AI Models
Guide

The versatile all-rounder with dynamic internal routing.

• GPT-6 Astra (Sept 3, 2026): flagship at $10/$50 per 1M, and the first model rated Critical for cybersecurity under OpenAI's Preparedness Framework. A follow-up GPT-6.1 Astra was shelved Sept 28 over safety concerns
• GPT-6.1 Sol (Sept 29), GPT-6.1 Terra (Oct 26), and GPT-6 Luna: 1.05M context, replacing the GPT-5.6 tiers. Sol is $2/$10 per 1M with near-Astra agentic coding and computer use; Terra is $1/$5 for computer use, visual reasoning, and everyday office work; Luna is $0.10/$0.50
• In ChatGPT: GPT-6 became the default on Oct 7 (Sol on paid plans, Luna on Free and Go). GPT-6 Pro added parallel agents on Oct 16
• GPT-Live-1 / mini (July 8, 2026): full-duplex voice models with real-time translation, now the default ChatGPT voice mode

The developer favorite for coding, reasoning, and safety.

• Opus 5.5 (Sept 22, 2026): flagship, 1M context, 128K output, $4/$20 per 1M. Performs at Fable 5.1 level on most work at about 40% lower cost than Opus 5; model ID claude-opus-5-5
• Fable 5.1 (Sept 1, 2026): most capable model, 1M context, 128K output, $10/$50 per 1M, and about 25-45% cheaper than Fable 5 on typical work
• Sonnet 5.5 (Sept 28, 2026): $2/$10 per 1M, 1M context, 30%+ faster than Sonnet 5; model ID claude-sonnet-5-5
• Haiku 5.5 (Oct 7, 2026): cheapest tier at $0.10/$0.50 per 1M for prompts up to 100K, with 1M context

Multimodal powerhouse with top benchmark breadth.

• Gemini 4 Carbon (Oct 28, 2026): Google's newest model, focused on cybersecurity and long-horizon work. So far it is limited to selected cyber defenders through the Fairwind Program, with no public release date
• Gemini 4 Argon (Sept 30, 2026): Google's generally available flagship, focused on coding, agentic work, research, and cybersecurity — 1M-token output (up from 64K). Generally available since Oct 13 to paid Gemini API developers, AI Pro and Ultra subscribers, and Gemini for Business. $2/$10 per 1M is introductory through Nov 30; from Dec 1 it becomes $4/$20
• Gemini 3.8 Flash (Sept 2, 2026): new default workhorse, built on 3.7 Flash — beats 3.7 on every benchmark Google published, 1M context, 64K output, same $0.75/$3.75 per 1M pricing through Dec 31, 2026
• Gemini 3.5 Flash-Lite and Flash Cyber (July 21, 2026): cost-efficient high-volume and defensive-security variants
• Gemini 3.1 Ultra (May 2026): top reasoning tier — 2M context, natively multimodal
• Gemini 3.1 Pro (Feb 2026): deployed Pro-tier flagship — 94.3% GPQA Diamond, 1M context, $2/$12 per 1M
• Gemini 3.1 Flash Live and Flash TTS round out the audio lineup; Omni Flash (May 19, 2026) adds any-to-any input with video output

Real-time data meets raw reasoning power.

• Grok 4.3 (Apr 30, 2026): cost-efficient API flagship — $1.25/$2.50 per 1M, always-on reasoning, native video input up to 5-minute clips
• Grok 4 Heavy: premium multi-agent variant — 256K context, first model to score 50.7% on Humanity's Last Exam; gated behind the $300/month SuperGrok Heavy tier
• Grok 4 Heavy also scored 100% on AIME 2025 — often mistakenly attributed to the base Grok 4.x model
• Real-time integration with X (Twitter) for current events
• Grok 4.7 (Sept 21, 2026): supersedes Grok 4.6 — $2/$6 per 1M; Fast variant runs 2× faster at 2× the price
• Grok Voice 3.0 (Oct 14, 2026): speech-to-speech model for voice agents, $0.08 per minute

Search-native AI built for grounded, cited answers.

• Custom Sonar model fine-tuned for search-grounded factuality and citation accuracy
• Runs at 1,200 tokens/sec on Cerebras inference hardware
• Model family: Sonar, Sonar Pro, Reasoning Pro, Deep Research
• Matches GPT-4o on user satisfaction benchmarks

Not an LLM — a System One model that outputs typed, schema-conformant decisions instead of generated text.

• Early access Sept 15, 2026 — founded by Diogo Almeida
• "Unstructured state in, typed probabilistic decisions out" — built for classification, routing, scoring, and extraction, not chat
• Outputs generated in parallel rather than token-by-token, so schema conformance is structural, not prompted — can't hallucinate or produce a type error by design
• 70-500ms end-to-end — 40-200× faster than a comparable frontier LLM on the same task
• $0.042 per 1M input tokens, output free; trade-off is no text generation — structured decisions only

Frontier-class coding at one-tenth the price — built on Kimi K2.5 with Cursor's own post-training pipeline.

• Released May 18, 2026 — built on Kimi K2.5 with Cursor's own post-training
• 79.8% SWE-bench Multilingual, 69.3% Terminal-Bench 2.0 — matches Claude Opus 4.7 and GPT-5.5 on coding benchmarks
• $0.50/$2.50 per 1M — roughly 10× cheaper than Opus 4.7
• Trained with RL for complex, hundreds-of-steps tasks
• Fast variant is the default; background agents run tasks autonomously

Microsoft's own foundation model stack — independent of OpenAI, built for speech, voice, vision, and reasoning.

• MAI-Transcribe-2-Streaming (Oct 1, 2026): real-time speech-to-text in 60 languages with automatic language detection, $0.54 per audio hour (public preview)
• MAI-Voice-2.1 and Flash (Oct 1, 2026): text-to-speech in 23 languages with one voice carried across languages; Flash targets low-latency, high-volume use
• MAI-Image-2: high-quality image generation
• Phi-4-reasoning (Apr 10, 2026): compact reasoning model, strong math/logic at small-model cost
• MAI-Thinking-1 (Jun 2026): 1T/35B active MoE reasoning model — flagship of a seven-model MAI launch
• MAI-Code-1 / Flash (Jun 2026): 5B-param coding model trained on Copilot's production workflows
• All available on Microsoft Foundry — independent of OpenAI

Meta's closed-weight frontier models — a break from the open Llama strategy, with an open-weight sibling in Muse Spark Open and Muse Code as its terminal coding agent.

• Released April 8, 2026 — led by Alexandr Wang (ex-Scale AI)
• Multimodal reasoning with thought compression and parallel sub-agent orchestration
• Muse Spark 1.3 (Sept 2, 2026): improved coding/agentic performance for longer-horizon work — #6 on Artificial Analysis Intelligence Index, $1.25/$4.25 per 1M
• Muse Wonder (Oct 12, 2026): Meta's first frontier model — closed weights, 1M context, multimodal input, built for agents, $4/$20 per 1M via the Meta Model API and Muse Code. Muse Wonder Full, the higher tier, is only in Muse Code's $50 and $200 monthly plans
• Muse Spark Open (Oct 15, 2026): open-weight sibling — 816B total / 42B active MoE, 1M context, multimodal input. Released under the Muse Model License (free use and fine-tuning, but products must show "Built with Muse Spark Open"), so it is not OSI open source

Apple's native on-device + cloud model framework — privacy-preserving inference shipping with iOS/macOS developer APIs.

• AFM 3 family shipped at WWDC 2026 (June 8) — five models spanning on-device (Core, Core Advanced) and cloud (Cloud, Cloud Pro, ADM 3 Cloud for image gen)
• Multimodal image input added this generation; Python SDK released for developers
• Cloud Pro: hosted on Google Cloud/NVIDIA via Private Cloud Compute for heavier workloads; on-device tiers handle sensitive tasks locally
• Ships as native iOS/macOS developer APIs

AWS's enterprise foundation model family — multimodal, agentic, and natively integrated with Bedrock.

• Released December 2025 at AWS re:Invent — Amazon's flagship in-house model family
• Nova 2 Lite: 1M context with native MCP support — fast, cost-efficient tier
• Nova 2 Pro: tuned for complex reasoning and agentic workflows
• Nova 2 Sonic: speech-to-speech for low-latency voice apps
• Nova 2 Omni: unified multimodal across text, image, audio, and video
• Powers Nova Act (agentic browser, 90%+ task reliability) and Nova Forge for custom model training

Fine-tuned GPT-5.5 for enterprise defensive cybersecurity — gated behind OpenAI's Trusted Access for Cyber program.

• Released May 7, 2026 — fine-tune of GPT-5.5 for dual-use security research; supersedes GPT-5.4-Cyber, still available under the same program
• Lowered refusal thresholds for defensive cybersecurity tasks; native binary reverse engineering without source code
• Gated behind OpenAI's Trusted Access for Cyber (TAC) program; partners include CrowdStrike, Cloudflare, Palo Alto, Cisco, JPMorgan, Goldman Sachs
• No public API pricing; $10M in credits committed via the Cybersecurity Grant Program
• Direct counterpart to Claude Mythos 5 — a matched pair of restricted cyber models

Cybersecurity research model via Project Glasswing — Anthropic's counterpart to GPT-5.5-Cyber. Updated to Mythos 5.1 alongside Fable 5.1 on Sept 1, 2026.

• Originally launched April 6–7, 2026 as Mythos Preview; graduated to Mythos 5 on June 9, 2026 alongside Fable 5; updated to Mythos 5.1 on Sept 1, 2026 alongside Fable 5.1
• GPQA 0.9; 93.9% SWE-bench Verified; 97.6% USAMO 2026 (as measured on Mythos 5)
• Expanded June 2, 2026 from ~12 initial partners to 150+ organizations across 15+ countries for critical-infrastructure cybersecurity
• Same weights as Fable 5.1 — safety classifiers lifted only for Glasswing partners, not a separate model
• High-cost pricing retained to gate general use and prevent misuse

Restricted defensive-security variant of Gemini 3.8 Flash — gated to vetted cyber defenders via Google's Fairwind Program.

• Released September 2, 2026 alongside general-availability Gemini 3.8 Flash
• Restricted to governments, critical infrastructure operators, and software maintainers approved through the Fairwind Program — not on general release or a public price sheet
• 70%+ real-world vulnerability discovery rate; sits on the CWE-Bench Pareto frontier for patching speed
• Participating orgs must limit access to internal cybersecurity/incident-response/pentest teams and enforce MFA
• Third major lab (after OpenAI's GPT-5.5-Cyber and Anthropic's Mythos 5.1) to ship a gated defensive-cyber model

A meta-model that presents as a single API but dynamically routes tasks across a pool of frontier models internally — no hardcoded rules. Now ships in two tiers.

• Fugu Max and Fugu Ultra v2 (Sept 11, 2026) replace the original Fugu Ultra — same orchestration architecture split into a cost-efficient tier (Max) and a max-capability tier (Ultra v2)
• Fugu Ultra v2 hits 74.3 on DeepSWE and 48.3 on Chartography (vs. 27.3 for Opus 5, 29.5 for Fable 5) — and does it without Fable 5, Fable 5.1, or GPT-6-Astra in its own agent pool
• Pricing: Fugu Max $2/$6 per 1M; Fugu Ultra v2 $5/$30 per 1M standard ($0.50/1M cache reads), $10/$45/$1 above 272K context
• Can substitute models around export controls — built for research, cybersecurity, multi-step patent work
• No hardcoded routing rules — the meta-model selects and coordinates sub-models per task

The leading open-source model family.

• Llama 4 Scout: industry-leading 10M token context window
• Llama 4 Maverick: 17B active / 128 experts — outperforms GPT-4o and Gemini 2.0 Flash on key benchmarks
• Fully open weights; can be self-hosted for complete data control
• Llama 4 Behemoth (288B active): effectively shelved — repeated delays and MoE-routing failures at 2T scale; Meta's newer models are the Muse line, including the open-weight Muse Spark Open

Cost-redefining open-source frontier — V4.1-Flash outperforms the flagship V4-Pro.

• V4.1-Flash (Sept 10, 2026): 552B MoE at launch, 8B active on input / 16B active on output — new Causal Encoder-Decoder architecture, native multimodal image understanding, MIT license; beats the larger V4-Pro on cost, speed, and total runtime
• V4.1-Flash-1017 (Oct 17, 2026): updated checkpoint under the same deepseek-flash API name, MIT license — better benchmarks and shorter outputs than the Sept 10 release, 77.0 on DeepSWE
• $0.15/$0.60 per 1M off-peak, doubling to $0.30/$1.20 at peak hours — among the cheapest frontier-class models available
• V4-Pro stays available: DeepSeek said it will keep serving V4-Pro after Sept 14 at unchanged billing, after first planning to reroute it; the old V4-Flash and vision-preview model names now route to V4.1-Flash
• MIT license; native 1M context window; built-in agentic long-context and tool-use
• Entire V3/V4 lineage trained for under $6M — redefining AI cost efficiency

The enterprise-safe European model family, now spanning text, reasoning, code, and speech.

• Mistral Large 4 (Oct 6, 2026): EU AI Act-compliant flagship — 1.05T MoE, 49B active, 1M context, $1.36/$4.18 per 1M. Open weights on Hugging Face since Oct 29 under Mistral's own license
• Mistral Medium 3.5 (Apr 29, 2026): 128B dense, 256K context — retired Magistral (reasoning), Devstral 2 (coding), and Medium 3.1 (chat) into one unified model with configurable reasoning effort; 77.6% SWE-bench Verified, $1.50/$7.50 per 1M
• Mistral Small 4 (March 16): 119B/6.5B-active MoE unifying reasoning, vision (Pixtral), and coding in one endpoint
• Voxtral (March 26): open-source 4B text-to-speech, 9 languages, runs on consumer hardware
• Strong European data sovereignty guarantees across the full model family

Cost-efficient open-weight coding leader — 1M context and top SWE-bench Pro score per dollar.

• Released June 1, 2026 — 1M context window; 59.0% SWE-bench Pro; successor to M2.5 (Feb 12) and M2.7 (late April)
• $0.15/1M input, $1.15/1M output — among the cheapest frontier-class coding models
• M2.5 base: 230B MoE / 10B active, 80.2% SWE-Bench Verified; M3 extends context and benchmark lead

Open-weight frontier model from Moonshot — 2.8T MoE, with K3.1 Pro as the newest post-trained version.

• Kimi K3.1 Pro (Oct 23, 2026): post-trained successor on the same 2.8T MoE — $2/$10 per 1M (a third cheaper than K3), 1M context. K3.1 Pro Highspeed runs 2-3× faster at $4/$20
• Kimi K3 (July 16, 2026; full weights July 27): 2.8T parameter MoE, successor to K2.7 Code
• Open-weight, Modified MIT license; weights on Hugging Face
• Kimi Code CLI agent rivals Claude Code and Gemini CLI

Ant Group's open-weight agent and coding model — 560B MoE with 25B active parameters.

• Ling 3.1 Flash (Sept 30, 2026): 560B total / 25B active MoE, 1M context, built for coding, terminal, and tool-calling work
• Weights released on Hugging Face under the MIT license on Oct 14, 2026

Frontier-class model on a MIT license — now with a fast multimodal variant.

• GLM-5.3 (Aug 14, 2026): scaling-post-training upgrade over GLM-5.2 — same ~744B parameter MoE (40B active), 1M context extended to 128K output; text-only, tuned for code, agents, and cyber-defense
• GLM-5.3-Flash (Aug 26, 2026): first native multimodal model in the GLM-5 series — 320B/18B active MoE, hybrid sparse+linear attention, 1M context, $0.15/$0.50 per 1M (~1/10th GLM-5.3's price)
• Released under MIT license; trained entirely on Huawei Ascend chips (zero NVIDIA GPUs)
• Priced roughly 6x cheaper than comparable proprietary models

Open-weight agentic model for long-horizon research, built on a GLM-5.2 base — appeared on Hugging Face before its own technical report.

• Released Sept 11, 2026, with no announcement; the 140-author technical report followed three days later
• 744B-parameter MoE built on the GLM-5.2 foundation; MIT license, open weights on Hugging Face
• Trained via a "Verifiable Experience Pipeline" — connects tool-mediated interactions to executable environments and externally verified outcomes
• Vendor claims top score on 5 of 16 benchmarks, incl. 59.6% SWE-bench Pro (vs. 74.7% for Fable 5.1) — none independently verified by Artificial Analysis yet
• Built for carrying scientific work from a published method to executable experiments and a reproducible report

NVIDIA's open agentic reasoning stack — Nano, Super, Ultra 550B, and the new 3.5 Super VL and Bolt models.

• Nemotron 3.5 (Oct 15, 2026): Super VL, a 120B open vision-language model for enterprise agents, and Bolt 74B-A9B, a 2-bit 24GB build for local RTX and DGX workloads. Bolt E4B (Oct 23), a roughly 4GB build derived from Gemma 4 for edge devices, runs on about 6GB of RAM. Nemotron 3.5 Transcribe (Oct 28) is an open-weight speech-to-text model for recorded audio: 40 languages, translation to and from English, and speaker labels for up to eight speakers
• Nemotron 3 Ultra 550B (June 4, 2026): 550B/55B active MoE, hybrid Mamba-Transformer, 1M context; 48/100 AA Intelligence Index at launch — the strongest open US model benchmarked at the time
• Nemotron 3 Nano Omni: 30B MoE omni-modal model unifying vision, audio, and text
• Nano and Super sizes released March 2026; Super peers with Llama 4 Maverick on benchmarks
• Weights, data, and recipes public on Hugging Face; available via Amazon Bedrock and NVIDIA NIM

Open-weight models from Gemini 3 research — optimized for on-device and frontier-class performance.

• Four Apache 2.0 models: E2B (2.3B), E4B (4.5B), 26B MoE (4B active), 31B dense (Apr 2, 2026)
• 31B ranks #3 on Arena AI leaderboard at 1452 Elo — outperforms models 20× its size
• E2B/E4B optimized for on-device Android: up to 4× faster and 60% less battery than prior Gemma
• All models natively multimodal; larger variants support 256K context

Cohere's first fully open-licensed frontier model — enterprise-tuned MoE for retrieval-augmented generation.

• Command A+ (May 20, 2026): 218B/25B active MoE, Apache 2.0, runs on two H100s — supersedes Command A as Cohere's flagship
• Purpose-built for RAG: strong grounding, citation accuracy, and document comprehension
• Cost-efficient enterprise tier; competitive on retrieval benchmarks against larger models
• Available via Cohere API and on major cloud marketplaces

The multilingual giant — Qwen3.8-Max is now the flagship, succeeding Qwen 3.7 Max; open-weight Qwen3.5 remains available.

• Qwen3.8-Max (GA Aug 2, 2026): flagship succeeding Qwen 3.7 Max — 2.4T/95B active MoE; hosted API is multimodal, 1M context, $2/$6 per 1M; open weights (Aug 12) under a restricted revenue-share license, text-only, 262K context
• Qwen3.8-27B (Aug 13–14, 2026): separate, genuinely self-hostable companion — 27B dense, multimodal with vision encoder, 262K context, Apache 2.0
• Qwen3.5 (Feb 2026): open-weight, 397B params, available for self-hosting; Qwen3.5-Omni adds native audio/video/text (256K context, 113 languages)
• Qwen3-Coder: 69.6% SWE-Bench Verified; Qwen3 Coder Next is the follow-up

The first commercial subquadratic LLM — a clean break from quadratic-attention transformers.

• Released May 5, 2026 — first production model built on Subquadratic Sparse Attention (SSA), scaling ~linearly with context length
• 12M-token production context window
• Matches Claude Opus-class performance on coding benchmarks at roughly 1/5th the compute cost, with attention up to 52× faster at scale
• Architecturally distinct from every other model on this page — the most significant new attention mechanism shipped in 2026
• Backed by $29M seed round (May 2026); positioning as infrastructure for very-long-context agents