Model Timeline

Organization: OpenAI. Parameters: 1.5 billion. Architecture: Decoder-only Transformer. Training data: WebText (40GB of web pages filtered by Reddit upvote quality).

Key contribution: Demonstrated that language models trained at sufficient scale exhibit zero-shot task performance. GPT-2 could generate coherent multi-paragraph text, summarize articles, and perform rudimentary translation without task-specific fine-tuning. OpenAI initially withheld the full model citing misuse concerns, sparking the first major debate about responsible release in the LLM era.

Organization: OpenAI. Parameters: 175 billion. Architecture: Decoder-only Transformer (96 layers, 96 heads, $d_{\text{model}} = 12288$). Training data: Approximately 300 billion tokens from Common Crawl, WebText2, Books, and Wikipedia.

Key contribution: In-context learning (ICL). GPT-3 showed that by simply placing examples in the prompt (few-shot), a frozen model could perform tasks it was never explicitly trained on. This was a qualitative shift: instead of fine-tuning a model per task, you write a prompt. The paper also established smooth scaling laws relating model size to performance.

Organization: OpenAI. Base model lineage: GPT-3.5, itself a fine-tuned descendant of InstructGPT (Ouyang et al. 2022), which was in turn built on GPT-3. The training pipeline was: (1) supervised fine-tuning on human-written demonstrations, (2) reward model training on human preference comparisons, (3) PPO optimization against the reward model. ChatGPT is not a simple fine-tune of base GPT-3; it is the end product of this multi-step alignment pipeline. Key innovation: Reinforcement Learning from Human Feedback (RLHF).

Key contribution: The consumer product that brought LLMs to mainstream awareness. The RLHF pipeline (drawing on policy gradient methods) made the model substantially more helpful, harmless, and conversational than the base GPT-3.5.

Organization: OpenAI. Parameters: Not officially disclosed. Architecture: Not officially disclosed.

Architecture rumors (not confirmed). External reporting (SemiAnalysis and other secondary sources) has described GPT-4 as a mixture-of-experts Transformer in roughly the 1.8 trillion total-parameter range with a small number of active experts per token. OpenAI has never confirmed these numbers and they should be treated as unverified leaks, not factual architecture details.

Key contribution: Multimodal input (text and images). Substantially improved reasoning, factuality, and instruction following over GPT-3.5. Reported to score in the 90th percentile on the bar exam; this figure has been contested. Martinez (2024, arXiv:2308.07312, "Re-Evaluating GPT-4's Bar Exam Performance") reanalyzed the result and placed the actual percentile at roughly 48-68th, depending on the testing cohort and exam form. The claim should be read as "strong performance on legal reasoning tasks," not as a settled empirical benchmark. Demonstrated that scaling plus RLHF plus careful data curation can produce models with broad expert-level knowledge across many domains.

Organization: Anthropic. Models (2023-2024): Claude 1 (March 2023), Claude 2 (July 2023), Claude 3 family (March 2024: Haiku, Sonnet, Opus), Claude 3.5 Sonnet (June 2024), Claude 3.5 Sonnet upgraded with computer use (October 2024).

Key innovation: Constitutional AI (CAI). Instead of relying solely on human preference labels for alignment, CAI uses a set of written principles (a "constitution") to guide the model. The model critiques and revises its own outputs according to these principles, then trains on the revised outputs. This makes the alignment process more transparent and scalable than pure RLHF.

Computer use milestone (October 2024): Claude 3.5 Sonnet became the first frontier model to support computer use in public beta: the ability to interpret screenshots and generate mouse/keyboard actions to control desktop software. This marked the beginning of agentic computer interaction as a model capability, distinct from text-only tool use.

Anthropic has emphasized safety research alongside capability, including work on interpretability (circuit analysis, dictionary learning), honest calibration, and refusal behavior.

Organization: Meta. Models: Llama 1 (Feb 2023, 7B-65B), Llama 2 (July 2023, 7B-70B), Llama 3 (April 2024, 8B and 70B), Llama 3.1 (July 2024, 8B-405B), Llama 4 Scout and Maverick (April 5, 2025).

Key contribution: Established the open-weight model ecosystem. Llama 1 showed that smaller models trained on more tokens could match larger models (the Chinchilla scaling insight applied aggressively). Llama 2 added RLHF and a permissive license. Llama 3 pushed quality to near-frontier levels. The 405B Llama 3.1 model demonstrated that open-weight models could compete with proprietary systems on many benchmarks.

Llama 4 (April 2025): The first Llama generation to use Mixture-of-Experts and native multimodal pretraining. Scout (109B total, 17B active, 16 experts, 10M context window) fits on a single H100 GPU with int4 quantization. Maverick (approximately 400B total, 17B active, 128 experts, 1M context window for instruct variant) targets frontier-level performance. Both models were pretrained on interleaved text, images, and video as first-class inputs.

The Llama family spawned an enormous ecosystem of fine-tunes, quantizations, and derivative models (Alpaca, Vicuna, Code Llama, and many others).

Organization: Alibaba Cloud. Models: Qwen 1 (2023), Qwen 1.5 (2024), Qwen 2 (June 2024), Qwen 2.5 (September 2024), Qwen 3 (April 28, 2025).

Key contribution: Strong multilingual performance across 119 languages, particularly in Chinese and English. Qwen models are open-weight with Apache 2.0 licensing.

Qwen 3 (April 2025): Introduced hybrid thinking modes (switching between chain-of-thought reasoning and direct response). Dense variants from 0.6B to 32B; MoE variants including Qwen3-30B-A3B (30B total, 3B active) and Qwen3-235B-A22B (235B total, 22B active). Also released a proprietary Qwen3-Max with over 1 trillion parameters, available via API. Trained on 36 trillion tokens. Demonstrated competitive performance across coding, mathematics, and general language tasks. The Qwen-VL series added vision capabilities.

Organization: DeepSeek AI (China). Key models: DeepSeek-V2 (May 2024, introduced MLA), DeepSeek-V3 (December 2024, 671B total, 37B active), DeepSeek-R1 (January 2025), DeepSeek-V3.1 (August 19, 2025).

Key innovations:

• Mixture-of-Experts (MoE): DeepSeek-V3 uses MoE with 671 billion total parameters but only 37 billion active per token, achieving frontier performance at dramatically lower inference cost.
• Multi-head Latent Attention (MLA): Introduced in V2, this KV-cache compression technique projects keys and values into a low-dimensional latent space, reducing memory requirements for long sequences.
• Reasoning models: DeepSeek-R1 demonstrated that chain-of-thought reasoning can emerge from reinforcement learning on reasoning tasks, achieving competitive performance with OpenAI o1 on math and coding benchmarks. Released open-weight, enabling community study of reasoning model internals.
• Hybrid reasoning (V3.1): DeepSeek-V3.1 (August 2025) introduced a hybrid architecture that combines thinking and non-thinking modes in a single model, switchable via the chat template. Maintains the 671B/37B MoE structure with 128K context.

DeepSeek demonstrated that architectural innovation (MoE, MLA) and RL-based training can compensate for having less compute than the largest Western labs.

Organization: Google DeepMind. Models: Gemini 1.0 (December 2023), Gemini 1.5 (February 2024), Gemini 2.0 (December 2024), Gemini 2.5 Pro (March 25, 2025; GA June 2025), Gemini 3 Pro (November 18, 2025).

Key innovations:

• Natively multimodal: Trained from the ground up on interleaved text, images, audio, and video, rather than bolting vision onto a text model.
• Long context: Gemini 1.5 Pro demonstrated reliable processing of up to 1 million tokens (later extended to 2 million), enabling processing of entire codebases, hours of video, or thousands of pages of documents in a single context window.
• Mixture-of-Experts: Gemini 1.5 uses a MoE architecture for efficiency.
• Reasoning model era (Gemini 2.5): Gemini 2.5 Pro was Google's first explicitly designated thinking model, using chain-of-thought reasoning before responding. Led benchmarks on math, coding, and multimodal tasks at time of release.
• Gemini 3 (November 2025): Released simultaneously across the Gemini app, Google AI Studio, and Vertex AI. Gemini 3 Pro and Gemini 3 Deep Think were the flagship variants. First time Google launched a new Gemini model across multiple products on day one.

Organization: Moonshot AI (China). Key models: Kimi k1.5 (2024-2025), Kimi K2 (July 11, 2025), Kimi K2.5 (January 27, 2026).

Kimi k1.5 (2024-2025): Pioneered extremely long context windows for consumer applications. Demonstrated that long context is not just a benchmark metric but enables qualitatively different use patterns: processing entire books, long legal documents, and extended conversation histories. Extended RL-based reasoning in a long-context setting.

Kimi K2 (July 2025): Open-weight release (Apache 2.0). Strong performance on coding benchmarks. Instruct update in September 2025 doubled context window to 256K and improved agentic coding. K2 Thinking (November 2025) added dedicated chain-of-thought reasoning capability.

Kimi K2.5 (January 2026): Native multimodal architecture supporting visual and text input, thinking and non-thinking modes, dialogue and agent tasks. 1 trillion parameters. Competitive on agent, coding, image, and video tasks.

Organization: Anthropic.

Claude Opus 4 and Sonnet 4 (May 22, 2025): The Claude 4 generation introduced hybrid models offering both near-instant responses and extended thinking modes for deeper reasoning in a single model. Opus 4 focused on sustained performance on complex, long-running agentic and coding tasks. Sonnet 4 provided superior coding and reasoning relative to Sonnet 3.7.

Claude Opus 4.1 (August 5, 2025): Incremental improvement on agentic tasks, real-world coding, and reasoning. Achieved 74.5% on SWE-bench Verified. Extended thinking capability expanded to 64,000 tokens.

Claude Sonnet 4.5 (September 29, 2025): Highest-intelligence Sonnet at its price point. Added 1M token context in beta. Best-in-class coding performance at the Sonnet tier.

Claude Opus 4.5 (November 24, 2025): Next full Opus release, continuing Anthropic's focus on agentic workflows and extended reasoning.

Claude Opus 4.7 (April 16, 2026): Current generally-available Opus tier. Continued improvements in reasoning, coding, and agentic task performance.

Claude Mythos Preview (April 7, 2026): Anthropic's most capable frontier model to date, publicly documented via the Claude Mythos Preview System Card (April 7, 2026). Per the card, Mythos Preview's large increase in capabilities led Anthropic to not make it generally available; instead it is being used as part of a defensive cybersecurity program with a limited set of partners. Findings from the Mythos System Card (RSP evaluations, cyber, alignment assessment, model-welfare assessment) are being used to inform future Claude releases and their safeguards. Not a public deployment, but a publicly documented capability datapoint.

Organization: OpenAI. Release: August 7, 2025, via livestream event. Available in ChatGPT, Microsoft Copilot, and the OpenAI API.

Key contribution: Highest published scores at release on benchmarks testing mathematics, programming, finance, and multimodal understanding. Improvements over GPT-4o in faster response times, coding and writing quality, accuracy on health-related questions, and reduced hallucination rates. Publicly accessible across tiers.

Organization: OpenAI. Release: April 23, 2026 in ChatGPT for Plus, Pro, Business, and Enterprise tiers and in Codex; API rollout (gpt-5.5 and gpt-5.5-pro) on April 24, 2026. Internal codename "Spud."

Key contribution: OpenAI's most capable mainline model at release, with emphasis on coding, debugging, online research, document and spreadsheet creation, and multi-tool agentic workflows. OpenAI states GPT-5.5 matches GPT-5.4 per-token serving latency while using fewer tokens to complete the same tasks, so the headline change is token efficiency at the same latency budget rather than a new latency floor. Continues the GPT-5.x trend of folding specialist capabilities (reasoning, Codex-class coding, computer use, tool search) into the mainline product surface.

Pricing at release: gpt-5.5 at 5 USD per million input tokens and 30 USD per million output tokens. gpt-5.5-pro at 30 USD / 180 USD per million input / output tokens, restricted to Pro, Business, and Enterprise.

What remains undisclosed. Parameter count, dense-vs-MoE architecture, training data composition, and training compute. OpenAI publishes a system card focused on capability and safety evaluations, not architectural disclosure.

Organization: xAI (Elon Musk). Models: Grok 1 (2023), Grok 2 (August 2024), Grok 3 (February 17, 2025), Grok 4 (July 9, 2025), Grok 4.1 (November 17, 2025).

Key contribution: Grok 3 achieved frontier-level performance on reasoning and coding benchmarks, established xAI as a serious frontier lab. Grok 4 and 4 Heavy released July 2025 with claimed benchmark outperformance over rivals. Real-time integration with X (Twitter) platform provides live information access distinct from other frontier models.

Organization: Mistral AI (France). European frontier lab producing both open-weight and proprietary models.

2025 releases: Mistral Large 24.11 (November 2024, 123B dense), Codestral 25.01 (January 2025, code-focused, led Lmsys Copilot Arena at launch), Mistral Small 3 (January 2025), Codestral (August 2025 refresh), Mistral 3 family (December 2025: full open-weight frontier model plus nine smaller offline-capable variants).

Key contribution: Demonstrated that efficient dense models (not requiring MoE) at 7B-123B scale can be competitive with much larger models when trained on high-quality data. Sliding window attention and grouped-query attention introduced in early Mistral 7B remain widely used. Mistral 3 positioned as a serious open-weight European alternative to US and Chinese MoE giants.

Organization: Zhipu AI (China; Z.ai). Release: February 11, 2026. Parameters: 744B total, 40B active (MoE, vs GLM-4.5 which was 355B/32B). Training data: 28.5T tokens (vs 23T for GLM-4.5).

Release date: February 11, 2026, open-source on Hugging Face; the parameter table below lists GLM-5 under year 2026 accordingly.

Key contribution: Native multimodal, 128K context, designed for agentic tasks. At release approached Claude Opus 4.5 on coding benchmarks and surpassed Gemini 3 Pro on some benchmarks. First major model release from Zhipu after the company went public. Established GLM series as a serious frontier competitor from the Chinese open ecosystem.