially improve long-running agentic coding, multistep research, knowledge work, scientific problem solving, and self-verification.[28] Fable 5.1 also introduced more precise cybersecurity and biology safeguards and substantially reduced the price of prompt-cache reads.[28]

Claude 5 (2026)

The fifth generation of Anthropic's mainstream model lineup, alongside the higher-capability Fable/Mythos branch.

  • Claude Sonnet 5 (June 2026): The most agentic Sonnet model to date, released as a substantial improvement over Sonnet 4.6, with performance approaching that of Opus 4.8 at lower cost. It can autonomously plan, operate browsers and terminals, call tools, and execute extended workflows. Sonnet 5 introduced a standard 1-million-token context window, up to 128,000 output tokens, adaptive thinking, and granular reasoning-effort controls.[12]
  • Claude Opus 5 (July 2026): A step-change upgrade over Opus 4.8 and the first fifth-generation Opus model. It focuses on deep reasoning, complex agentic coding, enterprise work, long-horizon autonomous tasks, code review, computer use, and multi-agent coordination. Anthropic described it as approaching the frontier intelligence of Fable 5 at half the per-token price. Opus 5 has a 1-million-token context window, supports up to 128,000 output tokens, uses thinking by default, and expands the reasoning-effort scale through low, medium, high, xhigh, and max levels.[29][30]
  • Current lineup: As of September 2026, Anthropic lists Claude Fable 5.1, Claude Opus 5, Claude Sonnet 5, and Claude Haiku 4.5 as its current principal models. No Claude Haiku 5 has yet been released; Haiku 4.5 therefore remains the current low-latency Haiku model.[23]

Summary Table of Claude Generations

Evolution of Key Characteristics of Claude Models
Generation / Model Line Release Year Key Versions Max. Context Window Key Innovations
Claude 1 2023 Claude, Claude Instant, Claude 1.3 9,000 → 100,000 tokens First public Claude assistant family; context expanded from 9K to 100K tokens during 2023.
Claude 2 2023 Claude 2, Claude 2.1 100,000 → 200,000 tokens Stronger coding, mathematics and reasoning; improved instruction following; 200K context introduced with Claude 2.1.
Claude 3 2024 Opus 3, Sonnet 3, Haiku 3 200,000 tokens Introduction of the Opus/Sonnet/Haiku capability tiers; native vision input; major gains in reasoning, coding and multilingual performance.
Claude 3.5 2024 Sonnet 3.5, Haiku 3.5 200,000 tokens Higher intelligence at lower cost; major gains in coding and visual reasoning; introduction of Artifacts.
Claude 3.6 2024 Sonnet 3.6 200,000 tokens Upgraded Sonnet model with stronger coding and agentic tool use; first Claude model with general-purpose Computer Use in public beta.
Claude 3.7 2025 Sonnet 3.7 200,000 tokens First hybrid-reasoning Claude model, combining standard responses with extended thinking; stronger agentic behavior; Claude Code research preview.
Claude 4 2025 Opus 4, Sonnet 4 200,000 tokens Major advance in agentic coding and long-running workflows; stronger parallel tool use, planning and sustained multi-step work.
Claude 4.1 2025 Opus 4.1 200,000 tokens Incremental upgrade to Opus 4 with stronger agentic task performance, coding and real-world reasoning.
Claude 4.5 2025 Sonnet 4.5, Haiku 4.5, Opus 4.5 200,000 tokens Stronger computer use and agent orchestration; improved coding and long-horizon work; effort controls and context-management improvements.
Claude 4.6 2026 Opus 4.6, Sonnet 4.6 1,000,000 tokens 1M-token context as a standard capability; adaptive thinking; stronger long-context reasoning, computer use, coding and long-running agentic workflows; transition to dateless pinned model IDs.
Claude 4.7 2026 Opus 4.7 1,000,000 tokens Improved performance on difficult software-engineering and long-running agentic tasks; stronger self-verification and high-resolution vision.
Claude 4.8 2026 Opus 4.8 1,000,000 tokens Improved reliability and judgment in agentic work; Fast mode at approximately 2.5× normal speed; effort control and dynamic workflows for large-scale parallel agent orchestration.
Claude Fable 2026 Fable 5, Fable 5.1 1,000,000 tokens Generally available Mythos-class line above Opus in capability; introduced fifth-generation Claude intelligence for the hardest coding, knowledge-work, research and long-running agentic tasks. Fable 5.1 adds stronger autonomous execution, verification and scientific-research capabilities with more precise cybersecurity and biology safeguards.
Claude Mythos 2026 Mythos Preview, Mythos 5, Mythos 5.1 1,000,000 tokens Restricted-access Mythos-class line for frontier cybersecurity, biology and advanced research. Mythos 5 and Fable 5 share the same underlying model, as do Mythos 5.1 and Fable 5.1; Mythos variants provide reduced safeguards for vetted organizations through trusted-access programs.
Claude Opus 5 2026 Opus 5 1,000,000 tokens Fifth-generation Opus model for complex agentic coding and enterprise work; stronger long-horizon reasoning and autonomous execution; performance approaching Fable-class intelligence at lower cost.
Claude Sonnet 5 2026 Sonnet 5 1,000,000 tokens Most agentic Sonnet generation; adaptive thinking; autonomous browser and terminal use; major gains in reasoning, coding, tool use and knowledge work while retaining lower latency and cost than Opus.
Claude Haiku 2024–present Haiku 3, Haiku 3.5, Haiku 4.5 200,000 tokens Fastest and lowest-cost Claude line, optimized for low-latency and high-volume workloads. Haiku 4.5 remains the current Haiku model and provides near-frontier intelligence with extended thinking.

Models

No. Date Model Type Note Source System / Model Card
1 2023-03-14 Claude Claude First public Claude assistant; offered together with Claude Instant. Anthropic —
2 2023-03-14 Claude Instant Instant Faster, lighter, lower-cost version of the first public Claude assistant. Anthropic —
3 2023-04 Claude 1.3 Claude Later first-generation Claude model; used as the comparison baseline for Claude 2. Anthropic Claude 2 Model Card
4 2023-07-11 Claude 2 Claude Second-generation Claude model with improved coding, math, reasoning, longer responses, and 100K-token context. Anthropic Claude 2 Model Card
5 2023-08-09 Claude Instant 1.2 Instant Updated Claude Instant model with improved math, coding, reasoning, safety, and structured responses. Anthropic —
6 2023-11-21 Claude 2.1 Claude Claude 2 update with a 200K-token context window, lower hallucination rates, system prompts, and beta tool use. Anthropic Claude 2 Model Card
7 2024-03-04 Claude 3 Opus Opus Most capable model in the Claude 3 family. Anthropic Claude 3 Model Card
8 2024-03-04 Claude 3 Sonnet Sonnet Mid-tier Claude 3 model balancing intelligence and speed. Anthropic Claude 3 Model Card
9 2024-03-13 Claude 3 Haiku Haiku Fastest and lowest-cost model in the Claude 3 family. Anthropic Claude 3 Model Card
10 2024-06-20 Claude 3.5 Sonnet Sonnet First Claude 3.5 model; higher intelligence than Claude 3 Opus with Sonnet-tier speed and cost. Anthropic Claude 3.5 Sonnet Model Card Addendum
11 2024-10-22 Claude 3.5 Sonnet (new) / Claude Sonnet 3.6 Sonnet Upgraded October 2024 version of Claude 3.5 Sonnet with improved coding, tool use, and computer-use capability; later Anthropic publications refer to this model as Claude Sonnet 3.6. Anthropic; later Anthropic nomenclature Claude 3.5 Haiku and Upgraded Claude 3.5 Sonnet Model Card Addendum
12 2024-10-22 Claude 3.5 Haiku Haiku Faster and cheaper Claude 3.5-family model. Anthropic Claude 3.5 Haiku and Upgraded Claude 3.5 Sonnet Model Card Addendum
13 2025-02-24 Claude 3.7 Sonnet Sonnet First hybrid reasoning Claude model, with standard and extended-thinking modes. Anthropic Claude 3.7 Sonnet System Card
14 2025-05-22 Claude Opus 4 Opus Opus-class hybrid reasoning model for coding, advanced reasoning, and long-running agent workflows. Anthropic Claude 4 System Card
15 2025-05-22 Claude Sonnet 4 Sonnet Sonnet-class hybrid reasoning model; upgrade from Claude 3.7 Sonnet. Anthropic Claude 4 System Card
16 2025-08-05 Claude Opus 4.1 Opus Incremental upgrade to Claude Opus 4 with improvements in agentic tasks, coding, and reasoning. Anthropic Claude Opus 4.1 System Card
17 2025-09-29 Claude Sonnet 4.5 Sonnet Sonnet model focused on coding, complex agents, computer use, reasoning, and math. Anthropic Claude Sonnet 4.5 System Card
18 2025-10-15 Claude Haiku 4.5 Haiku Fast, cost-efficient Haiku model with near-frontier performance. Anthropic Claude Haiku 4.5 System Card
19 2025-11-24 Claude Opus 4.5 Opus Opus model for coding, agents, computer use, deep research, spreadsheets, and slides. Anthropic Claude Opus 4.5 System Card
20 2026-02-05 Claude Opus 4.6 Opus Opus upgrade with stronger coding, cybersecurity investigation, long-running agentic work, and 1M-token context in beta. Anthropic Claude Opus 4.6 System Card
21 2026-02-17 Claude Sonnet 4.6 Sonnet Sonnet upgrade across coding, computer use, long-context reasoning, agent planning, knowledge work, and design. Anthropic Claude Sonnet 4.6 System Card
22 2026-04-07 Claude Mythos Preview Mythos Invitation-only research-preview model for defensive cybersecurity workflows through Project Glasswing. Anthropic Claude Mythos Preview System Card
23 2026-04-16 Claude Opus 4.7 Opus Opus-tier upgrade over Opus 4.6 with stronger coding, vision, and long-running professional work. Anthropic Claude Opus 4.7 System Card
24 2026-05-28 Claude Opus 4.8 Opus Opus-tier upgrade for complex reasoning, long-horizon agentic coding, and high-autonomy work. Anthropic Claude Opus 4.8 System Card
25 2026-06-09 Claude Fable 5 Fable Mythos-class model made available for general use with additional safeguards. Access to the Mythos-class models (Fable 5 and Mythos 5) was suspended on 2026-06-12 under a U.S. government export-control directive; the models were not deprecated or retired. The U.S. Department of Commerce lifted the export controls on 2026-06-30, and Anthropic began restoring Fable 5 globally on 2026-07-01 (Anthropic statement). Anthropic Claude Fable 5 and Mythos 5 System Card
26 2026-06-09 Claude Mythos 5 Mythos Limited-availability Mythos-class model through Project Glasswing; same underlying model class as Fable 5 with different safeguards/access. Access to both Mythos-class models was suspended on 2026-06-12 under a U.S. government export-control directive. The export controls were lifted on 2026-06-30; Mythos 5 access is being restored to a set of approved U.S. organizations (following U.S. government approval on 2026-06-26), with Anthropic continuing to work to expand access through Project Glasswing (Anthropic statement). Anthropic Claude Fable 5 and Mythos 5 System Card
27 2026-06-30 Claude Sonnet 5 Sonnet Most agentic Sonnet model at launch; major upgrade in reasoning, coding, tool use, and knowledge work, with autonomous browser and terminal use and performance approaching Opus 4.8 at lower cost. Anthropic Claude Sonnet 5 System Card
28 2026-07-24 Claude Opus 5 Opus Fifth-generation Opus model for long-running agents, coding, reasoning, and professional knowledge work; approaches Fable 5 frontier intelligence at lower cost and adds stronger self-verification and iterative problem solving. Anthropic Claude Opus 5 System Card
29 2026-09-01 Claude Fable 5.1 Fable Generally available Mythos-class model and successor to Fable 5; major gains in coding, knowledge work, scientific research, and long-running autonomous problem solving. Fable 5.1 is the same underlying model as Mythos 5.1, with additional cybersecurity and biology safeguards for general availability. Anthropic Claude Fable 5.1 and Mythos 5.1 System Card
30 2026-09-01 Claude Mythos 5.1 Mythos Restricted-access Mythos-class model for vetted cybersecurity and life-sciences users; same underlying model as Fable 5.1 but with more permissive safeguards. It improves on Mythos 5 in cybersecurity, biology, scientific research, and long-running agentic work. Anthropic Claude Fable 5.1 and Mythos 5.1 System Card

Benchmarks

No. Date Model Typea MMLUb GSM8K HumanEval MATH GPQA-Dc MMMU SWE-Vd SWE-Proe
1 2023-03-14 Claude Claude
2 2023-03-14 Claude Instant Instant
3 2023-04 Claude 1.3 Claude 77.0f 85.2 56.0
4 2023-07-11 Claude 2 Claude 78.5 88.0 71.2
5 2023-08-09 Claude Instant 1.2 Instant 86.7 58.7
6 2023-11-21 Claude 2.1 Claude
7 2024-03-04 Claude 3 Opus Opus 86.8 95.0 84.9 60.1 50.4 59.4
8 2024-03-04 Claude 3 Sonnet Sonnet 79.0 92.3 73.0 43.1 40.4 53.1
9 2024-03-13 Claude 3 Haiku Haiku 75.2 88.9 75.9 38.9 33.3 50.2
10 2024-06-20g Claude 3.5 Sonnet Sonnet 88.7 96.4 92.0 71.1 59.4 68.3 33.4h
11 2024-10-22 Claude 3.5 Sonnet (new) Sonnet 88.7 93.7 78.3 65.0 70.4 49.0
12 2024-10-22 Claude 3.5 Haiku Haiku 77.6 88.1 69.2 41.6 40.6
13 2025-02-24 Claude 3.7 Sonnet Sonnet 62.3i
14 2025-05-22 Claude Opus 4 Opus 72.5j
15 2025-05-22 Claude Sonnet 4 Sonnet 72.7j
16 2025-08-05 Claude Opus 4.1 Opus 74.5
17 2025-09-29 Claude Sonnet 4.5 Sonnet 77.2k
18 2025-10-15 Claude Haiku 4.5 Haiku 73.3
19 2025-11-24 Claude Opus 4.5 Opus 80.7 80.9l
20 2026-02-05 Claude Opus 4.6 Opus 91.3 80.8l 53.4
21 2026-02-17 Claude Sonnet 4.6 Sonnet 89.9 79.6 58.1
22 2026-04-07 Claude Mythos Preview Mythos 94.6 93.9 77.8
23 2026-04-16 Claude Opus 4.7 Opus 94.2 87.6m 64.3
24 2026-05-28 Claude Opus 4.8 Opus 93.6n 88.6 69.2
25 2026-06-09 Claude Fable 5 Fable 95.0o 80.0
26 2026-06-09 Claude Mythos 5 Mythos 94.1q 95.5o 80.3
27 2026-06-30 Claude Sonnet 5 Sonnet 85.2p 63.2p

Table notes

  • a Tiers follow Anthropic's naming. Before Claude 3 there were no named tiers: "Claude" is the flagship line (Claude 1/2), "Instant" the fast, low-cost line (retrospectively the Haiku lineage). From Claude 3 on, the family splits into Haiku, Sonnet and Opus. "Mythos" and "Fable" form the Mythos-class tier above Opus; Fable 5 and Mythos 5 share one underlying model.
  • b Early-era figures (rows 3–12) are Anthropic's own, from model cards/addenda; settings differ by benchmark and generation. MMLU: 5-shot CoT for Claude 1.3 and Claude 2 (rows 3–4), plain 5-shot (not the higher 5-shot CoT variant) for the Claude 3 and Claude 3.5 rows (7–12). GSM8K is shown only where Anthropic reported it and uses 0-shot CoT; the October 2024 Claude 3.5 addendum reports MGSM rather than GSM8K, so the GSM8K cells for rows 11–12 are left blank. HumanEval (Codex): 0-shot. MATH: 0-shot for Claude 3 (rows 7–9), 0-shot CoT for the Claude 3.5 rows (10–12). GPQA-D: 0-shot CoT. MMMU: validation split, 0-shot with chain-of-thought where reported — for rows 10–11, but not for Claude 3.5 Haiku (row 12), which launched text-only. See the Claude 2 model card, the Claude 3 model card, and the June/October Claude 3.5 addenda for exact settings.
  • c Modern GPQA-D figures (rows 20–24 and 26, where reported) were run under newer Anthropic evaluation settings, often with adaptive/extended thinking at high or max effort, and are not directly comparable to the 0-shot CoT figures in earlier rows. Anthropic notes that GPQA-D has become saturated for top models.
  • d SWE-V figures are Anthropic-reported and generally use Anthropic's simple SWE-bench scaffold (bash + string-replacement file editing) on the full 500-problem set. Trials, prompt additions, context length, thinking budget, and high-/adaptive-compute settings vary by release (see row-specific notes); they are not comparable to independent leaderboards (e.g. vals.ai), which use a different harness.
  • e SWE-Pro (SWE-bench Pro, Scale AI) is a harder set. Figures shown are from Anthropic's own harness; on Scale's standardized leaderboard the same models score markedly lower (e.g. Opus 4.6 ≈ 47–52%).
  • f Row 3 (Claude 1.3) figures come from the comparison table in the Claude 2 model card, where "Claude 1.3" is a named column.
  • g The API snapshot ID is claude-3-5-sonnet-20240620 (June 20, 2024); Anthropic's announcement page is dated June 21, 2024. This table uses the snapshot date.
  • h 33.4 (Claude 3.5 Sonnet) is the June 2024 release figure. The upgraded version (Oct 2024, row 11) scores 49.0.
  • i 62.3 (Claude 3.7 Sonnet) is on the full 500-problem set. On the n=489 subset: 63.7 without a scaffold pass and 70.3 with one.
  • j 72.5 / 72.7 (Opus 4 / Sonnet 4) are standard mode. In "high compute" (parallel attempts plus internal candidate selection): 79.4 / 80.2.
  • k 77.2 (Sonnet 4.5) is standard mode, 200K context, averaged over 10 trials, no test-time compute. "High compute" gives 82.0; the 1M configuration gives 78.2.
  • l 80.9 / 80.8 (Opus 4.5 / Opus 4.6) are both Anthropic figures but under different configs: 4.5 with no thinking budget, averaged over 5 trials; 4.6 with adaptive thinking, max effort, averaged over 25 trials (exactly 80.84; 81.42 with a prompt modification). The ~0.1 difference is noise, not a regression.
  • m 87.6 (Opus 4.7) is from Anthropic's system card; independent vals.ai reports 82.0. The jump over Opus 4.6 largely reflects a grader change, not only the model.
  • n 93.6 (Opus 4.8) — the benchmark is saturated (~93–94% for top models) and differences are within statistical noise; this cell may also be left blank.
  • o Fable 5 and Mythos 5 share the same underlying model but are reported separately in the system card: Fable 5's public safeguards route high-risk requests to a Claude Opus 4.8 fallback, which slightly lowers its measured scores. On SWE-bench Verified / SWE-bench Pro, Fable 5 scores 95.0 / 80.0 and Mythos 5 scores 95.5 / 80.3. Anthropic's combined launch chart ("Claude Mythos 5 / Fable 5") shows the higher Mythos-class figure (80.3).
  • p The Claude Sonnet 5 system card reports SWE-bench Verified = 85.2 and SWE-bench Pro = 63.2, averaged over five trials with Anthropic's standard SWE-bench configuration. Anthropic's comparison for the Pro figure is Opus 4.8's 69.2.
  • q Claude Mythos 5 scored 94.1% on GPQA Diamond in the Fable 5 / Mythos 5 system card. Anthropic did not publish a separate Fable 5 GPQA-D figure and describes the benchmark as saturated, stating it plans to stop reporting future performance on it.

Application and Availability

Claude models are available through several channels:

  • Web interface claude.ai: Provides free access (to the Sonnet model) and paid subscriptions (Pro, Max) with access to more powerful models (Opus) and higher limits.
  • Developer API: Anthropic offers a commercial API that allows developers to integrate Claude into third-party applications. Prices vary depending on the model (Haiku is the cheapest, Opus is the most expensive).
  • Cloud Platforms: Claude is available through Amazon Bedrock and Google Cloud Vertex AI, which simplifies its use in enterprise environments.
  • Integrations: Claude is integrated into popular services such as Slack, Notion, and Quora (in the Poe chatbot).

Literature

  • Anthropic (2022). Constitutional AI: Harmlessness from AI Feedback. arXiv:2212.08073.
  • Anthropic (2023). Claude’s Constitution.
  • Anthropic (2024, March). Introducing the next generation of Claude.
  • Anthropic (2025, May). Introducing Claude 4.
  • Ouyang, L. et al. (2022). Training Language Models to Follow Instructions with Human Feedback. arXiv:2203.02155.
  • Bai, Y. et al. (2022). Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback. arXiv:2204.05862.
  • Bai, Y. et al. (2022). Constitutional AI: Harmlessness from AI Feedback. arXiv:2212.08073.
  • Bulatov, A. et al. (2023). Scaling Transformer to 1M Tokens and Beyond with RMT. arXiv:2304.11062.
  • Jimenez, C. E. et al. (2023). SWE-bench: Can Language Models Resolve Real-World GitHub Issues?. arXiv:2310.06770.
  • Yuan, W. et al. (2024). Self-Rewarding Language Models. arXiv:2401.10020.
  • Yang, A. et al. (2024). Context Parallelism for Scalable Million-Token Inference. arXiv:2411.01783.
  • Miranda, L. J. V. et al. (2024). Hybrid Preferences: Learning to Route Instances for Human vs AI Feedback. arXiv:2410.19133.
  • Chittepu, Y. et al. (2025). Reinforcement Learning from Human Feedback with High-Confidence Safety Constraints. arXiv:2506.08266.
  • Yuan, W. et al. (2025). Process-based Self-Rewarding Language Models. arXiv:2503.03746.
  • Yang, B. et al. (2025). Long Context Windows in Generative AI: An AI Atlas Report. [31] (tech-report, open review).

References

  1. ↑ “Dario Amodei”. In Wikipedia [1]
  2. ↑ “Daniela Amodei”. In Wikipedia [2]
  3. ↑ 3.0 3.1 3.2 3.3 Anthropic Claude Model Release Timeline (Opus / Sonnet / Haiku family tree). [3]
  4. ↑ 4.0 4.1 “Claude Version History (2023–2026): Complete Timeline”. [4]
  5. ↑ Anthropic, “Claude 2”. [5]
  6. ↑ Anthropic, “Releasing Claude Instant 1.2”. [6]
  7. ↑ 7.0 7.1 Anthropic, “Introducing the next generation of Claude”. [7]
  8. ↑ Anthropic, “The Claude 3 Model Family: Opus, Sonnet, Haiku” (model card). [8]
  9. ↑ Anthropic, “Introducing Claude 3.5 Sonnet”. [9]
  10. ↑ “What Are Artifacts and How Do I Use Them? | Anthropic Help Center”. [10]
  11. ↑ 11.0 11.1 Anthropic, “Introducing computer use, a new Claude 3.5 Sonnet, and Claude 3.5 Haiku”. [11]
  12. ↑ 12.0 12.1 Anthropic, “Introducing Claude Sonnet 5”. [12]
  13. ↑ Anthropic, “Introducing Claude 4”. [13]
  14. ↑ Anthropic, “Claude Opus 4.1”. [14]
  15. ↑ Anthropic, “Introducing Claude Sonnet 4.5”. [15]
  16. ↑ Anthropic, “Introducing Claude Haiku 4.5”. [16]
  17. ↑ Anthropic, “Introducing Claude Opus 4.5”. [17]
  18. ↑ “Claude Opus 4.5: Benchmarks, Agents, Tools, and More”, DataCamp. [18]
  19. ↑ Anthropic, “Claude Opus 4.6”. [19]
  20. ↑ Anthropic, “Introducing Sonnet 4.6”. [20]
  21. ↑ Anthropic, “Introducing Claude Opus 4.7”. [21]
  22. ↑ Anthropic, “Introducing Claude Opus 4.8”. [22]
  23. ↑ 23.0 23.1 “Models overview”, Claude Platform Docs. [23]
  24. ↑ Anthropic, “Assessing Claude Mythos Preview’s cybersecurity capabilities”. [24]
  25. ↑ Anthropic, “Claude Fable 5 and Claude Mythos 5”. [25]
  26. ↑ Claude Platform Docs, “Introducing Claude Fable 5 and Claude Mythos 5”. [26]
  27. ↑ Anthropic, “Redeploying Fable 5”. [27]
  28. ↑ 28.0 28.1 Anthropic, “Claude Fable 5.1 and Claude Mythos 5.1”. [28]
  29. ↑ Anthropic, “Introducing Claude Opus 5”. [29]
  30. ↑ Claude Platform Docs, “What's new in Claude Opus 5”. [30]
Retrieved from "https://systems-analysis.info/eng/index.php?title=Claude_(Anthropic)&oldid=90"