Skip to content
Shared AI Researchan open archive

Model tracking45 releases

Frontier and notable model releases

A running record of the models that defined the field: who released them, when, under what access terms, and why they mattered.

Data compiled as of July 2026. Corrections by pull request are welcome.

45

2026

ModelDeveloperReleasedAccessContextSignificance
Claude Opus 5AnthropicJul 2026Proprietary1MOpus-tier Claude 5 release; close to Fable 5 on agentic coding at half the price, with an optional faster “fast mode”.
Kimi K3Moonshot AIJul 2026Open weights1M2.8T-parameter MoE, the largest open-weight release to date; weights published July 27 under a bespoke (non-OSI) license.
GPT-5.6 (Sol, Terra, Luna)OpenAIJul 2026Proprietary1MThree-tier family distilled from a single base run; replaced GPT-5.5 across ChatGPT and the API.
Grok 4.5xAI (SpaceX)Jul 2026Proprietary500KFirst Grok built specifically for agentic coding; first major release since xAI folded into SpaceX in February 2026.
Claude Sonnet 5AnthropicJun 2026Proprietary1MDefault mid-tier Claude 5 model; near-Opus-4.8 performance at lower prices, and the new free-tier model on Claude.ai.
GLM-5.2Zhipu AI (Z.ai)Jun 2026Open weights1M753B MoE under an MIT license; at release the strongest open-weight coding model, at a fraction of frontier API prices.
Claude Fable 5 / Mythos 5AnthropicJun 2026Proprietary1MFirst public Mythos-class model, a tier above Opus; access was suspended for most of June under a U.S. export-control order before being restored; the less-restricted Mythos 5 variant is limited to approved organizations.
Qwen3.7-MaxAlibabaMay 2026Proprietary1MAPI-only flagship built for long-horizon agent runs; marked Alibaba’s shift away from open weights for its top tier.
Gemini 3.5 FlashGoogle DeepMindMay 2026Proprietary—I/O 2026 workhorse that beat Gemini 3.1 Pro on coding and agentic benchmarks; shipped while the 3.5 Pro flagship was repeatedly delayed.
DeepSeek V4 (Pro, Flash)DeepSeekApr 2026Open weights1MMIT-licensed MoE pair (V4-Pro 1.6T/49B active, V4-Flash 284B); previewed in April, stable in July; its thinking mode absorbed the R1 reasoning line.
GPT-5.5OpenAIApr 2026Proprietary—Agent-focused flagship pitched as a step toward an AI “super app”; GPT-5.5 Instant became the ChatGPT default in May.
Muse SparkMetaApr 2026Proprietary—First model from Meta Superintelligence Labs and Meta’s first closed-weight flagship, ending its open-release strategy.
GPT-5.4OpenAIMar 2026Proprietary—Mainline GPT-5 refresh on OpenAI’s quickened release cadence; its xHigh reasoning tier led agentic tool-use benchmarks.
Gemini 3.1 ProGoogle DeepMindFeb 2026Proprietary1MIncremental Pro flagship update; native multimodal input and output across text, image, audio, video, and code.
Claude Opus 4.6–4.8AnthropicFeb 2026Proprietary—Rapid-cadence Opus point releases (4.6 in February, 4.7 in April, 4.8 in May) bridging the fourth generation to Claude 5.

2025

ModelDeveloperReleasedAccessContextSignificance
GPT-5.2OpenAIDec 2025Proprietary400KInstant, Thinking, and Pro variants; OpenAI’s rapid answer to Gemini 3, with stronger long-context and enterprise coding.
Mistral Large 3Mistral AIDec 2025Open weights256K675B MoE (41B active) multimodal flagship under Apache 2.0; Europe’s strongest open-weight release.
Claude Opus 4.5AnthropicNov 2025Proprietary200KFlagship coding and agentic model; substantial price cut from prior Opus tiers.
Gemini 3 ProGoogle DeepMindNov 2025Proprietary1MLed most reasoning and multimodal benchmarks at launch; deep integration into Search.
GPT-5.1OpenAINov 2025Proprietary400KRefinement of GPT-5 with adaptive reasoning effort and improved instruction following.
Claude Sonnet 4.5AnthropicSep 2025Proprietary200K–1MMid-tier flagship focused on long-horizon agentic coding; 1M-token context in beta.
GPT-5OpenAIAug 2025Proprietary400KUnified router across fast and reasoning modes; replaced the GPT-4 line in ChatGPT.
gpt-oss-120b / 20bOpenAIAug 2025Open weights—OpenAI’s first open-weight release since GPT-2; Apache-2.0 licensed MoE reasoning models.
Grok 4xAIJul 2025Proprietary256KReasoning-first flagship trained on the Colossus cluster; strong benchmark results.
Kimi K2Moonshot AIJul 2025Open weights—1T-parameter MoE (32B active); at release the strongest open agentic/coding model.
Claude Opus 4 / Sonnet 4AnthropicMay 2025Proprietary200KFourth-generation flagships; extended thinking with tool use during reasoning.
Qwen 3AlibabaApr 2025Open weights—Dense and MoE family (0.6B–235B) with switchable thinking mode; broadly adopted base for fine-tunes.
Llama 4 (Scout, Maverick)MetaApr 2025Open weightsup to 10M (Scout)Natively multimodal MoE family; mixed reception relative to open-weight competitors.
o3 / o4-miniOpenAIApr 2025Proprietary200KReasoning models with full tool use during chain of thought; o3 topped many benchmarks at launch.
Gemini 2.5 ProGoogle DeepMindMar 2025Proprietary1MThinking-by-default flagship; long-context multimodal reasoning at competitive prices.
Claude 3.7 SonnetAnthropicFeb 2025Proprietary200KFirst hybrid reasoning model: a single model with a controllable extended-thinking budget.
DeepSeek-R1DeepSeekJan 2025Open weights128KOpen reasoning model rivaling o1 at a much lower reported training cost; triggered a market-wide repricing of AI capex assumptions.

2024

ModelDeveloperReleasedAccessContextSignificance
DeepSeek-V3DeepSeekDec 2024Open weights128K671B MoE (37B active); the reported ~$5.6M compute cost covered only the final training run and is widely considered to understate total cost; basis for R1.
o1OpenAIDec 2024Proprietary200KFirst production reasoning model line (previewed September 2024); established test-time compute scaling.
Gemini 2.0 FlashGoogle DeepMindDec 2024Proprietary1MFast agentic workhorse model; native tool use and multimodal output.
Claude 3.5 SonnetAnthropicJun 2024Proprietary200KOutperformed the larger Opus tier; the October update added computer use — the first frontier agent able to operate a GUI.
GPT-4oOpenAIMay 2024Proprietary128KNatively multimodal (“omni”) flagship with real-time voice; free-tier default in ChatGPT.
Llama 3 / 3.1MetaApr 2024Open weights128K (3.1)Llama 3.1 405B (July 2024) was the first open-weight model at rough parity with frontier proprietary models.
Claude 3 (Opus, Sonnet, Haiku)AnthropicMar 2024Proprietary200KThree-tier family; Opus was the first model to clearly match GPT-4.
Gemini 1.5 ProGoogle DeepMindFeb 2024Proprietary1M–2MBroke the long-context barrier with near-perfect million-token recall.

2023

ModelDeveloperReleasedAccessContextSignificance
Mixtral 8x7BMistral AIDec 2023Open weights32KSparse MoE that beat much larger dense models; mainstreamed MoE in open models.
GPT-4 TurboOpenAINov 2023Proprietary128KCheaper, faster GPT-4 with 128K context; launched alongside the GPT Store and Assistants API.
Llama 2MetaJul 2023Open weights4KFirst openly licensed Llama for commercial use; seeded the open-model ecosystem.
GPT-4OpenAIMar 2023Proprietary8K–32KDefined the frontier for over a year; passed professional exams that stumped GPT-3.5.

2022

ModelDeveloperReleasedAccessContextSignificance
ChatGPT (GPT-3.5)OpenAINov 2022Proprietary4KThe consumer breakout: 100M users in two months; started the current investment cycle.