AI Release Index
Transformer Base
The field gains a parallelizable architecture that scales with compute and data.
GPT-1
One pretrained model begins to replace task-specific language systems.
BERT
Pretrained encoders become standard infrastructure for language understanding.
GPT-2
Language generation moves from benchmark artifact to public capability debate.
T5
Task interfaces collapse into a single generative abstraction.
GPT-3
In-context learning changes how developers adapt models to new tasks.
DALL·E
Natural language becomes a viable interface for image construction.
CLIP
A shared representation unlocks zero-shot vision and prompt-guided image systems.
LaMDA
Conversational quality is separated into measurable post-training objectives.
Codex
Model-assisted coding moves into daily software development.
Jurassic-1
Developers gain another route to frontier-scale text generation.
Chinchilla
Training budgets move from parameter count toward compute-optimal data allocation.
PaLM
Large-cluster training demonstrates another route up the scaling curve.
DALL·E 2
High-fidelity generation turns prompt writing into a creative practice.
Stable Diffusion 1
Generative imaging becomes inspectable, modifiable, and locally deployable.
Whisper
High-quality multilingual transcription becomes commodity infrastructure.
ChatGPT
A research trajectory becomes a global software platform in one release.
LLaMA 1
Frontier-adjacent base models move onto independent hardware and research stacks.
Claude 1
A distinct alignment method becomes visible through a production model.
GPT-4
General-purpose model performance reaches professional knowledge-work thresholds.
Falcon 40B
Commercially usable weights widen the set of deployable base models.
Genie 1
World modeling connects generative video with learned interaction.
Claude 2
Document-scale prompting becomes a standard frontier feature.
LLaMA 2
A licensed model family standardizes the open deployment market.
SDXL 1.0
Open image tooling consolidates around a higher-resolution base model.
Mistral 7B
Useful model quality moves into a single-GPU deployment envelope.
DALL·E 3
Image prompting becomes a conversational, model-mediated workflow.
Qwen 72B
High-capacity bilingual weights broaden the geography of open-model development.
Gemini 1.0
A major platform standardizes around a natively multimodal model family.
Mixtral 8×7B
Sparse routing becomes a practical open-weight efficiency strategy.
Gemini 1.5 Pro
Whole repositories and hours of media fit inside a single model call.
Claude 3 Opus
Model tiering becomes a deliberate systems-design choice.
Grok-1
A new frontier lab places a very large sparse model in public hands.
DBRX
The data-platform stack extends directly into foundation-model production.
Command R+
Retrieval-grounded generation becomes a first-order model capability.
Llama 3
The open ecosystem converges on a stronger common base family.
Phi-3 Mini
Useful reasoning quality enters constrained local hardware.
GPT-4o
Multimodal interaction moves from batch composition to live model behavior.
Kling 1.0
High-duration video synthesis moves beyond closed research previews.
Qwen2 72B
A strong multilingual family reaches more local and regional deployments.
Gen-3 Alpha
Generative video becomes part of a repeatable editing workflow.
Claude 3.5 Sonnet
Generated output moves from chat text into persistent interactive artifacts.
Mistral Large 2
Large multilingual capability remains available for self-hosted inference.
FLUX.1
Open image generation reorganizes around a new architecture and vendor.
Grok 2
Frontier model deployment connects directly to a real-time social graph.
OpenAI o1
Scaling shifts from pretraining alone to deliberate inference-time computation.
Qwen2.5 72B
One open family spans a wide set of deployment budgets and languages.
Llama 3.2
Multimodal weights arrive in the most widely deployed open family.
GLM-4-Plus
A major regional model stack integrates the full assistant toolchain.
Amazon Nova Pro
The largest cloud platforms now each operate a native frontier model family.
DeepSeek-V3
Frontier-adjacent quality arrives with an unusually low reported training bill.
DeepSeek-R1
Inference-time reasoning becomes inspectable, distillable, and locally deployable.
Claude 3.7 Sonnet
Reasoning depth becomes an adjustable parameter inside one model.
Gemma 3
A capable multimodal family brings long-context deployment to smaller local systems.
Gemini 2.5 Pro
Deliberate reasoning scales across both long context and native multimodality.
Llama 4 Maverick
The dominant open family changes architecture while preserving its deployment surface.
OpenAI o3
Tool-mediated reasoning becomes a unified production model behavior.
Qwen3 235B
Hybrid reasoning becomes available across a broad open checkpoint family.
Mistral Medium 3
Frontier-adjacent performance moves into a less expensive enterprise serving tier.
Veo 3
Text-to-video begins to produce a complete audiovisual clip in one pass.
Devstral
Agentic coding behavior becomes available in a deployable open model.
Qwen VLo
A single model begins to treat visual perception and creation as one workflow.
Grok 4
xAI establishes a new flagship generation for reasoning and tool-use workloads.
Voxtral
High-quality speech understanding gains an open-weight deployment path.
Qwen3-Coder
Repository-scale coding agents gain a strong open model and first-party CLI.
Codestral 25.08
Code completion and agent workflows consolidate around one enterprise model stack.
GPT-5
Model routing hides capability-tier selection inside a single assistant surface.
Grok Code Fast 1
Coding agents get a dedicated low-latency model rather than a general flagship.
GPT-5.1
Reasoning effort adapts more directly to task difficulty and latency needs.
Grok 4.1
xAI tightens flagship instruction following, reasoning, and interaction quality.
Gemini 3 Pro
A new Gemini generation arrives across Search, the app, APIs, and agent tooling at once.
GPT-5.1-Codex-Max
Coding models move from short edits toward durable, long-horizon execution.
DeepSeek-V3.2
One openly available family spans direct answers and deliberate reasoning.
Mistral 3
Mistral renews the open deployment path across multiple capability sizes.
Devstral 2
Open coding agents get a second-generation model and native operating harness.
GPT-5.2
The flagship is optimized around complete professional deliverables, not isolated answers.
Gemini 3 Flash
The fast tier surpasses the prior generation's larger model on many workloads.
Mistral OCR 3
Document parsing matures into a dedicated production model category.
Claude Opus 4.6
Agent work shifts from short tool loops toward durable multi-hour execution.
GPT-5.3-Codex-Spark
Model speed crosses a thousand tokens per second and changes the feel of coding agents.
Claude Sonnet 4.6
Stronger agent performance arrives in Anthropic's broadly deployed Sonnet tier.
Gemini 3.1 Pro
A mid-generation update more than doubles Google's reported ARC-AGI-2 result.
GPT-5.4
Computer operation becomes native to a general-purpose frontier model.
Mistral Small 4
One Apache-licensed model unifies multimodality, reasoning, and coding agents.
GPT-5.4 mini / nano
Smaller models become capable enough to act as fast coding and computer-use subagents.
Voxtral TTS
Production voice-agent speech gains a fast open-weight model option.
Grok 4.20
The Grok family continues its rapid cadence toward tool-mediated work.
Muse Spark
Visual reasoning, tools, and multi-agent orchestration are trained into one Meta model.
Claude Opus 4.7
Frontier coding agents become more consistent about verifying their own work.
GPT-5.5
General model progress is measured increasingly by complete, economically valuable work.
DeepSeek-V4 Pro / Flash
DeepSeek formalizes separate high-capability and high-throughput production tiers.
Gemini 3.5 Flash
A fast model reaches flagship-class agent performance while serving at scale.
Gemini Omni Flash
Video generation becomes an editable multimodal conversation rather than a one-shot prompt.
Mistral Medium 3.5
A production model ships alongside the remote coding and work agents built to use it.
Claude Opus 4.8
Browser and computer agents become reliable enough for codebase-scale workflows.
Gemma 4 12B
A current-generation Google model remains small enough for practical local deployment.
Claude Fable 5
Anthropic expands beyond the familiar Haiku, Sonnet, and Opus tier structure.
Claude Mythos 5
A new top tier shifts Anthropic's frontier beyond the Opus naming line.
Mistral OCR 4
Reliable document extraction becomes core infrastructure for enterprise agents.
Claude Sonnet 5
Opus-class agent behavior becomes available at the lower-cost Sonnet tier.
Robostral Navigate
Mistral extends its model portfolio from digital agents into embodied navigation.
GPT-5.6 Sol / Terra / Luna
A flagship generation spans frontier, balanced, and low-cost tiers under one release.
Grok 4.5
xAI keeps pace in the shift from chat models to general-purpose agents.
Claude Opus 5
Anthropic's strongest broadly deployed model becomes a more efficient everyday workhorse.
Gemini 3.6 Flash
The high-throughput Gemini tier continues to absorb frontier reasoning capability.
Shieldstral
One small open model can adapt moderation decisions to policies supplied at inference time.
GPT-5.6-Cyber
Advanced cyber capability moves into a purpose-built model with controlled access for defenders.