AI Models
Guide

The versatile all-rounder with dynamic internal routing.

• GPT-5.6 Sol / Terra / Luna (July 9, 2026): OpenAI's current flagship API family, now broadly available after an early access period — 1M token context, with an internal router selecting the right sub-model per request. Sol (flagship): $5/$30 per 1M tokens — OpenAI's best coding model, adds 'Sol Ultra' multi-agent mode. Terra (balanced): $2.50/$15 per 1M tokens. Luna (fast/cheap): $1/$6 per 1M tokens.
• GPT-5.5 Instant (May 5, 2026): the ChatGPT default for all tiers — leaner variant tuned for low latency (~30% fewer words/lines per response), distinct from the API flagship family. This is what most consumer users actually hit
• Available to Plus, Pro, Business, and Enterprise via ChatGPT and Codex; native computer use and tool calling for agentic automation
• GPT-Live-1 / GPT-Live-1 mini (July 8, 2026): full-duplex voice models — listen and speak simultaneously, real-time translation, nine remastered voices; agentic delegation routes hard queries to GPT-5.5; GPT-Live-1 is now the default voice mode for Plus/Pro, mini is the Free default

The developer favorite for coding, reasoning, and safety.

• Claude Opus 5 (July 24, 2026): new flagship — 1M context, 128K max output (300K on Batch API); $5/1M input, $25/1M output; delivers frontier intelligence close to Fable 5 at roughly half the price (within 0.5% of Fable 5 on CursorBench at half the cost); the recommended starting model for agentic coding and enterprise work; effort defaults to high; model ID: claude-opus-5
• Fable 5 (June 9, 2026): Anthropic's most capable model — 1M context, 128K max output, $10/1M input, $50/1M output; 95% SWE-bench Verified, 80% SWE-bench Pro; the pick for max-capability workloads over Opus 5
• Sonnet 5 (June 30, 2026): most agentic Sonnet yet — the default model on Free and Pro tiers; near-flagship performance at a fraction of the cost; $2/1M input, $10/1M output (intro pricing through Aug 31, 2026); model ID: claude-sonnet-5
• Dynamic Workflows: run parallel subagents on independent subtasks and stitch results back together
• Fast mode across the Opus line is ~2.5× faster at ~2-3× the standard-inference price
• Haiku 4.5: smallest tier — hits ~90% of Sonnet 5's agentic coding performance at a fraction of the cost; ideal for high-volume tasks
• Leads human-preference leaderboards; strong ARC-AGI-2 scores
• Claude Mythos 5 (released June 9, 2026 alongside Fable 5): cybersecurity research model via Project Glasswing — 150+ organizations across 15+ countries for critical-infrastructure use cases. Same weights as Fable 5 — differs only in that safety classifiers are lifted for Project Glasswing partners. Not a separate model.

Multimodal powerhouse with top benchmark breadth.

• Gemini 3.5 Pro (Google I/O, May 20, 2026): Google's next frontier flagship — has now missed three consecutive GA deadlines (June 30, then July 17); no confirmed launch date, reportedly rebuilt from scratch after recursive tool-calling failures and hallucination rates falling short of GPT-5.6. Still in limited Vertex AI enterprise preview. Leads Gemini 3.1 Pro on reasoning benchmarks; 2M context window; Deep Think reasoning mode
• Gemini 3.6 Flash (July 21, 2026): new default model in the Gemini app, superseding 3.5 Flash — 17% fewer output tokens for equivalent tasks; DeepSWE coding benchmark up to 49% (from 37%), OSWorld-Verified up to 83.0% (from 78.4%); knowledge cutoff advanced to March 2026; $1.50/1M input, $7.50/1M output (cut from $9.00); 1M input context, 64,000 max output tokens; native multimodal input, thinking controls, built-in Computer Use as a client-side tool
• Gemini 3.5 Flash-Lite (new gen) and Gemini 3.5 Flash Cyber (July 21, 2026): released alongside 3.6 Flash — Flash-Lite shares the 1M context / 64K output spec for cost-efficient high-volume workloads; Flash Cyber is a defensive-security-focused variant
• Gemini 4 teased same day (July 21, 2026) — no release date confirmed
• Gemini 3.1 Ultra (May 2026): current top reasoning tier — 2M-token context, natively multimodal across text, image, audio, and video
• Gemini 3.1 Pro (Feb 2026): still Google's deployed Pro-tier flagship pending 3.5 Pro's release — 94.3% GPQA Diamond, 77.1% ARC-AGI-2, 1M context, $2/1M input, $12/1M output; deep Google Workspace integration
• Gemini 3.1 Flash Live (audio/voice, 90+ languages) and Flash TTS (native text-to-speech) round out the audio lineup
• Gemini Spark (May 19, 2026): Google's 24/7 persistent AI agent built on Gemini 3.5 Flash — always-on assistant layer rolling out across Google products as the default interactive experience

Real-time data meets raw reasoning power.

• Grok 4.3 (Apr 30, 2026): cost-efficient API flagship — $1.25/1M input, $2.50/1M output; always-on reasoning, native video input up to 5-minute clips, PDF / spreadsheet / deck generation, ~20% cheaper per benchmark run than prior versions
• Grok 4 Heavy: premium multi-agent variant — 256K context, first model to score 50.7% on Humanity's Last Exam (text-only with tools); gated behind the $300/month SuperGrok Heavy tier. This is the model the 'multi-agent specialists that debate before responding' description actually refers to
• Grok 4 Heavy also scored 100% on AIME 2025 — widely (and incorrectly) attributed to the base Grok 4.x model
• Real-time integration with X (Twitter) for current events
• Grok 4.5 (launched July 8, 2026): public release of the V9 MoE architecture — 1.5T parameters, 500K context window, $2/$6 per 1M tokens (>60% cheaper than Opus 4.8/GPT-5.5); trained on agentic coding data (Cursor sessions); EU rollout is partial as of late July — live via Cursor, but the direct API console is still restricted for EU users
• Grok Voice: voice interface model released June 4, 2026
• Grok 5 still slipping — ~6T parameter MoE on xAI's Colossus 2 supercluster; Polymarket gives ~33% probability of a pre-July ship, with a 1.5M-token context window targeted

Search-native AI built for grounded, cited answers.

• Custom Sonar model fine-tuned for search-grounded factuality and citation accuracy (base model has evolved since launch — check Perplexity docs for the current version)
• 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

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 base with ~85% of compute spent on Cursor's post-training stack
• 79.8% SWE-bench Multilingual, 63.2% CursorBench v3.1, 69.3% Terminal-Bench 2.0 — matches Claude Opus 4.7 and GPT-5.5 on coding benchmarks
• Standard tier at $0.50/1M input, $2.50/1M output — roughly 10× cheaper than Opus 4.7
• Trained with RL to solve complex tasks requiring hundreds of sequential actions
• Fast variant remains the default; background agents run tasks autonomously while you work

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

• MAI-Transcribe-1: speech-to-text across 25 languages; outperforms Whisper-large-v3 on accuracy
• MAI-Voice-1: generates 60s of audio in 1s; supports voice cloning
• MAI-Image-2: high-quality image generation
• Phi-4-reasoning (Apr 10, 2026): compact reasoning model — strong math/logic performance at small-model cost
• MAI-Thinking-1 (Jun 2026): 1T total / 35B active MoE reasoning model — flagship of a seven-model MAI launch; competes at frontier reasoning tier
• MAI-Code-1 / MAI-Code-1-Flash (Build 2026, Jun 2): 5B-param in-house coding model trained on GitHub Copilot's production workflows; rolled out across Copilot plans and VS Code
• All available on Microsoft Foundry — signals Microsoft building foundational AI independent of OpenAI (Apr 2, 2026)

Meta's first closed-weight frontier model — a deliberate break from the open Llama strategy, now with Meta's first paid developer API.

• Released April 8, 2026 — led by Alexandr Wang (formerly Scale AI) following Meta's $14.3B investment
• Multimodal reasoning with thought compression and parallel sub-agent orchestration
• Scored 52 on the Artificial Analysis Intelligence Index — top 5 globally
• Muse Spark 1.1 (July 9, 2026): major upgrade — 1M token context, agentic coding focus (tool use, computer use, multi-agent orchestration), and Meta's first-ever paid developer API at $1.25/$4.25 per 1M input/output tokens; claims top scores on MCP Atlas, JobBench, and Humanity's Last Exam
• Closed weights: a strategic departure from Meta's open-source Llama lineage
• Positions Meta directly against OpenAI, Anthropic, and Google in the closed-frontier tier

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: AFM 3 Core (on-device, 3B dense), AFM 3 Core Advanced (20B sparse), AFM 3 Cloud, ADM 3 Cloud (image generation), and AFM 3 Cloud Pro
• Multimodal image input support added in this generation; Python SDK released alongside for developer access
• AFM 3 Cloud Pro: cloud inference tier hosted on Google Cloud/NVIDIA infrastructure via Private Cloud Compute for heavier workloads; on-device tiers handle sensitive tasks locally
• Ships as native iOS/macOS developer APIs — direct integration into the Apple ecosystem

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-token context with native MCP support — fast and cost-efficient tier
• Nova 2 Pro: tuned for complex reasoning and agentic workflows
• Nova 2 Sonic: speech-to-speech model for low-latency voice apps
• Nova 2 Omni: unified multimodal across text, image, audio, and video
• Powers Nova Act — agentic browser service claiming 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 (Apr 14, 2026), which remains 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; enterprise partners include CrowdStrike, Cloudflare, Palo Alto Networks, Cisco, JPMorgan, Goldman Sachs
• No public API pricing; $10M in API credits committed via Cybersecurity Grant Program
• Direct counterpart to Anthropic's Claude Mythos 5 — part of a matched pair of restricted cyber-focused models

Cybersecurity research model via Project Glasswing — Anthropic's counterpart to GPT-5.5-Cyber. Graduated from Preview to full release June 9, 2026.

• Originally launched April 6–7, 2026 as Claude Mythos Preview via Project Glasswing; graduated to Claude Mythos 5 on June 9, 2026 alongside Fable 5
• GPQA 0.9; 93.9% SWE-bench Verified; 97.6% USAMO 2026
• June 2, 2026: Anthropic expanded access from the initial ~12 partners (AWS, Apple, Cisco, CrowdStrike, Google, JPMorgan, Microsoft, NVIDIA, others) to 150+ organizations across 15+ countries, targeted at critical-infrastructure cybersecurity
• Same weights as Fable 5 — differs only in that safety classifiers are lifted for Project Glasswing partners. Not a separate model.
• High-cost pricing retained to gate general use and prevent misuse
• Cybersecurity-focused capabilities not available in standard Claude models

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

• Released June 22, 2026 — based on Sakana's ICLR 2026 papers: TRINITY coordinator + Conductor orchestration architecture
• 73.7% SWE-Bench Pro at launch — above Opus 4.8 (69.2%), below Fable 5 (80.3%); v1.1 (July 2026) gained up to 7.9 points on ProgramBench/Terminal-Bench 2.1 at the same pricing, so this figure is likely conservative today
• Pricing: $5/$30 per 1M tokens standard; $10/$45 above 272K context
• Can substitute models around export controls — designed for research, cybersecurity, multi-step patent work
• Notable: no hardcoded routing rules — the meta-model selects and coordinates sub-models dynamically 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 (originally April, then June), MoE-routing and chunked-attention failures at 2T scale; Meta pivoting to closed-weight Muse Spark; no formal cancellation but no release expected near-term
• Llama 4.5: Meta has separately announced a Llama 4.5 open-weight release for 2026 — the next step in the open Llama line, distinct from the closed-weight Muse Spark family; no release date confirmed as of July 2026

Cost-redefining open-source frontier — released April 24, 2026 in two variants.

• V4-Pro: 1.6T total parameters / 49B active MoE; $0.435/1M input, $0.87/1M output (permanent price cut from launch pricing as of May 22, 2026) — cheapest frontier-class open model
• V4-Flash: 284B total / 13B active; $0.14/1M input, $0.28/1M output — ultra-budget tier
• MIT license; native 1M context window; built-in agentic long-context and tool-use
• Engram conditional memory and Manifold-Constrained Hyper-Connections for improved long-context performance
• 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 3: EU AI Act-compliant flagship for regulated industries (finance, healthcare, gov); 675B total MoE
• Magistral: Mistral's reasoning model — multilingual, transparent chain-of-thought
• Mistral Small 4 (March 16): 119B/6.5B-active MoE unifying reasoning (Magistral), vision (Pixtral), and coding (Devstral) in one endpoint
• Devstral 2 (Dec 9, 2025): 72.2% SWE-bench Verified — top open agentic coding model; 123B params, 256K context; Modified MIT license with a commercial restriction for companies >$20M/month revenue (separate Mistral agreement required)
• Devstral Small 2 is deprecated (Feb 27, 2026 per Mistral's model card) — Mistral now points self-hosters to Devstral 2 directly
• 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 that closes the gap with US labs — 2.8T MoE, second only to Fable 5 and GPT-5.6 on most benchmarks.

• Released July 16, 2026 (full weights release July 27) — 2.8T parameter MoE, successor to K2.7 Code
• Outperforms every other model except Claude Fable 5 and GPT-5.6 on most benchmarks
• K2.7 Code (June 12, 2026): still available — 1T MoE coding-specialized variant of K2.6, 30% fewer thinking tokens than K2.6
• Open-weight, Modified MIT license; weights on Hugging Face
• Kimi Code CLI agent rivals Claude Code and Gemini CLI

Frontier-class model on a MIT license — now with 1M context.

• Released June 13, 2026 (Z.ai / Zhipu AI) — 1M context window; successor to GLM-5.1 (Apr 6)
• GLM-5.1 base: 744B parameter MoE (44B active); GPQA 0.9; 77.8% SWE-bench Verified
• Released under MIT license; trained entirely on Huawei Ascend chips (zero NVIDIA GPUs)
• Priced roughly 6x cheaper than comparable proprietary models

NVIDIA's open agentic reasoning stack — Nano, Super, and Ultra 550B sizes on Bedrock.

• Nemotron 3 Ultra 550B (June 4, 2026): 550B-parameter open MoE / ~55B active per token; hybrid Mamba-Transformer architecture for sub-quadratic long-context efficiency; 1M context; OpenMDW-1.1 license; 48/100 on Artificial Analysis Intelligence Index — strongest open US model benchmarked
• Nemotron 3 Nano Omni: 30B MoE omni-modal reasoning model — unifies vision, audio, and text; distinct from the Ultra 550B; available alongside Nano/Super sizes
• Weights, training data, and recipes all public on Hugging Face
• Earlier sizes (Nano, Super) released March 2026; available on Amazon Bedrock and NVIDIA NIM
• Nemotron 3 Super peers with Llama 4 Maverick on open-model benchmarks
• Optimized for NVIDIA hardware; available for self-hosting via NIM microservices

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 total parameters / 25B active MoE, fully Apache 2.0 licensed; runs on as few as two H100 GPUs; supersedes Command A (Apr 7, 2026 — 111B/11B active) 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 — Qwen 3.7 Max and Plus lead the current lineup; open-weight Qwen3.5 remains available; Qwen3.8-Max previewed and closing in.

• Qwen3.8-Max previewed July 19, 2026 — 2.4T parameters, Alibaba claims 'second only to Fable 5'; not yet GA, no independent benchmark table published, watch this space
• Qwen 3.7 Max (May 19, 2026): closed-weight flagship — 56.6 on Artificial Analysis Intelligence Index; $2.50/1M input, $7.50/1M output
• Qwen 3.7 Plus (GA June 1, 2026): closed-weight multimodal with vision, deep reasoning, and tool invocation; built on the Bailian platform
• Qwen3.5 (Feb 2026): open-weight, 397B parameters — available for self-hosting and fine-tuning
• Qwen3.5-Omni: native audio/video/text multimodal — Thinker architecture, 256K context, 113-language speech recognition
• Qwen3.6-Plus (Apr 2, 2026): closed-source API-only, 1M context; still available alongside 3.7 lineup
• 0.6B to 235B open-weight range; Qwen3-Max (1T+) API-only; supports 119 languages
• Qwen3-Coder achieves 69.6% on SWE-Bench Verified; Qwen3 Coder Next (Feb 4, 2026) 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), which scales roughly linearly with context length rather than quadratically
• 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