AI Model Timeline
2025 – 2026

Largest open-source model ever released at 2.8 trillion parameters with 1M-token context. API launched July 16; full weights dropped July 27. Rivals top US frontier systems on coding and reasoning benchmarks.

OpenAI's latest GPT-5 generation released as three named variants — Sol (speed), Luna (balanced), Terra (deep reasoning) — continuing the unified-router architecture.

Major Grok upgrade with tripled parameter count, targeting coding leadership. Strong SWE-bench performance and expanded agentic capabilities.

Broadly available Claude 5 model combining the next-generation architecture with the speed and cost profile developers expect from the Sonnet tier.

First models in the new Claude 5 Mythos-class tier. Mythos 5 is limited to approved organizations. Fable 5 marks a new generation of long-context reasoning architecture.

Open-weight coding model under MIT license with 1M-token context. Part of Z.ai's GLM Coding Plan — NIST conducted a CAISI safety assessment at release, signaling growing international recognition of Chinese open-weight labs.

Coding-specialized iteration of the K2 line, competing in the rapidly crowding open-weight coding model space against Qwen3-Coder and DeepSeek V4-Flash.

Final major iteration in the Opus 4.x line before the Claude 5 generation. Delivered Anthropic's highest benchmark scores to that point.

Frontier-class multimodal model optimized for agentic and coding use cases. Significant step up from Medium 3.1 in both capability and vision understanding.

Two open-weight models launched simultaneously — Pro and Flash — on Hugging Face, API, and chat. 1M context, full tool use, and API pricing well below US competitors.

Added native computer-use capabilities — browser control, form-filling, and workflow execution. Marks GPT-5's evolution into an agentic operating model.

Continued Anthropic's fast iteration on the Opus tier with enhanced performance on complex agentic and multi-turn tasks.

Open-sourced two months after initial release, part of Z.ai's strategy of staged openness. Continued the GLM-5 generation's improvements in multilingual instruction following and long-context reasoning.

Iterative upgrade in the K2 line with strengthened multi-step agentic reasoning. Moonshot shipped incremental versions rapidly in early 2026 ahead of the K3 generation.

Next-generation small model continuing Mistral's shift to permissive open licensing. Efficient on modest hardware with strong multilingual and instruction-following performance.

Significant capability upgrade continuing OpenAI's rapid iteration cadence on the GPT-5 line, with improvements to instruction following and tool use reliability.

1 trillion parameter MoE with 32B active parameters, adding multimodal vision-language understanding and advanced agentic capabilities over K2. Maintained the 1M-token context window from its predecessor.

397B-A17B MoE open-weight model building on Qwen3's hybrid-reasoning architecture. Also released a closed proprietary Qwen3.5-Plus variant for enterprise deployments.

Next-generation launch from Z.ai (formerly Zhipu AI), introducing a new capability tier with improved multilingual reasoning, coding, and long-context performance across Chinese and English.

Sonnet 4.6 became the default free model on Claude.ai. Competitive benchmark results across coding, analysis, and long-context tasks.

Sparse MoE model with 675B total parameters under Apache 2.0 — a major licensing shift from Mistral's earlier restrictive terms. 256K context, frontier-class performance on coding and reasoning.

Further refinement of the GPT-5 line with stronger reasoning consistency and expanded multimodal capabilities.

Final major release in the GLM-4.x line before the GLM-5 generation. Refined multilingual instruction-following and reasoning, maintaining Z.ai's rapid quarterly release cadence.

Reasoning variant of K2 adding extended chain-of-thought via a $4.6M training run. Delivered top-tier math and coding results in the open-weight class.

Grok 4 iteration released broadly to all grok.com users with expanded tool use, improved coding, and tighter X platform integration.

Enhanced Opus with stronger multi-step reasoning and improved agent capabilities for complex, long-horizon tasks.

Point release to GPT-5 with improved instruction-following, reduced hallucination rate, and faster inference under the unified model router.

Compact instruction-tuned SLM for on-device deployment. Optimized for latency-sensitive and privacy-first enterprise scenarios.

Fastest and most cost-efficient Claude model, targeting high-throughput production use cases and real-time applications.

48B MoE with only 3B active parameters using Kimi Delta Attention (KDA) — a novel linear attention mechanism that significantly reduces memory usage and improves generation throughput versus standard attention.

Experimental checkpoint showing significant gains in coding and math. Later V3.2-Speciale won gold at IMO, IOI, and ICPC 2026.

Iterative improvement in the GLM-4 family with stronger Chinese-English bilingual instruction following. Z.ai maintained a brisk sub-quarterly release cadence throughout 2025.

First DeepSeek model with dual-mode functionality — thinking and non-thinking — within a single model endpoint, matching a pattern set by Qwen3.

Incremental improvement to the Opus line with enhanced multi-step agentic task handling and better instruction-following across long documents.

OpenAI unified its model lines into a single GPT-5 endpoint with an intelligent router that decides how much to think per request. Paired a fast response mode with a deep reasoning mode seamlessly.

xAI's fourth-generation flagship with enhanced reasoning and multimodal understanding. Trained on expanded Colossus cluster capacity.

Massive trillion-parameter MoE open-weight model trained on 15.5T tokens. Released under a modified MIT license with 1M context, strong agentic and coding capability.

Alibaba's flagship agentic coding model — 480B-parameter MoE with 35B active, native 256K context extrapolating to 1M. Open-source release set a new bar for open coding models, challenging DeepSeek Coder V3 and Kimi K2 on SWE-bench.

First 2025 GLM release from Zhipu AI (later rebranded Z.ai), continuing their strong bilingual Chinese-English instruction model line. Competitive with Western mid-tier models on multilingual and reasoning tasks.

Full production rollout of the 2.5 family. Flash became the default Gemini model at Google I/O; Flash-Lite optimized for cost-sensitive high-volume workloads.

Significant reasoning upgrade to the R1 family, achieving 97.3% on MATH-500. Continued DeepSeek's strategy of rapid open-weights iteration.

Positioned between Small and Large tiers, offering GPT-4-class performance with significantly reduced inference cost for enterprise deployments.

Claude 4 generation launch. Sonnet 4 replaced 3.7 as the default API model; Opus 4 targeted complex long-context and multi-step agent workflows.

Small reasoning models outperforming models 5–10× larger on STEM tasks. Proved that distilled reasoning capability scales to sub-10B parameter models.

Six dense + two MoE models, all fully open-sourced at launch. Hybrid reasoning mode (thinking/non-thinking) within a single model. The 235B-A22B MoE set new open-weight benchmarks.

o3 delivers frontier reasoning; o4-mini brings reasoning at speed and lower cost. Both support tool use natively within the reasoning chain.

Meta's natively multimodal MoE flagship. Scout (17B active / 109B total) ships with a record 10M-token context; Maverick (17B active / 400B total) targets instruction and coding tasks.

Google's most capable reasoning model to date at launch. Topped multiple coding and math benchmarks with native thinking mode and 1M-token context.

Four-size family of open-weight models with multimodal capabilities. Strong on-device performance; the 27B model rivaled proprietary models from 2024.

Alibaba's reasoning-focused open-weight model that matched DeepSeek-R1 at 1/21 the parameter count. Apache 2.0, fits on a single 24GB GPU with quantization — made frontier-class reasoning accessible to consumer hardware.

Small-but-mighty SLMs punching above their weight class. Phi-4-multimodal integrates text, vision, and speech into a single efficient model.

xAI's flagship model trained on the Colossus supercluster. Included DeepSearch web-grounding and Think mode for step-by-step reasoning.

Anthropic's first hybrid-reasoning model with extended-thinking capability. Customers can toggle between standard and deep-reasoning modes on a single endpoint.

Moonshot AI's multimodal long-thinking model released the same day as DeepSeek R1. Used reinforcement learning to achieve strong reasoning performance in a frontier API offering with 128K context.

Shocked the industry — GPT-o1-level reasoning at a fraction of the cost. MIT license. Sparked the open-weight reasoning race and triggered a global AI policy reaction.

Compact reasoning model offering o1-class performance at lower cost and latency. Available across API tiers and ChatGPT Plus/Pro.

Small, efficient instruction model released under Apache 2.0. Strong multilingual performance optimized for edge deployments.