borg Download

Borg model directory

Models in Borg

This directory lists the live models available in Borg, grouped by provider so you can compare their published specifications and find the right lane. Every listed model is included in one Borg subscription; Borg Fusion routes between models for each step of a task, while you can also choose a lane directly. Benchmark figures, where shown, come from vendor publications linked on the model pages, not from an independent ranking. Borg speed figures are our own gateway measurements, refreshed weekly. Open a model page for its sources; unknown specifications are marked Not published.

How subscription pricing compares · Compare models

Borg gateway · 30-day median

Fastest first token (gateway proxy)

  1. Gemini 2.5 Flash-Lite776 ms
  2. Claude Haiku 4.5889 ms
  3. Grok 4.71037 ms
  4. Gemini 3.1 Flash-Lite1141 ms
  5. Grok 4.31457 ms

Fastest tokens per second (whole turn)

  1. Gemini 2.5 Flash142.4 output tokens/s (whole turn)
  2. Gemini 3.1 Flash-Lite112.9 output tokens/s (whole turn)
  3. Gemini 3 Flash107.6 output tokens/s (whole turn)
  4. Gemini 2.5 Pro95.3 output tokens/s (whole turn)
  5. Claude Sonnet 5.576.7 output tokens/s (whole turn)

Based on gateway measurements; first token is the first streamed response event, and tokens per second includes the whole turn.

Search by model name, lane id or use case.

Anthropic models

Anthropic develops the Claude model family for writing, coding, and analysis. Borg offers access to Claude models alongside other providers through its own routing layer.

6 models carried · 6 reasoning models · largest published context window: 1,000,000 tokens Vendor source

ModelContextReasoningReleaseBest forBorg speedBench
Claude Fable 5.11,000,000 tokens Vendor sourceYes—long-horizon coding agents tackling difficult repositories55.9 output tokens/s (whole turn)—
Claude Opus 5.51,000,000 tokens Vendor sourceYes2026-09-22 Vendor sourcelong-running agentic coding across repositories60 output tokens/s (whole turn)—
Claude Sonnet 5.51,000,000 tokens Vendor sourceYes2026-09-28 Vendor sourcewell-scoped bug fixes and code reviews76.7 output tokens/s (whole turn)—
Claude Haiku 4.5200,000 tokens Vendor sourceYes2025-10-15 Vendor sourcehigh-volume classification and tagging64.9 output tokens/s (whole turn)—
Claude Opus 4.61,000,000 tokens Vendor sourceYes2026-02-05 Vendor sourcemaintaining workflows calibrated to Opus 4.635.5 output tokens/s (whole turn)—
Claude Sonnet 4.61,000,000 tokens Vendor sourceYes2026-02-17 Vendor sourcecomputer-use workflows in office applications46.4 output tokens/s (whole turn)—

OpenAI models

OpenAI develops the GPT family and open-weight gpt-oss models. Borg exposes supported GPT lanes for direct selection as well as routed use.

9 models carried · 9 reasoning models · largest published context window: 1,050,000 tokens Vendor source

ModelContextReasoningReleaseBest forBorg speedBench
GPT-5.51,050,000 tokens Vendor sourceYes2026-04-24 Vendor sourcedeep document analysis24.9 output tokens/s (whole turn)—
GPT-5.6 Luna1,050,000 tokens Vendor sourceYes2026-07-09 Vendor sourcehigh-volume classification17.2 output tokens/s (whole turn)73%
GPT-5.6 Sol1,050,000 tokens Vendor sourceYes2026-07-09 Vendor sourcecomplex professional reasoning14.3 output tokens/s (whole turn)—
GPT-5.6 Terra1,050,000 tokens Vendor sourceYes2026-07-09 Vendor sourcedocument review with judgment18.7 output tokens/s (whole turn)—
GPT-6 Astra1,050,000 tokens Vendor sourceYes2026-09-03 Vendor sourcehard coding assignments16 output tokens/s (whole turn)67%
GPT-6 Luna1,050,000 tokens Vendor sourceYes2026-09-22 Vendor sourcebulk classification15.3 output tokens/s (whole turn)—
GPT-6 Sol1,050,000 tokens Vendor sourceYes2026-09-22 Vendor sourceagentic coding across repos11.3 output tokens/s (whole turn)—
GPT-6.1 Sol1,050,000 tokens Vendor sourceYes2026-09-29 Vendor sourcecoding with tools17.7 output tokens/s (whole turn)—
gpt-oss-120b131,072 tokens Vendor sourceYes2025-08-05 Vendor sourceprivate self-hosted reasoning46.8 output tokens/s (whole turn)—

Google models

Google develops Gemini models across Pro, Flash, and Flash-Lite tiers. Borg includes several Gemini generations and gateway effort variants.

12 models carried · 12 reasoning models · largest published context window: 1,048,576 tokens Vendor source

ModelContextReasoningReleaseBest forBorg speedBench
Gemini 2.5 Flash1,048,576 tokens Vendor sourceYes2025-04-17 Vendor sourcemultimodal document and video summaries142.4 output tokens/s (whole turn)—
Gemini 2.5 Flash-Lite1,048,576 tokens Vendor sourceYes2025-06-17 Vendor sourcecheap high-volume classification62.2 output tokens/s (whole turn)—
Gemini 2.5 Pro1,048,576 tokens Vendor sourceYes2025-03-25 Vendor sourcelong-repository code analysis95.3 output tokens/s (whole turn)—
Gemini 3 Flash1,048,576 tokens Vendor sourceYes2025-12-17 Vendor sourcevisual and spatial reasoning107.6 output tokens/s (whole turn)—
Gemini 3.1 Flash Image131,072 tokens Vendor sourceYes2026-02-26 Vendor sourcesearch-grounded image creation——
Gemini 3.1 Flash-Lite1,048,576 tokens Vendor sourceYes2026-03-03 Vendor sourcelarge-scale translation112.9 output tokens/s (whole turn)—
Gemini 3.1 Pro1,048,576 tokens Vendor sourceYes2026-02-19 Vendor sourcecomplex software engineering——
Gemini 3.5 Flash1,048,576 tokens Vendor sourceYes2026-05-19 Vendor sourceiterative coding agents——
Gemini 3.5 Flash-Lite1,048,576 tokens Vendor sourceYes2026-07-21 Vendor sourcehigh-throughput document parsing59.2 output tokens/s (whole turn)—
Gemini 3.6 Flash1,048,576 tokens Vendor sourceYes2026-07-21 Vendor sourcerapid agentic coding loops56.8 output tokens/s (whole turn)—
Gemini 3.7 Flash1,048,576 tokens Vendor sourceYes2026-08-13 Vendor sourcevideo-informed agent workflows72.8 output tokens/s (whole turn)—
Gemini 3.8 Flash1,048,576 tokens Vendor sourceYes2026-09-02 Vendor sourcelong-horizon coding agents29.2 output tokens/s (whole turn)—

xAI models

xAI develops Grok models for reasoning, coding, and multimodal tasks. Borg exposes available Grok lanes alongside other model families.

4 models carried · 4 reasoning models · largest published context window: 1,000,000 tokens Vendor source

ModelContextReasoningReleaseBest forBorg speedBench
Grok 4.31,000,000 tokens Vendor sourceYes—high-volume instruction-following tasks31.9 output tokens/s (whole turn)—
Grok 4.5500,000 tokens Vendor sourceYes2026-07-08 Vendor sourceagentic coding with tools5.9 output tokens/s (whole turn)—
Grok 4.6500,000 tokens Vendor sourceYes2026-08-12 Vendor sourcemulti-step codebase changes13.6 output tokens/s (whole turn)—
Grok 4.7500,000 tokens Vendor sourceYes2026-09-21 Vendor sourcelong-running agentic coding33 output tokens/s (whole turn)—

Moonshot AI models

Moonshot AI develops the Kimi family of models. Borg carries Kimi lanes for general and coding work.

2 models carried · 2 reasoning models · largest published context window: 1,048,576 tokens Vendor source

ModelContextReasoningReleaseBest forBorg speedBench
Kimi K31,048,576 tokens Vendor sourceYes2026-07-16 Vendor sourcelong-horizon coding across large repositories18.1 output tokens/s (whole turn)—
Kimi K2.7 Code262,144 tokens Vendor sourceYes—multi-file coding with iterative test runs24.2 output tokens/s (whole turn)—

Z.ai models

Z.ai develops the GLM family of language models. Borg offers a selection of GLM generations for direct use and routing.

9 models carried · 9 reasoning models · largest published context window: 1,048,576 tokens Vendor source

ModelContextReasoningReleaseBest forBorg speedBench
GLM-4.5131,072 tokens Vendor sourceYes2025-07-28 Vendor sourcetool-assisted software engineering——
GLM-4.5-Air131,072 tokens Vendor sourceYes2025-07-28 Vendor sourcecost-conscious coding-agent steps12.1 output tokens/s (whole turn)—
GLM-4.6200K tokens (vendor-stated) Vendor sourceYes2025-09-30 Vendor sourcelarge-repository code review——
GLM-4.7200K tokens (vendor-stated) Vendor sourceYes2025-12-22 Vendor sourceend-to-end application prototypes40.6 output tokens/s (whole turn)—
GLM-5200K tokens (vendor-stated) Vendor sourceYes2026-02-12 Vendor sourcemulti-stage systems engineering33.3 output tokens/s (whole turn)—
GLM-5-Turbo200K tokens (vendor-stated) Vendor sourceYes—OpenClaw agent automation16.4 output tokens/s (whole turn)60%
GLM-5.1200K tokens (vendor-stated) Vendor sourceYes2026-04-07 Vendor sourcelong-running autonomous coding——
GLM-5.21,048,576 tokens Vendor sourceYes2026-06-16 Vendor sourcewhole-repository architecture audits25.7 output tokens/s (whole turn)47%
GLM-5.31,048,576 tokens Vendor sourceYes2026-08-18 Vendor sourcecomplex coding-agent tasks25.2 output tokens/s (whole turn)53%

Meta models

Meta develops AI models including the Muse family. Borg lists Muse Spark lanes, whose public names require separate verification.

3 models carried · 3 reasoning models · largest published context window: 1,048,576 tokens Vendor source

ModelContextReasoningReleaseBest forBorg speedBench
Muse Spark 1.11,048,576 tokens Vendor sourceYes2026-07-09 Vendor sourcecomputer-use agents spanning several applications——
Muse Spark 1.21,048,576 tokens Vendor sourceYes2026-08-05 Vendor sourcemulti-file refactors in large repositories——
Muse Spark 1.31,048,576 tokens Vendor sourceYes2026-09-02 Vendor sourcelong-horizon coding with changing requirements26.2 output tokens/s (whole turn)—

Borg models

Borg brings multiple model providers into one subscription. Its Fusion routing profiles select an underlying model for each step, while Atlas Max provides a high-capacity lane.

5 models carried · 0 reasoning models · largest published context window: Not published

ModelContextReasoningReleaseBest forBorg speedBench
Borg Fusion———multi-step tasks——
Borg Fusion Coding———multi-step coding work——
Borg Fusion Design———multi-step design work——
Borg Fusion Fast———multi-step time-sensitive tasks——
Atlas Max———multi-step demanding, high-capacity tasks——