The AI Loadout
Get Daily Dispatch
🧠 DIRECTORY 03

AI Models Directory: Frontier and Local

Frontier cloud models from every major lab, plus open-weight models you can run on your own computer. Prices and specs checked against official vendor pages.

New to local models? Start with how to run AI locally, or pick the Local filter below.

AnthropicContext: 1M tokens (128K max output)

Claude Opus 5.5

Hardest coding and agentic tasks, large refactors, long-horizon work in Claude Code.

Benchmark Notes: Per GenAI Brief (Oct 2026): first on Terminal-Bench 4.0 at 64.8% (Claude Code, max effort) and tied first on DeepSWE v1.1 at 72.3%. Leads the Arena WebDev board at 1813 per Fello AI.
In: $4.00 / Out: $20.00Model Specs →
AnthropicContext: 1M tokens (128K max output)

Claude Sonnet 5.5

Everyday coding, writing and drafting at half the price of Opus 5.5.

Benchmark Notes: Per Fello AI (Oct 2026): leads Artificial Analysis' Terminal-Bench 4.0 run at 63.6% and LiveBench Coding at 91.4. Anthropic halved its cache read price to $0.10 per 1M tokens on Oct 7, 2026.
In: $2.00 / Out: $10.00Model Specs →
AnthropicContext: 1M tokens (128K max output)

Claude Haiku 5.5

High-volume work: summarization, subagents and browser use.

Benchmark Notes: Anthropic calls it its cheapest, fastest and most capable small model to date (announcement, Oct 7, 2026).
In: $0.10 (prompts up to 100K) / Out: $0.50 (prompts up to 100K)Model Specs →
OpenAIContext: 1.05M tokens (128K max output)

GPT-6.1 Sol

Best value agentic coding, especially through Codex.

Benchmark Notes: Per GenAI Brief (Oct 2026): ties Claude Opus 5.5 on DeepSWE v1.1 at 72.3% and scores 58.2% on Terminal-Bench 4.0 (Codex, max) at roughly one seventh of the cost per trial.
In: $2.00 / Out: $10.00Model Specs →
GoogleContext: 1M tokens (unified input and output)

Gemini 4 Argon

Google's flagship reasoning model for complex coding, agentic, math and science work.

Benchmark Notes: Released Oct 13, 2026. Gemini API paid tier only (not on the free tier or the Google AI Pro API plan); in the Gemini app for Google AI Pro and Ultra subscribers. Generally available in GitHub Copilot (Pro+, Max, Business, Enterprise). Trails the top Claude models on Terminal-Bench per GenAI Brief.
In: $2.00 (intro, $4.00 from Dec 1) / Out: $10.00 (intro, $20.00 from Dec 1)Model Specs →
OpenAIContext: 1.05M tokens (922K max input, 128K max output)

GPT-6 Astra

OpenAI's most capable model: complex reasoning, coding, computer use, research and document creation.

Benchmark Notes: Per OpenAI: prompts over 272K input tokens cost 2x input and 1.5x output for the whole request; Batch and Flex are half price; Fast mode is 2x. Reasoning effort low to max. In ChatGPT it powers GPT-6 Pro (Pro, Business, Enterprise) and is included in Work and Codex on Plus. Also on Microsoft Azure and Amazon Bedrock.
In: $10.00 / Out: $50.00Model Specs →
OpenAIContext: 1.05M tokens (128K max output)

GPT-6 Luna

Cheap, high-volume work with a clear goal: extraction, classification, structured summaries, subagents.

Benchmark Notes: OpenAI's most efficient GPT-6 model. Cached input $0.01 per 1M tokens.
In: $0.10 / Out: $0.50Model Specs →
AnthropicContext: 1M tokens (128K max output)

Claude Fable 5.1

Anthropic's top model for demanding reasoning and long-horizon agentic work, when Opus 5.5 at higher effort still falls short.

Benchmark Notes: Adaptive thinking is always on, default effort high. Cache reads cost 2.5% of the input price.
In: $10.00 / Out: $50.00Model Specs →
GoogleContext: 1,048,576 tokens (65,536 max output)

Gemini 3.8 Flash

Google's fast, low-cost everyday model, with a free tier in the Gemini API.

Benchmark Notes: Introductory price through Dec 31, 2026, then $1.50 in and $7.50 out per 1M tokens (output includes thinking tokens).
In: $0.75 (intro) / Out: $3.75 (intro)Model Specs →
xAIContext: 500K tokens

Grok 4.7

xAI's flagship for coding, agentic tool use and knowledge work, with built-in web and X search.

Benchmark Notes: Per xAI docs: prompts over 200K tokens cost $4 in and $12 out; cached input $0.50. Reasoning effort low to xhigh (default high). Knowledge cutoff May 2026. Available in Cursor on all plans.
In: $2.00 / Out: $6.00Model Specs →
MetaContext: 1,048,576 tokens

Muse Spark 1.3

Meta's hosted flagship, tuned for agentic workflows (multi-step tool, browser and long-horizon tasks), with text, image, video, audio and PDF input.

Benchmark Notes: Cached input $0.15. A cheaper Contributor mode ($0.10 in, $0.20 out) lets Meta use your prompts and outputs for training.
In: $1.25 / Out: $4.25Model Specs →
MistralContext: 1M tokens

Mistral Large 4

Mistral's open-weight multimodal flagship for general work, with function calling and agents.

Benchmark Notes: Public preview from Oct 6, 2026. Mixture of experts with 52B active of 1.05T total parameters. Launch pricing is 50% off for two weeks. Open weights announced as coming soon.
In: $1.36 ($0.68 launch sale) / Out: $4.18 ($2.09 launch sale)Model Specs →
DeepSeekContext: 1M tokens (384K max output)

DeepSeek V4 Pro

DeepSeek's strongest API model for reasoning and coding at a low price. No vision.

Benchmark Notes: Thinking and non-thinking modes. Off-peak prices are half of peak. DeepSeek V4.1 Flash (model name deepseek-flash) is the cheaper option at $0.30 in and $1.20 out at peak, with vision.
In: $1.32 peak / $0.66 off-peak / Out: $3.96 peak / $1.98 off-peakModel Specs →
Alibaba QwenContext: 1M tokens

Qwen3.8 Max

Alibaba's flagship for engineering-scale coding, multi-tool agents and vision.

Benchmark Notes: International endpoint prices shown; global-scope endpoints list $1.65 in and $4.951 out. Thinking and non-thinking modes. Snapshot qwen3.8-max-0902.
In: $2.00 / Out: $6.00Model Specs →
MoonshotContext: 1,048,576 tokens

Kimi K3

Moonshot's open-weight flagship for long-horizon coding and knowledge work.

Benchmark Notes: Always reasons; effort low, high or max. Cached input $0.30. 2.8 trillion parameters with full weights released.
In: $3.00 / Out: $15.00Model Specs →
Zhipu (Z.ai)Context: 1M tokens (128K max output)

GLM-5.3

Z.ai's open-weight flagship for agentic coding and repository-scale work. Text only.

Benchmark Notes: Reasoning always on (low, high, max). Cached input $0.26. Also on Amazon Bedrock (launched Oct 5, 2026) and hosted by Mistral.
In: $1.40 / Out: $4.40Model Specs →
AmazonContext: See Amazon Bedrock docs

Amazon Nova Premier

Amazon's most capable Nova model on Amazon Bedrock, for teams already on AWS that want complex reasoning and agent work billed with their AWS account.

Benchmark Notes: On-demand price in US East (N. Virginia) per the Amazon Bedrock pricing page. Nova 2 Lite is the cheaper option at $0.30 in and $2.50 out. Nova 2 Pro is in preview.
In: $2.50 / Out: $12.50Model Specs →
CohereContext: 128K tokens (64K max output)

Cohere Command A+

Enterprise agents, RAG, tool use and translation across 48 languages, on your own GPUs or via Cohere.

Benchmark Notes: Cohere's first mixture of experts model (command-a-plus-05-2026), with vision input and reasoning. Apache 2.0 license; runs on two H100s or one B200. Listed as Free on cohere.com/pricing.
In: Free on Cohere API (rate limited) / Out: Free on Cohere API (rate limited)Model Specs →
Local · GoogleContext: 128K (E2B, E4B) / 256K (12B, 26B, 31B)

Gemma 4 (local)

Best all-round local family: E4B for laptops, 26B A4B for strong reasoning and code on a 24 GB GPU or 32 GB Mac.

Benchmark Notes: Image input on every size; audio on E2B and E4B. Plan on RAM or VRAM of roughly the download size plus 2 to 4 GB for context.
ollama run gemma4:e4bModel Specs →
Local · Alibaba QwenContext: 256K

Qwen 3.5 (local)

Wide size range from 0.8B to 122B. The 9B is a strong laptop coder; 27B and 35B for workstation use.

Benchmark Notes: Vision, tools and thinking. Plan on RAM or VRAM of roughly the download size plus 2 to 4 GB for context.
ollama run qwen3.5:9bModel Specs →
Local · Alibaba QwenContext: 256K

Qwen 3.8 27B (local)

Newest Qwen open model for coding, research and long agentic tasks on a 24 to 32 GB machine.

Benchmark Notes: Vision, tools and thinking. Plan on RAM or VRAM of roughly the download size plus 2 to 4 GB for context.
ollama run qwen3.8Model Specs →
Local · OpenAIContext: 128K

gpt-oss (local)

OpenAI's open-weight reasoning models. 20B runs on a 16 GB GPU or Mac; 120B needs about 80 GB.

Benchmark Notes: Tools and thinking, text only. Plan on RAM or VRAM of roughly the download size plus 2 to 4 GB for context.
ollama run gpt-oss:20bModel Specs →
Local · MetaContext: 128K

Muse Glimmer 30B (local)

Meta's open model for always-on local agents: tool use, long tasks and failure recovery on a single GPU.

Benchmark Notes: Vision, tools and thinking. Plan on RAM or VRAM of roughly the download size plus 2 to 4 GB for context.
ollama run muse-glimmerModel Specs →
Local · MetaContext: 128K

Llama 3.3 70B (local)

Proven general model if you have 48 GB or more of memory. Older than Muse Glimmer but widely supported.

Benchmark Notes: Tools, text only. Plan on RAM or VRAM of roughly the download size plus 2 to 4 GB for context.
ollama run llama3.3Model Specs →
Local · MistralContext: 128K

Mistral Small 3.2 24B (local)

Solid general assistant with vision and reliable function calling on a 24 GB machine.

Benchmark Notes: Vision and tools. Plan on RAM or VRAM of roughly the download size plus 2 to 4 GB for context.
ollama run mistral-small3.2Model Specs →
Local · MistralContext: 384K

Devstral Small 2 24B (local)

Local coding agent: explores codebases, edits multiple files and drives tools.

Benchmark Notes: Vision and tools. Plan on RAM or VRAM of roughly the download size plus 2 to 4 GB for context.
ollama run devstral-small-2Model Specs →
Local · MistralContext: 256K

Ministral 3 (local)

Small edge models (3B, 8B, 14B) with vision for low-memory laptops.

Benchmark Notes: Vision and tools. Plan on RAM or VRAM of roughly the download size plus 2 to 4 GB for context.
ollama run ministral-3Model Specs →
Local · DeepSeekContext: 128K

DeepSeek-R1 distills (local)

Step-by-step reasoning and math on modest hardware. The 8B (R1-0528 on Qwen3) fits a 16 GB laptop.

Benchmark Notes: Tools and thinking. The full 671B model needs about 404 GB and is not practical locally. Plan on RAM or VRAM of roughly the download size plus 2 to 4 GB for context.
ollama run deepseek-r1:8bModel Specs →
AnthropicContext: 200,000 tokens

Claude 3.5 SonnetSuperseded

Autonomous software engineering, multi-file code editing, tool use orchestration, and complex architectural design.

Benchmark Notes: Industry benchmark leader on SWE-bench Verified (~49% to ~53%) and HumanEval coding evaluations.
In: $3.00 / Out: $15.00Model Specs →
AnthropicContext: 200,000 tokens

Claude 3.7 Sonnet (Hybrid Reasoning)Superseded

Complex engineering challenges requiring controllable extended thinking alongside rapid tool execution.

Benchmark Notes: Sets frontier high scores in agentic software repair and mathematical multi-step proofs with adjustable thinking budgets.
In: $3.00 / Out: $15.00Model Specs →
OpenAIContext: 128,000 tokens

GPT-4oSuperseded

Omni-modal voice workflows, vision inspection, fast general-purpose conversational queries, and low-latency API tasks.

Benchmark Notes: Strong multimodal vision performance with rapid token streaming latency.
In: $2.50 / Out: $10.00Model Specs →
OpenAIContext: 200,000 tokens

OpenAI o1Superseded

Deep competitive programming, complex algorithmic optimization, cryptography proofs, and scientific analysis.

Benchmark Notes: Top-tier performance on Codeforces competitions (89th percentile) and AIME mathematics competitions.
In: $15.00 / Out: $60.00Model Specs →
OpenAIContext: 200,000 tokens

OpenAI o3-miniSuperseded

Cost-effective reasoning, fast code verification, unit test generation, and structured JSON parsing.

Benchmark Notes: Exceeds o1-preview on STEM benchmarks with faster inference and low token pricing.
In: $1.10 / Out: $4.40Model Specs →
GoogleContext: 2,000,000 tokens

Gemini 1.5 ProSuperseded

Massive code repository ingestion, hour-long video analysis, book-length document synthesis, and multi-hour audio transcripts.

Benchmark Notes: World-record 2M token context window with near 100% needle-in-a-haystack retrieval accuracy.
In: $3.50 / Out: $10.50Model Specs →
GoogleContext: 1,000,000 tokens

Gemini 2.0 FlashSuperseded

Ultra-low latency web applications, real-time agent loops, high-volume classification, and multimodal streaming.

Benchmark Notes: Sub-second response time with native multimodal generation capabilities.
In: $0.10 / Out: $0.40Model Specs →
DeepSeekContext: 128,000 tokens

DeepSeek-R1Superseded

Open-weights reasoning, local offline deployment on high-RAM rigs, math reasoning, and verifiable coding problems.

Benchmark Notes: Open-weights reasoning model matching frontier proprietary benchmarks at a fraction of the inference cost.
In: $0.55 / Out: $2.19Model Specs →
DeepSeekContext: 128,000 tokens

DeepSeek-V3Superseded

High-throughput text generation, cost-sensitive code completion, summarization, and data extraction pipelines.

Benchmark Notes: 671B parameter Mixture-of-Experts model with high generation speed and affordable API rates.
In: $0.14 / Out: $0.28Model Specs →
MetaContext: 128,000 tokens

Llama 3.3 70B InstructSuperseded

Enterprise self-hosting, on-premise compliance workloads, local development via Ollama, and agent pipelines.

Benchmark Notes: Matches previous 405B model benchmark capabilities in a compact 70B footprint suitable for single-node hosting.
In: $0.60 / Out: $0.60Model Specs →
AnthropicContext: 200,000 tokens

Claude 3.5 HaikuSuperseded

Rapid coding completions, agent sub-routines, fast text transformations, and lightweight MCP tool calling.

Benchmark Notes: Matches original Claude 3 Opus benchmarks on coding tasks while running with significant speedups.
In: $0.80 / Out: $4.00Model Specs →