Google DeepMind's mathematical-reasoning model that formally proves theorems; the AlphaProof Nexus version tackles Erdős problems.

Anthropic's Mythos-class multimodal Claude model (text, vision, code) made safe for general use — strong at software engineering, knowledge work, long-context reasoning, scientific research and protein design, with safeguards that fall back to Claude Opus 4.8 for sensitive domains. Available via the Claude API and claude.ai. Priced at $10 / $50 per million input / output tokens.

Claude Fable 5.1 is Anthropic's September 1, 2026 upgrade to Fable 5, state of the art on coding, knowledge work and long-running agentic problem-solving — 52.6% on Terminal-Bench-Science 0.1 (vs 24.7% for Fable 5), 73.4% on CursorBench 3.2.0 at max effort and 65.0% on Humanity's Last Exam with tools. Generally available on Claude apps, the API, AWS, Google Cloud and Microsoft Azure; a 75% cut in cache-read pricing (to $0.25 per million tokens) makes typical workloads about 25% cheaper than Fable 5.

The same underlying multimodal Claude model as Fable 5, but with safeguards lifted in some domains (e.g. cybersecurity, biology), offered to vetted users through Anthropic's Project Glasswing trusted-access program. Text, vision and code. Priced at $10 / $50 per million input / output tokens.

Claude Mythos 5.1 is the same underlying model as Claude Fable 5.1, released September 1, 2026, but with safeguards lifted for verified work in cybersecurity and the life sciences — including a 60% reduction in false-positive refusals in the cyber domain and advanced protein-design capability (high-affinity binder hit rates near 50%). It is not generally available: access runs only through Anthropic's Cyber Verification Program for defensive security professionals and the Life Sciences Verification Program for US-based researchers.

DeepSeek V4 — DeepSeek's frontier model with a million-token context window built for long-running agentic workloads.

DeepSeek-V4-Flash — the fast, low-cost tier of DeepSeek's V4 family, released at $0.28 per million tokens and upgraded (0731) with stronger agentic and coding performance.

DeepSeek-V4-Pro — the flagship tier of DeepSeek's trillion-scale V4 family, a MoE model aimed at frontier reasoning, coding and agentic workloads above the cheaper V4 Flash tier.

DeepSeek V4.1 Flash is DeepSeek's September 2026 update to its fast, low-cost V4 Flash model. It supports a 1-million-token context window and adds an FP4 KV cache and cross-layer attention reuse to cut the memory cost of long-context, agent-style workloads; early testers reported it reaching about 98% of GPT-6 Astra's score on the OpenDesign Arena at roughly 1.4% of the cost.

The GLM (General Language Model) family from Zhipu AI / Z.ai — open-weight Chinese LLMs (GLM-4.5/4.6/5/5.1/5.2 plus GLM-OCR, Air, Flash and Turbo variants) known for strong coding and long-context performance under permissive (MIT) licenses.

GLM-5.2 — Zhipu AI's flagship GLM model tuned for long-horizon, multi-step agentic tasks.

Most capable and efficient frontier model for professional work.

OpenAI's GPT-5.5 — the default ChatGPT model (Instant), with Pro and Thinking variants for paid plans. Lower hallucination in law, medicine and finance, and stronger STEM reasoning. A GPT-5.5-Cyber variant targets cybersecurity.

OpenAI's GPT-5.6 family (previewed 2026) — tiered Sol, Terra and Luna models with new reasoning modes; reportedly beats Anthropic's Mythos 5. Initial rollout was limited at US government request.

GPT-5.6 Luna — the low-cost, high-throughput tier of OpenAI's GPT-5.6 family, whose price cuts drove the 2026 inference price war.

GPT-5.6 Sol — the flagship tier of OpenAI's GPT-5.6 family, positioned as its highest-capability reasoning model and benchmarked against Claude Opus 5 and Fable 5.

GPT-5.6 Terra — the mid tier of OpenAI's GPT-5.6 family, sitting between Luna and Sol on price and capability, generally available via the OpenAI API and Amazon Bedrock.

GPT-6 Astra is OpenAI's flagship model, released on September 3, 2026 and positioned as a computer-use and coding model rather than a chat model. It ships with a ~1.05M-token context window, reports 72.6% on OSWorld V2-Offline, and is the first OpenAI model to reach the Critical cybersecurity capability level under the company's Preparedness Framework — a milestone OpenAI leadership has described as the start of an AGI era.

Gemini 3.7 Flash is Google's fast, low-cost Gemini tier released in August 2026, aimed at coding and agentic workloads at $0.75 per million input tokens — arriving three weeks after the previous Flash release.

Gemini 3.8 Flash is Google's fast, low-cost Gemini model released on September 2, 2026 — its third Flash model in six weeks, arriving shortly after Gemini 3.7 Flash. Google says it "works harder" by performing more reasoning steps on complex tasks, which can raise costs. It shipped alongside Gemini 3.8 Flash Cyber, a cybersecurity variant of the same underlying model that differs only in its safety mitigations.

Google's fast, cost-efficient Gemini model tier, announced at Google I/O 2026.

Google's 12-billion-parameter open model in the Gemma 4 family — a compact, efficient multimodal LLM designed to run on a single GPU.

Grok is an AI assistant built by xAI. Chat, create images, write code, and get real-time answers from the web and X

Grok 4.5 — xAI's frontier large language model, the successor to Grok 4, with stronger reasoning, coding and real-time knowledge via X integration.

Grok 4.6 — xAI's frontier large language model and successor to Grok 4.5, a 500K-context model tuned for long-running agents, coding and real-time knowledge via X integration.

Claude Haiku 4.5 is our fastest, most cost-efficient model, matching Sonnet 4’s performance on coding, computer use, and agent tasks.

Granite is IBM's family of open, Apache 2.0 licensed language, vision and embedding models aimed at enterprise workloads.

Moonshot AI's flagship 1T-parameter open-weight LLM featuring 262K context window, long-horizon coding with up to 300 sub-agent swarms and 4,000 coordinated steps. Outperforms GPT-5.4 and Claude Opus 4.6 on SWE-Bench Pro (58.6). Supports multimodal input including vision.

Kimi K3 — Moonshot AI's next-generation Kimi model, reported to close the gap with Anthropic's Opus 4.8 and shipped as one of the largest open-source frontier models.

Meituan's 1.6-trillion-parameter open-source agentic coding model with a 1M-token context window (June 2026). The first trillion-parameter model claimed to be fully pre-trained AND served on domestic Chinese AI chips.

MiniMax M3 — MiniMax's long-context reasoning and agentic model, positioned for open-weight deployment on accelerated infrastructure.

Meta's Muse Glimmer — a 30B open-weights, multimodal agentic model built for always-on local agent workflows, small enough to run on a single consumer GPU.

Meta open-source AI video generation model for creating short video clips from text and image prompts.

Meta's Muse Spark 1.2 — the August 2026 update to Meta's paid Muse Spark foundation model, powering the Muse Code terminal agent with higher coding and Terminal-Bench scores at a low per-task price.

By far the most powerful AI model Anthropic ever developed.

NVIDIA's Nemotron 3 Nano — the compact, efficiency-focused tier of the Nemotron 3 open model family, including the Nano Omni multimodal variant for long-context document, audio and video agents.

NVIDIA's Nemotron 3 Ultra — the largest, highest-capability model in the Nemotron 3 open model family, tuned for advanced reasoning, agentic workflows and enterprise AI.

NVIDIA's Nemotron 3.5 Lightning — the speed-optimized tier of the Nemotron 3.5 open model family, built for fast, accurate specialized task execution in long-running agentic workflows.

NVIDIA's Nemotron 4 — the next major generation of the Nemotron open model family, following the Nemotron 3 and Nemotron 3.5 releases, aimed at reasoning-heavy and agentic enterprise AI workloads.

Claude Opus 4.6 is state-of-the-art across a wide range of coding and agentic capabilities.

Opus 4.7 is a notable improvement on Opus 4.6 in advanced software engineering, with particular gains on the most difficult tasks.

Anthropic's flagship Claude Opus 4.8 — the successor to Opus 4.7 with further improvements in advanced reasoning, coding and agentic capabilities.

Claude Opus 5 — Anthropic's most capable model in the Claude 5 family, the frontier successor to Opus 4.8 for the hardest reasoning, coding and agentic work.

Ox Alpha is an anonymous “stealth” AI model that appeared on OpenRouter on August 20, 2026 — a free-preview reasoning model aimed at coding and agentic work, with a 1,048,576-token context window, a 131,072-token output cap and text, image and video input. No lab has claimed it; community fingerprinting (tokenizer probes, API error strings, video-token budgets) points at Z.ai’s GLM family.

Alibaba Cloud's latest Qwen 3-series flagship LLM, the successor to Qwen 3.6 with improved reasoning, coding and agentic capabilities.

Alibaba Cloud's latest Qwen 3-series flagship LLM, the successor to Qwen 3.7 with improved reasoning, coding and agentic capabilities.

Sakana AI model that coordinates and orchestrates multiple models, matching frontier models on some benchmarks.

Hybrid reasoning model with superior intelligence for agents, featuring a 1M context window

Anthropic's Claude Sonnet 4.8, launched alongside Opus 4.8 — brings Opus-class quality into the mid tier with the new Dynamic Workflows tool and improved vision workflows.

Claude Sonnet 5 — Anthropic's balanced mid-tier model in the Claude 5 family, pairing strong reasoning and coding with fast, cost-efficient responses.

Watermelon — the codename for Meta's upcoming frontier AI model, reported to match OpenAI's GPT-5.5 on key benchmarks. Meta's bid to reclaim ground in the frontier LLM race under Alexandr Wang's superintelligence group.