LLM Field Guide

Current serving checkpoint of DeepSeek's fast model (DeepSeek-V4.1-Flash-1017, released 2026-10-17), which the deepseek-flash API alias now points to. All 11 vendor benchmarks improved over the 0924 build, including DeepSWE v1.1 (74.2→77.0), Terminal-Bench 4.0 (31.2→42.3), HLE (39.2→42.6) and Apex Bench v1.1 (64.6→72.1); outputs are shorter and tasks finish faster, cutting per-task cost by ~60%. Native visual input, 1M context, compressed KV cache, MIT-licensed open weights, and the top open-weight spot on Arena Code Arena WebDev (1670).

Meta's multimodal reasoning model released on September 3, 2026, built for long-running agentic, multi-agent and coding workflows. It natively accepts text, images, video and PDF documents and returns text, with a 1M-token context window (up to ~944K output tokens), adjustable reasoning effort, tool calling and structured outputs. Cached input is $0.15/M.

A lightweight multimodal version of Zhipu GLM-5.3 with a 1M context window, supporting image, video, file, and text inputs, with free cache storage for a limited time.

The second model in the Claude 5.5 family, released by Anthropic on September 28, 2026. A clear upgrade over Sonnet 5, it runs more than 30% faster and cuts costs by up to 30% on most tasks. Positioned as a faster, lower-cost complement to Opus 5.5, it excels at well-scoped everyday tasks, bug fixes, and document/slide/spreadsheet creation.

Anthropic's October 2026 lightweight flagship with a 1M-token context window. OpenRouter pricing of $0.10/$0.50 per million tokens is roughly a tenth of Haiku 4.5 ($1/$5), and it entered OpenRouter's fastest-growing token ranking at #2 in its first week — one of the cheapest million-context closed models available.

Anthropic's latest flagship Opus-tier model, released on September 22, 2026, matches Claude Fable 5.1 on most tasks while costing about 40% less to run than Opus 5. It focuses on long-horizon, highly autonomous agents, building/debugging/refactoring/code review for large-scale codebases, and end-to-end enterprise workflows; it supports cross-session memory learning, sub-agent orchestration, and Fast mode (up to 2.5x speed).

Xiaomi's October 2026 lightweight model with a 1.05M-token context window at $0.14/$0.28 per million tokens. It drew 10.9T tokens on OpenRouter in a week (#4) and entered the monthly ranking at #7, making it one of the highest-usage million-context models and one of the cheapest.

StepFun's October 2026 preview model with a 1M-token context window, priced at $1.00/$2.70 per million tokens. It hit #2 on OpenRouter's daily token ranking (2.59T) and led the weekly growth ranking at +309%, making it the fastest-rising Chinese model right now.

Zhipu AI's new-generation flagship model released on August 14, 2026, with 743B parameters. Built on the same base as GLM-5.2, its coding capability is improved by 50% over GLM-5.2 through post-training scaling (IndexShare, SAO, and the open-source Slime reinforcement learning framework). It ranks first among open-source models on Terminal Bench 3.0 and Agents' Last Exam (CLI), with agent capabilities on par with Kimi K3, Claude Opus 4.8, and Claude Fable 5. It is now available to all users of ZCode, AutoClaw, and the GLM Coding Plan; full weights will be open-sourced within two weeks, and the API is coming soon.

Anthropic's September 2026 update; Fable 5 delivers more than 2x the performance of Fable 5 (Frontier-Bench). Supports a new "adaptive reasoning" mode for handling complex multi-step tasks. The default model for Claude Max/Team Premium.

A 2.8T-parameter frontier flagship model released by Moonshot AI in July 2026, featuring an 896-expert MoE architecture with 16 experts activated per token, and unified pricing for 1M context with no tiered surcharges. Its KDA hybrid attention mechanism delivers 6.3x decoding speedup; its front-end programming capability ranks first globally, and its coding and reasoning match Claude Opus at half the cost. Open weights released on July 27, 2026.

The newest member of OpenAI's GPT-6.1 generation at $2.00/$10.00 per million tokens with a 1.05M-token context, positioned as the balanced mid-tier of the GPT-6.1 line (above Luna, below the Pro tier).

OpenAI's high-end GPT-6 tier at $10/$50 per million tokens with a 1M context, aimed at the most complex reasoning-heavy workloads where call volume is not the driver.

Mistral AI's October 2026 flagship with a 1M-token context window at $0.68/$2.09 per million tokens. It focuses on European compliance and price-performance, and reached #3 on OpenRouter's fastest-growing token ranking in its debut week.

A new-generation model family released by OpenAI on July 9, 2026, in three tiers: the flagship Sol, the mainline Terra, and the lightweight Luna. Luna is designed for low cost and high throughput, with about 62% fewer factual errors than GPT-5.5 Instant. Permanent price cuts take effect from July 30: Luna down 80%, Terra down 20%; Sol adds a Fast mode (double the price, 2.5x the speed).

The mid-tier of OpenAI's GPT-5.6 generation at $2.00/$10.00 per million tokens with a 1.05M-token context — the same price band as GPT-6.1 Sol and GPT-6 Sol.

Qwen3.8-Max is Alibaba Cloud's flagship multimodal model from the Qwen3.8 generation, released in August 2026. It uses a 2.4T total-parameter architecture with 95B active parameters (A95B) and a 1M-token context window, with maximum output of 65,536 tokens. Core capabilities include native vision, function calling, web search, context caching and structured output. Priced at 2.0 USD per million input tokens and 6.0 USD per million output tokens, it targets professional programming, office work, enterprise agents and long document processing, distinguishing itself from peers through its large sparse architecture and extended context.

OpenAI's newest ultra-light flagship (GPT-6 generation), priced at $0.10/$0.50 per million tokens with a 1.05M-token context — half the price of GPT-5.6 Luna ($0.20/$1.20) at the same context size. It is currently the cheapest million-context model in OpenAI's lineup.

A new-generation flagship model released and open-sourced by MiniMax on June 1, 2026, featuring a self-developed MSA sparse attention architecture (not MoE), 1M context, and native multimodality. SWE-Bench Pro 59.0% surpasses GPT-5.5, and BrowseComp 83.5 points surpasses Claude Opus 4.7. One week after release, it ranked third globally in OpenRouter call volume.

DeepSeek flagship MoE model (model version DeepSeek-V4-Pro-0813), supporting 1M context and 384K max output, with thinking mode enabled by default, and supporting JSON output, tool calling, Responses API, and Anthropic API-compatible format.

Released by SpaceXAI (formerly xAI) on July 9, 2026, built on a 1.5-trillion-parameter V9 base and jointly trained with Cursor. Token efficiency is 2x that of competitors, inference speed is 80 TPS, and SWE-bench Pro reaches 64.7%, surpassing GPT-5.5. Pricing is highly aggressive, at only 1/4 the cost of Opus.

The upper mid-tier of OpenAI's GPT-5.6 generation at $2.00/$12.00 per million tokens with a 1.05M-token context, sitting between Sol ($2/$10) and the Pro tier.

Released by Google on July 21, 2026, successor to 3.5 Flash. 17% fewer output tokens, 2x faster (303 tok/s), knowledge cutoff March 2026, built-in Computer Use.

A new-generation flagship large language model released and open-sourced by Tencent on August 28, 2026. MoE architecture with 770B total parameters / 49B activated, 1M context (max input 960K, max output 64K), focused on Agent capabilities of "planning, execution, and reliable delivery." Open-source preview version (Preview), not a GA stable release; now available on Tencent Cloud TokenHub and OpenRouter.

NVIDIA's October 2026 open-weight MoE model (550B total / 55B active). It drew 6.38T tokens on OpenRouter in a week (#7) and 22.5T over a month (+31%), and a free endpoint with a 1M context makes it a cost-effective pick for long-context batch workloads.

OpenAI's smallest and cheapest model at $0.05/$0.40 per million tokens with a 400K-token context — the lowest input price in the OpenAI lineup and roughly 40x cheaper than GPT-5.4 Mini on input.

Moonshot AI's coding model for code scenarios, following instructions more reliably in long contexts with a higher success rate on programming tasks; supports text, image, and video input, thinking mode only, 256K context.

Alibaba Cloud Bailian's new-generation lightweight model with native support for million-token context, suited for low-cost processing of ultra-long documents, large code repositories, and complex conversations.

Google's 2026 flagship model with 1M-token context and advanced multimodal capabilities. Released on July 15, 2026, focusing on enterprise applications and developer tools.

A cost-effective ERNIE model on Baidu Qianfan, with a 128K context; batch inference and prepaid volume packages can further reduce costs.

ByteDance's flagship large model released on February 14, 2026, achieving gold-medal level in IMO/CMO math competitions and ICPC programming contests, surpassing Gemini 3 Pro on the Putnam benchmark. All-around multimodal, scientific reasoning, and Agent capabilities, with over 100 million monthly active users.

Google's newest frontier model, released October 14, 2026, built for the hardest real-world tasks in coding, finance, law, cybersecurity and science. It ranks #1 on the Vals Index for economically valuable work, supports up to 1,048,576 output tokens sharing a single 1M-token context window with input and thinking, and cuts harmful-request violation rates by nearly half versus the previous Gemini generation.

Meta's flagship frontier model released October 13, 2026, which Mark Zuckerberg says puts the company back at the AI frontier. Meta claims it outperforms Anthropic's Claude Opus 5.5 and Fable 5.1 and OpenAI's GPT-6 Astra on some technical benchmarks, with strong advanced coding and improved multi-turn memory. It is closed-source, with the strongest tier sold directly to enterprises.

The lightweight, high-efficiency member of OpenAI's GPT-6.1 generation at $1.00/$5.00 per million tokens ($0.05 cached input) with a 1.05M-token context, built for visual reasoning, precise computer use and clear writing at half the price of GPT-5.6 Terra.

MiniMax's lightweight multimodal MoE model (428B total / ~23B active parameters), generally available on 2026-10-19. Native text, image and video input with a 1M-token context and up to 524K output tokens; reasoning is always on with five tunable effort levels, aimed at coding agents, computer use and security work. Flat $0.15/$0.60 per million tokens (no peak pricing, $0.015 cached input), about half of M3's base rate, with community-licensed open weights.

Alibaba Cloud's mid-tier closed-source multimodal model in the Qwen3.7 line (snapshot qwen3.7-plus-2026-05-26), generally available on Bailian since June 2026. Accepts text, image and video input, can read screens and drive GUIs, and writes code from visual references; 1M-token context (max input 991,808, max output 131,072) with thinking mode, web search and context caching. Billed in input-length tiers: up to 256K ¥2 input / ¥8 output per million tokens (currently 20% off, ¥1.6/¥6.4, cached hit ¥0.32); above 256K ¥6 / ¥24 (20% off ¥4.8/¥19.2).

Gemini 3.1 Pro Preview is Google's frontier reasoning model, built on a multimodal foundation and positioned for demanding, complex workflows. It offers a 1M-token context window and delivers enhanced software engineering performance, improved agentic reliability, and more efficient token usage. Priced at 2.0 USD per million input tokens and 12.0 USD per million output tokens, it suits advanced reasoning, coding, and agent-driven tasks.

GPT-5.2 is OpenAI's latest frontier-grade model in the GPT-5 series, delivering stronger agentic and long-context performance than GPT-5.1. It uses adaptive reasoning to allocate computation dynamically, responding quickly while scaling effort as needed. The model supports a 400K-token context window. Priced at 1.75 USD per million input tokens and 14.0 USD per million output tokens, it targets demanding agentic and long-context workloads.

GPT-5.1 is OpenAI's latest frontier-grade model in the GPT-5 series, delivering stronger general-purpose reasoning, improved instruction adherence, and a more natural conversational style than GPT-5. It uses adaptive reasoning and supports a 400K-token context window. Priced at 1.25 USD per million input tokens and 10.0 USD per million output tokens, it suits demanding general-purpose reasoning, instruction-following, and conversational workloads.

GPT-5.4 is OpenAI's latest frontier model, unifying the Codex and GPT lines into a single system. It offers a 1M+ token context window with 922K input and 128K output capacity, supporting long-document and large-codebase workloads. Priced at 2.5 USD per million input tokens and 15.0 USD per million output tokens, it targets demanding coding, reasoning, and agentic tasks within a unified architecture.

GPT-5.5 is OpenAI's frontier model built for complex professional workloads, succeeding GPT-5.4 with stronger reasoning, higher reliability, and improved token efficiency on hard tasks. It offers a 1.1M token context window and is priced at 5.0 USD per million input tokens and 30.0 USD per million output tokens, positioning it as a premium-tier option for demanding, long-context work.

GLM-5.1 is a Zhipu AI model that delivers a major leap in coding capability, with particularly significant gains in handling long-horizon tasks. Unlike previous models built around minute-level interactions, it can work independently and continuously on extended tasks. It offers a 205K token context window. Priced at 1.4 USD per million input tokens and 4.4 USD per million output tokens, it suits sustained coding and long-running autonomous work.

GLM-4.6 is Zhipu AI's successor to GLM-4.5, expanding the context window from 128K to 200K tokens to handle more complex inputs. It targets long-context reasoning and generation workloads, with pricing at 0.5 USD per million input tokens and 2.0 USD per million output tokens, positioning it as a cost-conscious option for extended-context tasks.

Grok 4.3 is a reasoning model from SpaceXAI. It accepts text and image inputs and produces text output, and is positioned for agentic workflows, instruction-following tasks, and applications requiring high factual accuracy. The model supports a 1M-token context window. Pricing is set at 1.25 USD per million input tokens and 2.5 USD per million output tokens, positioning it for cost-sensitive deployments at scale.

Gemini 3.5 Flash is Google's high-efficiency multimodal model, delivering near-Pro-level coding and reasoning at Flash-tier cost and speed. It is highly optimized for coding proficiency and parallel agentic execution, and supports a 1M-token context window. Priced at 1.5 USD per million input tokens and 9.0 USD per million output tokens, it targets cost-sensitive, high-throughput workloads.

Gemini 3.8 Flash is Google's most intelligent Flash model, delivering significant gains over 3.7 Flash in software engineering, agentic tasks, and multi-step reasoning. It offers a 1M-token context window and is priced at 0.825 USD per million input tokens and 4.125 USD per million output tokens, positioning it as a capable, cost-conscious option for demanding workloads. Typical uses include coding, agentic workflows, and complex multi-step reasoning.

Claude Opus 4.8 is Anthropic's most capable generally available model in the Opus family. It accepts text, image, and file inputs and produces text output, with reasoning support and a 1M-token context window. Priced at 5.0 USD per million input tokens and 25.0 USD per million output tokens, it is positioned as a premium flagship offering for demanding workloads that require strong reasoning and large-context processing.

Kimi K2.6 is Moonshot AI's next-generation multimodal model, positioned for long-horizon coding, coding-driven UI/UX generation, and multi-agent orchestration. It handles complex end-to-end coding tasks across Python, Rust, and Go within a 262K-token context window. Pricing is set at 0.65 USD per million input tokens and 3.41 USD per million output tokens, supporting sustained agentic and development workflows.