Large Language Models
Frontier and open-weight language models for chat, reasoning, coding and agents, compared by price, context window and benchmark.
Gemini 4 Argon
Google DeepMind's Sept 30, 2026 frontier model — DeepSWE v1.1 77.9%, AutomationBench 51.3%, LVBench 91.7%, 1M-token output limit, $2 / $10 introductory price. Limited access via the Fairwind Program; broad rollout pending.
Claude Opus 5.5
Anthropic's Sept 22, 2026 flagship — Fable 5.1-level performance on most work at $4 / $20, 40% cheaper than Opus 5 on typical workloads, 1M context, 30%+ faster output, Terminal-Bench 4.0 66.4%.
GPT-6
OpenAI's GPT-6 family — Astra ($10 / $50, Sept 3), Sol and Luna (Sept 22), GPT-6.1 Sol (Sept 29, near-Astra intelligence at $2 / $10) — 1.05M context, ARC-AGI-3 99.9%, FrontierMath Tier 4 98%.
Claude Fable 5.1
Anthropic's Sept 1, 2026 Mythos-class update — ~25% cheaper than Fable 5 on typical workloads via 75% lower cache reads ($0.25), 60% fewer cyber false positives, Terminal-Bench-Science 52.6% vs 24.7%. Same model as the restricted Mythos 5.1.
GLM-5.3
Z.ai's August 14, 2026 post-training upgrade of GLM-5.2 — the same 753B base with 50% better coding on its in-house bench, Terminal-Bench 3.0 up from 4.6 to 28.3, a 1M context, open weights, and the top CyberGym score reported at launch.
Muse Glimmer
Meta's first Apache 2.0 model — a 30B dense multimodal agent optimized for local deployment on consumer GPUs, with DFlash speculative decoding for fast inference.
Muse Spark 1.2
Meta's coding-focused flagship — co-trained with Muse Code terminal agent, TB 2.1 82.9%, $1.25/$4.25 standard pricing, 1M context, open weights announced.
Qwen3.8-Max
Alibaba's flagship — 2.4T-parameter sparse MoE (95B active), multimodal, top scores on agentic/research benchmarks, $2/$6 pricing, with open weights released on August 12, 2026.
Claude Opus 5
Anthropic's July 2026 flagship — near-Fable intelligence at Opus pricing ($5/$25), SOTA on Frontier-Bench, ARC-AGI-3 and OSWorld 2.0, with thinking on by default.
Gemini 3.6 Flash
Google's July 2026 workhorse model — 17% fewer output tokens than 3.5 Flash, cheaper at $1.50/$7.50, optimized for agentic coding and knowledge work.
Kimi K3
The world's first open 3T-class frontier model — 2.8T parameters (104B active MoE), 1M context, native vision, open weights under a revenue-tiered custom license.
Grok 4.5
xAI's July 2026 coding-and-agent flagship — 500K context, $2/$6 per 1M, default in Grok Build and Cursor, trained alongside Cursor.
Claude Sonnet 5
Anthropic's most agentic Sonnet yet — near-Opus quality for coding, agents and knowledge work at $2/$10 introductory pricing, and the new default model on claude.ai.
GPT-5.6
OpenAI's GPT-5.6 family — flagship Sol, balanced Terra and fast Luna — with max/ultra reasoning modes and tiered pricing. Generally available since July 9, 2026.
Sakana Fugu
Tokyo-based Sakana AI's June 2026 multi-agent orchestration system — shipped as a single OpenAI-compatible model that dynamically routes each task to a swappable pool of frontier and open models.
GLM-5.2
Z.ai's open-weight long-horizon flagship — a 753B-param MoE with a stable 1M context under an MIT license, beating GPT-5.5 on several coding benchmarks.
Kimi K2.7 Code
Moonshot's open-weight agentic coding flagship — a 1T-param MoE (32B active) with 256K context and native image/video input, thinking ~30% leaner than K2.6.
Claude Fable 5
Anthropic's first Mythos-class model for general use — state-of-the-art on nearly every benchmark, with safety classifiers that fall back to Opus 4.8 on sensitive topics. Suspended by US export controls June 12–30, back globally since July 1.
Claude Opus 4.8
The June 2026 overall intelligence leader and top model for agentic coding, with the most careful hallucination calibration of any frontier model.
DeepSeek V4-Pro
The open-weight value champion — near-frontier coding (SWE-V ~85%) under an MIT license with a 1M-token context at a fraction of closed-model prices.
GPT-5.5
OpenAI's unified reasoning-and-chat flagship — the strongest model for terminal-native agentic work, natively omnimodal across text, image, audio and video.
Gemini 3.1 Pro
The highest-value closed frontier model — top-tier reasoning and native multimodality at roughly $2 per 1M input tokens with a 1M-token context.
Claude Haiku 5.5
Anthropic's Oct 7, 2026 small model — from $0.10 / $0.50 per 1M tokens (about 75% cheaper than Haiku 4.5), first Haiku with adjustable effort, 1M context, OSWorld 2.1 offline 72.4%.
Mistral Large 4
Mistral's Oct 6, 2026 public preview — a 1.05T-parameter open-weight multimodal MoE (52B active) with 1M context, state-of-the-art among open models on cybersecurity, finance and law, DeepSWE v1.1 61.7%. Weights due end of October.
Claude Sonnet 5.5
Anthropic's Sept 28, 2026 workhorse — Terminal-Bench 4.0 70.6%, CursorBench 4.0 55.5%, 30%+ faster than Sonnet 5 and up to 30% cheaper per task at an unchanged $2 / $10, with cache reads halved to $0.10.
MiMo-V2.6-Pro
Xiaomi's September 22, 2026 open-weight omni-modal flagship — a 1.02T-parameter MoE (42B active) with text, image, video and audio input, a 1M context and an MIT license, trained with six days of large-scale RL and priced at $0.435 / $0.87 per 1M tokens.
Grok 4.7
xAI's Sept 21, 2026 flagship for coding and knowledge work — a new, larger base model at Grok 4.6's $2 / $6 price and speed, 500K context, xhigh reasoning, CursorBench 4.0 46.3%, Terminal-Bench 4.0 37.6%.
Qwen3.8-Omni-Flash
Alibaba's September 18, 2026 natively omni-modal model — text, image, audio and video in, 1M context, built for audio-video agents (video editing, music videos, long-meeting analysis) at $0.15 / $0.47 per 1M tokens, with a separate Realtime API for live voice and video.
DeepSeek V4.1-Flash
DeepSeek's Sept 10, 2026 open-weight multimodal MoE under MIT — 552B parameters with only 8B / 16B active, 1M context, a KV cache a quarter the size, DeepSWE v1.1 74.2% and Terminal-Bench 2.1 90.6%, from $0.15 / $0.60 per 1M tokens.
Gemini 3.8 Flash
Google's Sept 2, 2026 workhorse — best Flash yet for long-horizon coding and agents, HLE-Verified 54.9%, 1M input / 64K output, multimodal, $0.75 / $3.75 introductory price (rising to $1.50 / $7.50 in 2027).
GLM-5.3-Flash
Z.ai's August 26, 2026 natively multimodal open-weight model — 320B parameters with 18B active, a hybrid sparse-plus-linear attention design, 1M context and an MIT license, beating GLM-5.2 on DeepSWE v1.1 (63.4 vs 46.2) at one-tenth the price.
Qwen3.8-Flash-Next
Alibaba's August 26, 2026 open-weight multimodal MoE and architecture preview for Qwen4 — 125B parameters plus a 51B N-gram embedding with only 6B active per token, SWE-bench Pro 62.5, trained at about 1/9 the cost of Qwen3.7-Plus, hosted as qwen3.8-flash at $0.15 / $0.47.
Nemotron 3.5 Lightning
NVIDIA's highest-efficiency open model for long-running agentic AI workloads — built by the Nemotron Coalition, available with NeMo Switchyard for intelligent multi-model routing.
Inkling
Mira Murati's first model — a 975B/41B-active open-weight MoE with Apache 2.0 license, native text/image/audio, 1M context, designed as a base for fine-tuning.
MiniMax M3
The first open-weight model to combine frontier coding, a 1M-token context and native multimodality — a 428B MoE (22B active) built on sparse attention.
Ornith-1.0
Ornith-1.0 is DeepReinforce's open-source, self-improving LLM family for agentic coding — from a 9B edge model to a 397B MoE that rivals Claude Opus 4.7 on SWE-Bench.
Qwen3.7-Max
Alibaba's proprietary "Agent Frontier" flagship — top knowledge scores and long-horizon autonomy across a 1M-token context.
Grok 4.3
xAI's budget-frontier model — strong agentic tool calling and a self-reported low hallucination rate inside a 1M-token window at $1.25 / $2.50 per 1M.