AI Model Timeline

OpenAI's GPT-6 generation flagship, launched 2026-09-03 in a phased rollout: participants in OpenAI's cybersecurity access program first, then ChatGPT Plus, Pro, Business and Enterprise, the API (model id gpt-6-astra), Azure and AWS Bedrock. 1.05M-token context window, 128K max output, knowledge cutoff 2026-04-30, reasoning effort selectable from none to max. First OpenAI model classified 'Critical' for cybersecurity under the Preparedness Framework: the general release ships with offensive-cyber behaviour constrained, and a fuller cyber variant is limited to vetted defenders. Priced at $10/M input and $50/M output ($1.00 cached input, $12.50 cache write); requests above 272K input tokens are billed at 2x input and 1.5x output for the whole request, and a Fast mode costs 2x. OpenAI's launch table reports Terminal-Bench 4.0, Terminal-Bench-Science, OSWorld 2.0, FrontierMath Tier 4, DeepSWE v1.1 and HLE with tools, but no SWE-bench Pro, Terminal-Bench 2.x, MMLU-Pro or no-tools HLE figure, so the tracked fields here come from independent runs (Artificial Analysis, Vals AI, ARC Prize) where one exists.

Cybersecurity-specialised variant of Gemini 3.8 Flash and successor to Gemini 3.5 Flash Cyber, built to discover, validate and patch software vulnerabilities autonomously across codebases in 20 programming languages. Not a public model: access runs through Google's new Fairwind Program, which pairs the model with the CodeMender harness for government agencies and national cyber authorities, critical-infrastructure operators (healthcare, telecom, energy, finance), maintainers of widely used software platforms and vetted security partners, under operational requirements such as multi-factor authentication and access limited to internal security staff. Google says it will not be generally released. Vendor-reported, none of it a tracked academic benchmark: CyberGym (autonomous vulnerability discovery) above 3.5 Flash Cyber and larger frontier models, no figure published; a success rate above 70% on an internal 20-language vulnerability-discovery benchmark; CWE-Bench (Collinear) patching at 47.2% pass@1; and 2.6x more correct Chrome patches than the best commercial models in the Chrome Security team's own evaluation. No standard academic benchmarks were reported.

Successor to Claude Fable 5 in the Mythos-class tier, at the same $10/M input and $50/M output pricing but with cache reads cut to $0.25/M (2.5% of the input price instead of 10%). Launched alongside Claude Mythos 5.1, the same underlying model with safeguards lifted for Project Glasswing participants (tracked separately). Safety classifiers remain in place; refused requests can fall back server-side to Claude Opus 4.8 or Claude Opus 5. Adaptive thinking is always on, forced tool use is no longer supported, and all text output carries Anthropic's statistical watermark. Vendor-reported benchmarks: Terminal-Bench 4.0 55.8 (Fable 5: 42.0, Opus 5: 52.3), Terminal-Bench-Science 0.1 52.6, HLE 60.9 without tools (65.0 with tools), OSWorld 2.0 41.7 strict, CursorBench 3.2.0 73.4, AutomationBench 31.4.

World Labs' omni world model, pretrained from scratch to operate natively on text, images, video and 3D as a multimodal autoregressive diffusion transformer with every input in one shared spatial context. The headline capability is camera-controlled video, up to one minute at 1440p from one or more reference images, with the camera path supplied as a geometric input rather than described in a prompt. The same model reconstructs scenes into point clouds and 3D Gaussian splats, works with depth maps and camera poses, reframes multi-camera footage and supports parts of a real-to-sim robotics workflow. Early access for selected partners via a request form; no pricing, API or public date announced. World Labs reports that third-party human raters preferred Atlas over rival video models in 75 to 94 percent of pairwise trials depending on the competitor, a vendor-commissioned preference study recorded here as a note, not as a score.

Successor to Gemini Omni Flash (2026-06-30) for video generation and editing, released as the model id gemini-omni-1.1-flash. Takes any combination of text, image, audio and video as input and returns video. Scene extension now conditions on up to 10 seconds of preceding footage instead of the tail of the clip and continues in 10-second increments up to 40 seconds total; a first-and-last-frame control renders a single continuous shot between two supplied frames, and a video reference of up to 3 seconds carries character and style across generations. Native output resolution is 720p, with 1080p and 4K delivered as upscales, plus a new 360p draft mode that Google reports as up to 60% faster at a third of the 720p cost. Per-second API pricing: $0.03 (360p), $0.10 (720p), $0.15 (1080p), $0.30 (4K). Live via the Gemini API, Google AI Studio, Flow and the Gemini Enterprise Agent Platform; scene extension is rolling out to Google AI Plus, Pro and Ultra subscribers in the Gemini app globally. The predecessor's gemini-omni-flash-preview endpoint is scheduled for deprecation on 2026-09-30.

Post-training-only update to GLM-5.2: Z.ai leaves the 753B MoE base model unchanged and attributes the capability gains entirely to scaled post-training. Weights landed on Hugging Face on 2026-08-28, fourteen days after launch, once Z.ai's safety review completed — but under a bespoke GLM-5.3 License instead of the MIT license GLM-5.2 shipped under: use, modification, distribution, sublicensing, sale, deployment and fine-tuning are all permitted, but a company with more than $10B aggregate revenue over any 12 consecutive months must pass a Z.ai security review before hosting the model commercially. Individual users and smaller companies are unaffected. Z.ai reports its results on a new generation of agentic benchmarks rather than the axes tracked here (maximum thinking effort): Terminal-Bench 3.0 28.3 (GLM-5.2: 4.6), DeepSWE v1.1 66.9 (46.2), Agents' Last Exam 28.5 (23.8), CyberGym 84.5, plus AutomationBench, GDPVal-AA v2 and HLE with tools. Only Terminal-Bench 3.0 has a column here (its own axis, separate from the 2.x scale); the rest do not map onto the tracked benchmarks, and Z.ai published no GPQA, MMLU-Pro or SWE-bench Pro table.

Workhorse successor to Gemini 3.6 Flash, shipped three weeks after it. Google states the model was not trained from scratch — it replaces the predecessor via algorithmic improvements and user feedback. Same 1,048,576-token context window as 3.6 Flash, multimodal input (text, image, audio, video, files). Live through the Gemini API in AI Studio, Android Studio, Google Antigravity and the Gemini Enterprise Agent Platform. Vendor-reported coding deltas against 3.6 Flash: DeepSWE v1.1 49.0 to 65.3, FrontierCode 1.1 Main 34.4 to 43.6. Introductory API pricing of $0.75 / $3.75 per MTok runs until 2026-12-31, after which Google lists $1.50 / $7.50.

Reasoning-first flagship succeeding Grok 4.5, with a 500K-token context window and text + image input. Same $2 / $6 per million input/output tokens as Grok 4.5, with a 75% cache-hit discount.

Coding-focused update of the Muse Spark frontier family, released alongside and co-trained with the Muse Code terminal agent. Third Meta Superintelligence Labs model in four months. Meta evaluates it inside its own Muse Code harness at xhigh reasoning effort. Same API pricing as Muse Spark 1.1 at $1.25 / $4.25 per million input/output tokens, with cached input at $0.15; a contributor tier cuts the rate by over 90% in exchange for permission to train on prompts and completions. Rate limits reach 3.000 requests and 4M tokens per minute per team.

General-availability build of the flagship Max model, two weeks after the WAIC preview. 2.4T-parameter sparse Mixture-of-Experts, text/image/video in and text out, 1M-token context (max ~991K input, ~983K with thinking enabled; 65,536 output tokens). Served via Alibaba Cloud Model Studio with an OpenAI- and DashScope-compatible API at $2 / $6 per MTok flat across the full context. The active-parameter count is still undisclosed and no benchmark table has been published. Alibaba announced open weights for Qwen3.8-Max plus a smaller Qwen3.8-27B checkpoint for the following week — the first Qwen-Max-class model to go open-weight. Both have since shipped and are tracked as separate local entries: Qwen3.8-2.4T-A95B (2026-08-12, text-only and always-thinking, custom Qwen3.8-Max License) and Qwen3.8-27B (2026-08-14, Apache 2.0). The open checkpoint is not identical to this hosted API — it drops the vision path — so the benchmark values here, measured against the API, are not carried over to it.

Second generation of Google DeepMind's robotics stack, extending control from the upper body to the whole body: a humanoid can walk, crouch, lean and manipulate in one coordinated motion, and several robots can collaborate on a task. Ships as three models — Gemini Robotics-ER 2 (reasoning, available in Google AI Studio and in private preview on the Gemini Enterprise Agent Platform), plus the VLA and an on-device variant for early-access partners.

Unified multimodal flow model that generates image, video and audio from a single backbone and can be extended to predict robot actions. Rollout is staged: FLUX 3 Video with native synchronised audio (clips up to 20 seconds, aspect ratios from 9:16 to 21:9) and the action model are in application-based early access, image generation is announced for the following weeks, and an open-weight FLUX 3 Dev backbone for later in the year. No public API tier, pricing, parameter count or benchmark methodology published at launch. FLUX 3 Video now ranks #2 in the Text-to-Video Arena at ~1,496 Elo, behind Gemini Omni Flash (~1,512) and ahead of Dreamina Seedance 2.0 (~1,478) — the first independent quality signal for the model.

Workhorse model in the Gemini 3 series, based on Gemini 3.5 Flash, delivering better coding, knowledge work and multimodal performance at improved token-efficiency (Google reports ~17% fewer output tokens than 3.5 Flash). 1M-token context, 64K output, knowledge cutoff March 2026. GA at launch across the Gemini app, AI Studio and Gemini API. Vendor-reported (model card): SWE-Bench Pro 58.7, Terminal-Bench 2.1 78.0, OSWorld-Verified 83.0, MLE-Bench 63.9, CharXiv Reasoning (with tools) 89.4, GDM-MRCR v2 (128k) 91.8. Google's launch benchmark table also lists GPQA Diamond 90.4 (unchanged vs 3.5 Flash). Announced alongside 3.5 Flash-Lite and 3.5 Flash Cyber.

Security-specialized model fine-tuned from Gemini 3.5 Flash to identify, validate and patch software vulnerabilities at low cost per token. Deployed inside Google's CodeMender security agent, which calls it many times in rapid succession to explore alternative execution paths across a codebase. Availability is limited: an initial pilot for governments and trusted partners via CodeMender, not general API access. Google reports internal results of 55 confirmed unique vulnerabilities on Chrome's V8 JavaScript engine (vs 47 for standard 3.5 Flash and 36 for Claude Opus 4.6), a 42% improvement on long-range multi-turn cyber benchmarks over the Flash 3 predecessor, and frontier-competitive performance on CyberGym within CodeMender. No standard academic benchmarks were reported.

2.8T MoE (~50B active, 16 of 896 experts) native multimodal flagship with a 1M-token context window. Launched via API, app and playground on 2026-07-16; open weights scheduled for 2026-07-27.

Coding- and agentic-focused flagship from xAI, trained alongside Cursor. Available in Grok Build, Cursor and the console at $2 / $6 per million input/output tokens.

Pro tier of ByteDance's Seedream 5.0 image family (above Seedream 5.0 and 5.0 Lite). Adds reasoning- and online-search-guided generation, separable-layer editing, and text rendering in 10+ languages, aimed at high-density infographics and professional editing.

Tencent's Hunyuan 3 flagship: a 295B-parameter MoE activating 21B per token (192 routed experts with top-8 routing plus an always-active shared expert), 256K context window and three selectable reasoning-effort modes. Full release under Apache 2.0 — unlike the April preview license, without territorial restrictions.

Faster, lower-cost variant of Nano Banana 2 for rapid text-to-image generation (~4s per image).

Preview of OpenAI's GPT-5.6 frontier model (Sol tier), part of a Sol/Terra/Luna family in limited preview to trusted partners via Codex and the API. Sol priced at $5/M input, $30/M output tokens. Not yet generally available.

Lightweight, cost-efficient tier of OpenAI's GPT-5.6 Sol/Terra/Luna family, aimed at smaller and faster tasks. Lowest-cost option in the family at $1/M input, $6/M output tokens. Part of the limited preview from 2026-06-26; generally available 2026-07-09.

Beta preview of ByteDance's Seedance 2.5 video model, headline feature 30-second one-shot generation. Announced as a global enterprise beta at Volcano Engine 2026; public launch targeted for early July 2026.

Anthropic's first publicly available Mythos-class model, priced at $10/M input and $50/M output. Launched alongside the gated Claude Mythos 5 (same base model with safeguards lifted for a small set of cyberdefenders and infrastructure providers), which is not publicly available and is not tracked here. A safety system routes a minority of sessions (~5% on average) to Claude Opus 4.8. Vendor-reported benchmarks: SWE-bench Pro 80.3, HLE 59.0 (no tools; 64.5 with tools), Terminal-Bench 2.1 88.0.

Microsoft's first in-house coding model, part of the MAI family unveiled at Build 2026 and built on Maia 200 silicon without distillation from other labs. Live in GitHub Copilot and VS Code, and available to developers via OpenRouter, Fireworks and Baseten with tunable weights. Vendor-reported benchmarks: SWE-bench Pro 51.2, GPQA Diamond 84.6 (model card).

Microsoft AI's first flagship reasoning model, a medium-sized model in the first-party MAI family launched at Build 2026. Distributed via OpenRouter, Fireworks and Baseten.

Proprietary flagship in Alibaba's Qwen Max line. Closed weights — the Plus variant is open-sourced while Max stays closed, consistent with Alibaba's Max/Plus split since Qwen2.5-Max. 1M-token context (up from 256K in the Qwen3.6-Max preview), text-only input and output. MoE architecture details are not officially disclosed. Previewed on Arena AI on 2026-05-14, formally launched at the 2026 Alibaba Cloud Summit on 2026-05-20. Alibaba-reported benchmarks (verified post-launch via AA Intelligence Index v4.0 reproduction): GPQA Diamond 92.4, HLE 41.4, SWE-bench Verified 80.4, Terminal-Bench 2.0 69.7. AA Intelligence Index v4.0: 57 (initial Index run reported 56.6).

Fast-tier model in the Gemini 3.5 generation, announced at Google I/O 2026. GA at launch and set as the default model in the Gemini app and AI Mode in Google Search. Architecture is not publicly disclosed. Vendor-reported benchmarks: GPQA Diamond 90.4, MMMU-Pro 81.2, SWE-bench Verified 78. Independent (Artificial Analysis, ~5 days post-launch): HLE 40.2 (exact match to vendor figure), Terminal-Bench 2.1 76.2, AA Intelligence Index v4.0: 55. Distinct from 'Gemini Omni', a separate video/world model announced at the same event.

Default ChatGPT model, replacing GPT-5.3 Instant. API identifier 'chat-latest'. Hallucination-focused update with material gains on math and multimodal reasoning vs. its predecessor. Architecture not publicly disclosed.

Reasoning-first flagship with a 1M-token context window and native video input. Agentic-focused release at lower pricing than Grok 4.1.

Preview release, available via API and open weights. 1.6T total / 49B active MoE with hybrid attention and manifold-constrained hyper-connections. Reports 27% of single-token inference FLOPs and 10% of KV cache vs DeepSeek-V3.2 in a 1M-token setting.

Agentic-focused flagship with 1M context window (400K in Codex). Thinking and Pro variants available. Codename 'Spud'.

Open vision-language-action model for humanoid robots, released in early access under Apache 2.0 — the first fully commercially licensed model of the GR00T line. A vision-language backbone is paired with a diffusion transformer head that denoises continuous actions via flow matching. NVIDIA credits the generalization and language-following gains over N1.6 to 20,000 hours of EgoScale human egocentric video in pretraining. At 3B parameters it is small enough to run on local hardware, which is why it also appears in the Local tab.

Successor family to Qwen3.5 from the Qwen team, focused on agentic coding. 35B total / 3B active MoE with 262K context window and thinking preservation across conversation turns. Compatible with OpenClaw, Claude Code, and Cline. Apache 2.0 weights on HuggingFace.

First model from Meta Superintelligence Labs under Alexandr Wang. Powers Meta AI across Facebook, Instagram, WhatsApp, Messenger and Ray-Ban glasses. Natively multimodal (text/image/voice input, text output) with multi-agent 'Contemplating mode'. Closed-source departure from the Llama lineage. Free on meta.ai and the Meta AI app; API in private preview. Consumer-focused, strong in health and multimodal; coding gap acknowledged by Meta.

Video generation model from Alibaba's ATH AI Innovation Unit. 15B-parameter single-stream Transformer (40 layers) with native audio/video/text token fusion. #1 in both text-to-video and image-to-video on Artificial Analysis Video Arena (Elo blind test). Initially released anonymously; confirmed as Alibaba on April 10, 2026. Beta — no public API or released weights.

Major architectural overhaul on rewritten codebase. ~5x faster generation than V7, native 2K resolution, significantly improved text rendering, better prompt adherence for complex multi-element compositions.

Smallest, cheapest GPT-5.4 variant for classification, extraction, and cost-sensitive sub-agent tasks.

Fast everyday model with reduced hallucinations and improved web search.

Highest-capability model optimized for quality and depth over speed.

Google DeepMind's Nano Banana 2 (API gemini-3.1-flash-image-preview), third in the Nano Banana family after the original (Gemini 2.5 Flash, Aug 2025) and Nano Banana Pro (Gemini 3 Pro, Nov 2025). Brings Pro-tier image quality to the Flash architecture at roughly half the per-image API price (~$0.067 vs $0.134 for a 2K image).

Near-flagship performance at mid-tier pricing with 1M context window and advanced computer use.

Frontier-Flaggschiff der Seed-2.0-Familie (Pro/Lite/Mini/Code) von ByteDance, 256k Kontext, Schwerpunkt Reasoning, Coding und Agent-Tasks. Treibt die Doubao-App, Chinas meistgenutzten KI-Chatbot. Proprietär über Volcano Engine.

744B MoE (40B active) open-weight model with DeepSeek Sparse Attention and strong agentic/frontend coding performance.

Next-generation flagship model with adaptive thinking and agent teams.

High-definition text-to-video model with synchronized cinematic audio.

1T MoE (32B active) open-weight model with native multimodal vision and agent swarm support for up to 100 sub-agents.

Flash-tier model offering frontier-level reasoning at lower cost and higher speed. Predecessor to Gemini 3.5 Flash.

Runway's first General World Model: an autoregressive model built on top of Gen-4.5 that generates frame by frame, runs in real time at 24 fps and is steered interactively through camera pose, robot commands or audio. Ships as three variants, GWM Worlds for explorable environments, GWM Avatars for speaking characters with facial expression and lip sync, and GWM Robotics for synthetic robot training data, which Runway plans to merge into one model. Announced together with native audio for Gen-4.5 and pitched as more general than Google's Genie 3: a simulator for training agents in domains such as robotics and life sciences.

First simultaneous audio-visual generation model.

Successor to DeepSeek V3. 671B MoE (37B active) with DeepSeek Sparse Attention for long-context efficiency and large-scale agentic tool-use pipeline.

Next generation with improved photorealism and typography.

Advanced image generation model building upon Nano Banana.

Next major generation with agentic-first architecture.

Major update with "Instant" and "Thinking" modes for adaptive reasoning.

Odyssey's general-purpose world model, trained on a large corpus of general video and interaction data rather than on game engines or synthetic environments, which the company presents as evidence that basic physics, dynamics and behaviours can be learned from video alone. It generates interactive video that responds to typed input within about fifty milliseconds and keeps running for minutes rather than seconds; Odyssey describes the interaction like using a language model: you type, and the video responds. Publicly usable at experience.odyssey.ml, later extended by Odyssey-2 Pro (API access for embedding continuous simulations into applications) and Odyssey-2 Max.

Ultra-fast, edge-capable model.

Faster, cheaper generation with improved motion and style consistency.

Specialized coding model with architectural breakthroughs.

Next-generation multimodal image model that handles generation and instruction-based editing in one system, with high-definition output up to 4K.

Specialized high-speed coding model.

Open source release of a highly capable model.