CLAUDE CODE · OPUS 4.5 XHIGH & OPUS 5.5 HYBRID INFRASTRUCTURE

Stretch Claude Code limits 6x to 10x.

Keep Claude Opus for top-level architecture (~15% of tokens) and route 85% of sub-agent execution, multi-file edits, vision OCR, and 24/7 background tasks to our 18-model /v1 gateway + Qwen 3.8 27B lane. Dedicated NVMe server included.

View AI Coding Plans (from $30/mo) Compare Benchmarks & Pairing

OpenAI-Compatible /v1 24/7 Qwen 3.8 27B Dedicated NVMe VPS Included

AI Coding Plans

One OpenAI-compatible /v1 key for the full 30-model catalog: up to 1M-token context, a free 24/7 Qwen 3.8 27B coding lane, and an NVMe server included on every plan. Tokens reset every month.

Go

$30/ month

Up to 125M tokens, resets monthly

API benefits

Up to 125M tokens every month
All models, one key
Heavier models draw faster
Up to 1M-token context

Compute

4 GB RAM · 2 vCPU VPS
Full root on Windows or Ubuntu

Plus

$60/ month

Up to 250M tokens, resets monthly

API benefits

Up to 250M tokens every month
All models, one key
Heavier models draw faster
Up to 1M-token context

Compute

16 GB RAM · 8 vCPU VPS
Full root on Windows or Ubuntu

Pro

$125/ month

Up to 500M tokens, resets monthly

API benefits

Up to 500M tokens every month
All models, one key
Heavier models draw faster
Up to 1M-token context

Compute

32 GB RAM · 16 vCPU VPS
Full root on Windows or Ubuntu

Ultra

$250/ month

Up to 1B tokens, resets monthly

API benefits

Up to 1B tokens every month
All models, one key
Heavier models draw faster
Up to 1M-token context

Compute

64 GB RAM · 32 vCPU VPS
Full root on Windows or Ubuntu

Max

$500/ month

Up to 2B tokens, resets monthly

API benefits

Up to 2B tokens every month
All models, one key
Heavier models draw faster
Up to 1M-token context

Compute

128 GB RAM · 64 vCPU VPS
Full root on Windows or Ubuntu

Token packs one-time top-up · never expires · no subscription

1B tokens$250 one-timeBuy 1B

2B tokens$500 one-timeBuy 2B

Every plan includes an OpenAI-compatible key, monthly token reset, and a dedicated VPS. Rate stays locked in while active.

Included models

The Ultimate Fleet

30 models in the fleet - 23 live now, 7 coming soon. OpenAI-compatible, up to 1M-token context, with a server on every plan.

DeepSeek V4 Pro

Flagship 1T deep-reasoning model for architecture, scientific logic, and hard debugging.

DeepSeek V4 Pro 0813

Pinned 0813 DeepSeek V4 Pro checkpoint for deterministic long-running agent deployments.

DeepSeek V4.1 Flash 1017

High-throughput 1017 coding workhorse (90.6% Terminal-Bench 2.1, 384K max output).

DeepSeek Flash

Ultra-low-latency DeepSeek tier for rapid routing, triage, and high-concurrency sub-agents.

Qwen 3.8 Max

1T flagship for multimodal UI/OCR understanding, competition math, and algorithmic synthesis.

Qwen 3.8 Flash

Fast 397B multimodal throughput model for code, vision OCR, and streaming workflows.

Qwen 3.8 27B

Dedicated 24/7 unmetered builder lane (0% monthly token burn) on every plan.

GLM 5.3

Strict JSON schema adherence, reliable multi-tool calling, and multimodal agent control.

GLM 5.3 Flash

High-speed structured extraction & fast tool-calling GLM 5.3 tier.

GLM 5.2

Stable 355B previous-generation GLM lane for legacy structured prompts.

GLM 5.1

Proven GLM 5.1 baseline preserved for backward-compatible pipelines.

Kimi K3

Whole-repo ingestion, video + image + text comprehension, and formal proofs.

Kimi K3.1 Pro

Next-gen Kimi K3.1 Pro flagship for deep agentic planning and long-context coding.

Kimi K3.1 Pro Highspeed

Accelerated Kimi K3.1 Pro lane tuned for interactive speed in IDEs and swarms.

MiniMax-M3

Live 428B MoE agentic core (23B active) tuned for sustained multi-turn execution.

MiniMax-M3.1 Flash

Ultra-fast M3.1 Flash generation for rapid tool-calling and responsive sub-agents.

Doubao Seed 2.1 Pro

ByteDance 3.2T multimodal flagship: 98.3% AIME, 94.2% tau2-Bench, 89.5% Video-MME.

Doubao Seed 2.1 Turbo

High-speed Seed 2.1 tier trading peak size for fast interactive throughput.

Doubao Seed 2.1 Lite

Lightweight Seed 2.1 lane for fast classification, extraction, and scalable fan-out.

Doubao Seed 2.0 Pro

Proven Seed 2.0 Pro multimodal tier maintained live for compatibility.

Doubao Seed 2.0 Code

Code-specialized Doubao Seed model tuned for repository editing and tool chains.

Doubao Seed 2.0 Lite

Fast Seed 2.0 Lite workhorse for high-volume extraction and classification.

Doubao Seed 2.0 Mini

Ultra-light Seed 2.0 Mini for millisecond routing, parsing, and lightweight bots.

Kimi-2.7 & earlier

Legacy Moonshot K2/2.7 family queued for backward-compatible evaluation runs.

HY-4

Tencent Hunyuan 770B enterprise reasoning model (49B active MoE) queued next.

StepFun Step 5 Preview

StepFun 600B flagship preview (27B active) queued for high-throughput reasoning.

StepFun-3.7 and earlier

Earlier StepFun 3.7 / Step-3 Flash generations queued for compact fast lanes.

Qwen3.5-397b-a17b

Earlier Qwen sparse 397B (17B active) multimodal checkpoint queued for eval continuity.

MIMO (V2-Flash / Pro)

Xiaomi MiMo V2 Flash & Pro sparse reasoning series queued for upcoming activation.

MiniMax-M2.7 & earlier

Legacy MiniMax M2.7 10B-active lineage queued alongside live MiniMax-M3.

View model details

One OpenAI-compatible /v1 key reaches every model. Context up to 1M tokens, varies by model.

VERIFIED BENCHMARKS & TOKEN ECONOMICS

Claude Opus vs. BlueLeaf Hosting

Side-by-side coding benchmarks, output speeds, and monthly token economics across our 23-model gateway and proprietary flagships.

Category	Benchmark / Metric	BlueLeaf Hosting 23 Live Models + Free 27B + VPS	Claude Opus 4.5 (Xhigh) $5.00 / $25 MTok	Claude Opus 5.5 $4.00 / $20 MTok	Claude Fable 5.1 $10.00 / $50 MTok	GPT-6 Astra $4.00 / $16 MTok	Gemini 4 Argon $2.50 / $10 MTok
General & Science	AA Index v5.0Intelligence Index	34 (#1 in Fleet: Qwen 3.8 Max · #1 Open-Weights: 1017 at 32.5)	48	58	49	57	50
General & Science	GPQA Diamond & RiemannGraduate science & math logic	92.4% GPQA (deepseek-v4-pro · 90.9% 1017)	87.0% (GPQA)	69.6% (Riemann)	65.6% (Riemann)	72.0% (Riemann)	76.0% (Riemann)
Agentic Coding	DeepSWE v1.1 / FrontierSWE 2Multi-file repo engineering	74.2% (DeepSWE) (deepseek-v4.1-flash-1017)	59.0% (DeepSWE)	62.3% (FrontierSWE 2)	56.3% (FrontierSWE 2)	65.5% (FrontierSWE 2)	55.0% (FrontierSWE 2)
	LiveCodeBenchAlgorithmic & contest coding	93.5% (qwen3.8-max & deepseek-v4-pro)	88.8%	89.4%	82.4%	86.2%	84.0%
	Terminal-BenchCLI & shell execution	87.9% (v2.1) (V4 Pro · 26.8% v4.0 1017)	78.2% (v2.1)	59.6% (v4.0)	52.0% (v4.0)	59.1% (v4.0)	57.1% (v4.0)
Context & Speed	Context / Max Single OutputMax tokens generated per turn	1M / 384K Out (3x longer output than Claude)	1M / 128K	1M / 128K	1M / 128K	400K / 128K	1M / 65K
Context & Speed	Streaming ThroughputMedian output tokens / sec	175 - 300+ TPS (Live at status.blueleafhosting.com)	67 TPS	75 - 85 TPS	75 TPS	85 TPS	120 TPS
Economics & Compute	1B Monthly Tokens (70% Cached)90% Input / 10% Output blend	$250/mo Flat ($0.25/M effective · 0 overages)	$4,295/mo	$3,206/mo	$7,857/mo	$2,806/mo	$1,769/mo
	24/7 Builder LaneLinting, cron, bots, diffs	Included Free (qwen3.8-27b · 0 token burn)	None	None	None	None	None
	Dedicated NVMe VPS IncludedWindows Server 2022 or Ubuntu	4GB to 128GB RAM (2 to 64 vCPU AMD EPYC included)	$0 Hardware	$0 Hardware	$0 Hardware	$0 Hardware	$0 Hardware

View All 23 Live + 7 Queued Models Included With Every BlueLeaf Hosting Key (/v1/models)+

Canonical Model ID (accepted on /v1/chat/completions)	Best Role in Multi-Agent Pipeline	Context / Max Out	Modalities	token allowance Burn	Params
DeepSeek V4 Pro `deepseek-v4-pro`	Flagship 1T deep-reasoning model for architecture, scientific logic, and hard debugging.	1M / 384K	Text + Image	Pooled Monthly Bank	1T
DeepSeek V4 Pro 0813 `deepseek-v4-pro-0813`	Pinned 0813 DeepSeek V4 Pro checkpoint for deterministic long-running agent deployments.	1M / 384K	Text + Image	Pooled Monthly Bank	1T
DeepSeek V4.1 Flash 1017 `deepseek-v4.1-flash-1017`	High-throughput 1017 coding workhorse (90.6% Terminal-Bench 2.1, 384K max output).	1M / 384K	Text + Image	Pooled Monthly Bank	671B
DeepSeek Flash `deepseek-flash`	Ultra-low-latency DeepSeek tier for rapid routing, triage, and high-concurrency sub-agents.	1M / 128K	Text + Image	Pooled Monthly Bank	671B
Qwen 3.8 Max `qwen3.8-max`	1T flagship for multimodal UI/OCR understanding, competition math, and algorithmic synthesis.	1M / 131K	Text + Image	Pooled Monthly Bank	1T
Qwen 3.8 Flash `qwen3.8-flash`	Fast 397B multimodal throughput model for code, vision OCR, and streaming workflows.	1M / 131K	Text + Image	Pooled Monthly Bank	397B
`qwen3.8-27b` (aliased as `local` — 24/7 Builder Lane)	Dedicated 24/7 unmetered builder lane (0% monthly token burn) on every plan.	128K / 32K	Text	0% BURN (INCLUDED)	27B
GLM 5.3 `glm-5.3`	Strict JSON schema adherence, reliable multi-tool calling, and multimodal agent control.	1M / 128K	Text + Image	Pooled Monthly Bank	400B
GLM 5.3 Flash `glm-5.3-flash`	High-speed structured extraction & fast tool-calling GLM 5.3 tier.	1M / 128K	Text + Image	Pooled Monthly Bank	355B
GLM 5.2 `glm-5.2`	Stable 355B previous-generation GLM lane for legacy structured prompts.	1M / 128K	Text	Pooled Monthly Bank	355B
GLM 5.1 `glm-5.1`	Proven GLM 5.1 baseline preserved for backward-compatible pipelines.	1M / 128K	Text	Pooled Monthly Bank	355B
Kimi K3 `kimi-k3`	Whole-repo ingestion, video + image + text comprehension, and formal proofs.	1M / 128K	Text + Image + Video	Pooled Monthly Bank	1T
Kimi K3.1 Pro `kimi-k3.1-pro`	Next-gen Kimi K3.1 Pro flagship for deep agentic planning and long-context coding.	1M / 128K	Text + Image + Video	Pooled Monthly Bank	1T
Kimi K3.1 Pro Highspeed `kimi-k3.1-pro-highspeed`	Accelerated Kimi K3.1 Pro lane tuned for interactive speed in IDEs and swarms.	1M / 128K	Text + Image + Video	Pooled Monthly Bank	1T
MiniMax-M3 `minimax-m3`	Live 428B MoE agentic core (23B active) tuned for sustained multi-turn execution.	1M / 128K	Text + Image	Pooled Monthly Bank	428B
MiniMax-M3.1 Flash `minimax-m3.1-flash`	Ultra-fast M3.1 Flash generation for rapid tool-calling and responsive sub-agents.	1M / 128K	Text + Image	Pooled Monthly Bank	428B
Doubao Seed 2.1 Pro `seed-2-1-pro`	ByteDance 3.2T multimodal flagship: 98.3% AIME, 94.2% tau2-Bench, 89.5% Video-MME.	1M / 128K	Text + Image + Video	Pooled Monthly Bank	3.2T
Doubao Seed 2.1 Turbo `seed-2-1-turbo`	High-speed Seed 2.1 tier trading peak size for fast interactive throughput.	1M / 128K	Text + Image	Pooled Monthly Bank	1.2T
Doubao Seed 2.1 Lite `seed-2-1-lite`	Lightweight Seed 2.1 lane for fast classification, extraction, and scalable fan-out.	1M / 128K	Text + Image	Pooled Monthly Bank	400B
Doubao Seed 2.0 Pro `seed-2-0-pro`	Proven Seed 2.0 Pro multimodal tier maintained live for compatibility.	1M / 128K	Text + Image	Pooled Monthly Bank	2.4T
Doubao Seed 2.0 Code `seed-2-0-code`	Code-specialized Doubao Seed model tuned for repository editing and tool chains.	1M / 128K	Text + Image	Pooled Monthly Bank	680B
Doubao Seed 2.0 Lite `seed-2-0-lite`	Fast Seed 2.0 Lite workhorse for high-volume extraction and classification.	1M / 128K	Text + Image	Pooled Monthly Bank	320B
Doubao Seed 2.0 Mini `seed-2-0-mini`	Ultra-light Seed 2.0 Mini for millisecond routing, parsing, and lightweight bots.	1M / 128K	Text	Pooled Monthly Bank	120B
`Kimi-2.7 & earlier` COMING SOON	Legacy Moonshot K2/2.7 family queued for backward-compatible evaluation runs.	1M / TBA	Text + Image	Not live yet	1T
`HY-4` COMING SOON	Tencent Hunyuan 770B enterprise reasoning model (49B active MoE) queued next.	1M / TBA	Text + Image	Not live yet	770B
`StepFun Step 5 Preview` COMING SOON	StepFun 600B flagship preview (27B active) queued for high-throughput reasoning.	1M / TBA	Text	Not live yet	600B
`StepFun-3.7 and earlier` COMING SOON	Earlier StepFun 3.7 / Step-3 Flash generations queued for compact fast lanes.	128K / TBA	Text	Not live yet	196B
`Qwen3.5-397b-a17b` COMING SOON	Earlier Qwen sparse 397B (17B active) multimodal checkpoint queued for eval continuity.	1M / TBA	Text + Image	Not live yet	397B
`MIMO (V2-Flash / Pro)` COMING SOON	Xiaomi MiMo V2 Flash & Pro sparse reasoning series queued for upcoming activation.	1M / TBA	Text	Not live yet	309B
`MiniMax-M2.7 & earlier` COMING SOON	Legacy MiniMax M2.7 10B-active lineage queued alongside live MiniMax-M3.	1M / TBA	Text	Not live yet	230B

HYBRID INTEGRATION ARCHITECTURE · 15% PRIMARY / 85% SUB-AGENT SPLIT

How to Pair BlueLeaf Hosting (/v1) With Your Existing Setup

BREAK-EVEN: ~90M TOKENS/MO VS OPUS · ~35M VS FABLE

STEP 01 // ARCHITECTURAL SPLIT

Stretch Your Primary Quota 6x - 10x

Keep your primary model (Claude Opus, Fable 5.1, GPT-6 Astra, or Gemini) for top-level planning, ambiguous debugging, and final review (~15% of tokens). Offload 85% of sub-agent execution to your BlueLeaf Hosting token allowance and 100% of background loops to the qwen3.8-27b lane.

STEP 02 // TASK-SPECIFIC ROUTING

Route Each Job to Its Best Lane

Send multi-file edits to deepseek-v4.1-flash-1017 (74.2% DeepSWE, 384K output) or glm-5.3 for strict JSON. Send vision OCR to multimodal qwen3.8-max or kimi-k3, and 24/7 cron/linting to text-only qwen3.8-27b ($0 burn). Escalate to your primary model after 2 failed test runs.

STEP 03 // CLIENT & IDE WIRING

Zero-Conflict Setup in Minutes

Cline, Roo Code, DeepSeek Harness (dsh), and Hermes accept https://api.blueleafhosting.com/v1 natively. In Cursor (where the OpenAI Base URL override is global for Chat), run Cline/Roo/DSH inside Cursor for per-model routing. In Claude Code, point ANTHROPIC_BASE_URL at https://api.blueleafhosting.com/anthropic directly, or route sub-agents via LiteLLM or MCP.

OPTION A // CLAUDE CODE DIRECT /anthropic OR LITELLM PROXY

# 1. Direct Claude Code endpoint (zero proxy needed)
export ANTHROPIC_BASE_URL="https://api.blueleafhosting.com/anthropic"
export ANTHROPIC_AUTH_TOKEN="sk-ebot-your-founders-key"
export ANTHROPIC_MODEL="deepseek-v4-pro"
claude

# 2. Or keep Claude Opus for planning & route sub-agents (litellm_config.yaml):
model_list:
  - model_name: ebot-swe-fast
    litellm_params:
      model: openai/deepseek-v4.1-flash-1017
      api_base: https://api.blueleafhosting.com/v1
      api_key: os.environ/EBOT_API_KEY
  - model_name: ebot-math-vision
    litellm_params:
      model: openai/qwen3.8-max
      api_base: https://api.blueleafhosting.com/v1
      api_key: os.environ/EBOT_API_KEY
  - model_name: ebot-27b
    litellm_params:
      model: openai/qwen3.8-27b
      api_base: https://api.blueleafhosting.com/v1
      api_key: os.environ/EBOT_API_KEY

OPTION B // DIRECT OPENAI /v1 DROP-IN (CLINE · ROO · DSH · HERMES · CURSOR)

export OPENAI_BASE_URL="https://api.blueleafhosting.com/v1"
export OPENAI_API_KEY="sk-ebot-your-founders-key"

# Verified Hybrid ROI Formula (90% In / 10% Out, 70% Cached):
# Monthly Savings = 0.85 * (Monthly_MTok) * (Primary_Blended_$/MTok) - Ebot_Plan
# • Claude Fable 5.1 ($10/$50): $7,857/mo -> $1,429 Hybrid (Save $6,428/mo · 82%)
# • Claude Opus 4.5 ($5/$25): $4,295/mo -> $894 Hybrid (Save $3,401/mo · 79%)
# • Claude Opus 5.5 ($4/$20): $3,206/mo -> $731 Hybrid (Save $2,475/mo · 77%)
# • Every plan includes a dedicated 4GB - 128GB AMD EPYC NVMe VPS.

Every layer runs on hardware we control.

Your dedicated RTX PRO 6000 & AMD EPYC™ environment comes pre-configured with DeepSeek Harness, Hermes, and OpenClaw out of the box. Enjoy dedicated privacy, live state snapshots, and zero rate limits across your optimized AgentOS.

Preloaded Agent Stack

Windows or Ubuntu, preloaded with OpenWebUI, DeepSeek Harness, Hermes, and OpenClaw. Log in and deploy your swarms.

Local Tokens

Burst past 10,000 TPS on RTX PRO 6000 hardware. No rate limits, no API queues, no per-token fees.

Live Snapshots

Agent state is fragile. Automatic environment snapshots let you roll back to a known-good state when a swarm crashes.

Bare-Metal Isolation

Your weights, your prompts, your data. Bare-metal isolation means no shared tenants and no third-party access to your workloads.

Long-Context Memory

ECC memory paired with GPU acceleration, sized to hold up to 1M-token context windows in memory.

100 Gbps Backbone

Direct uplinks to 100 Gbps transit. Your agents can crawl, pull, and ingest data at scale without saturating the pipe.

PCIe Gen4 NVMe

Up to 14,000 MB/s reads and writes. Model weights load fast, and storage stops being the bottleneck.

Zen 4 Architecture

AMD EPYC™ 9274F processors with up to 48 cores, 4.3 GHz boost clocks, and large L3 cache for parallel agent workloads.

Grandfathered Pricing

The price, speeds, and limits on your plan today stay locked in while active. No forced migrations. No surprise changes.

99.99% Uptime SLA

Backed by redundant power and 100 Gbit Ashburn transit in our enterprise datacenter so your fleet stays online 24/7.

BIOS & OS Tuned

Internal OS optimization scripts and custom BIOS settings, tuned for system, GPU, and network throughput.

Rolling Model Upgrades

When new frontier models ship, we add them to your cluster. No migration and no plan change required.

01

One flat monthly price

Your tokens reset every month. No surprise per-token invoice at the end of the month.

02

No token cap on builder

Qwen 3.8 27B runs all day on a dedicated lane that never draws down your monthly tokens.

03

Server included

Full root on Windows or Ubuntu with NVMe storage. Your own machine, not a shared container.

04

One key, every model

A single OpenAI-compatible /v1 endpoint with up to 1M-token context across the live catalog.

Frequently Asked Questions.

Setup, model access, infrastructure, and support, answered before you pick a plan.

What is included with an AI Coding or AgentOS plan?

Each plan combines dedicated NVMe server resources (Windows Server 2022 or Ubuntu), a 24/7 Qwen 3.8 27B coding lane, and an OpenAI-compatible API key for all 23 live frontier models.

How do I connect this to Claude Code, Cursor, or Codex?

Cline, Roo Code, DeepSeek Harness (dsh), and Hermes connect directly via our OpenAI-compatible https://api.blueleafhosting.com/v1 URL. Inside Cursor, you can either enable the OpenAI Base URL override for Chat or run Cline/Roo/DSH alongside Cursor's native Tab & Agent models. For Claude Code, set ANTHROPIC_BASE_URL=https://api.blueleafhosting.com/anthropic directly, or route sub-agents via our LiteLLM proxy config or MCP.

Can I choose Windows or Ubuntu for my included server?

Yes. Every server can be delivered on Windows Server 2022 or Ubuntu LTS with full Administrator or root access.

How does model access change between tiers?

All five AI Coding Plans include all 23 live models and up to 1M-token context. Higher tiers scale your monthly token allowance (from 125M up to 2B tokens/month) and double your dedicated VPS RAM and vCPU cores.

How do monthly token resets work on 3-month and 6-month plans?

Your token allowance resets every 30 days on all billing cycles, while 3-month terms save 20% and 6-month terms save 25% off the monthly rate. The discount applies to price only - your token bank still resets every month.

Can you build a custom GPU or multi-server configuration?

Yes. Dedicated GPUs, multi-GPU systems, blade deployments, and custom token allowances can be scoped directly with our team. Start a custom quote.

Who handles support?

Our engineering team handles support directly: provisioning, infrastructure, networking, and account questions.

Need something custom?

Legacy machine, custom token allocation, multi-GPU cluster, or full cabinet. Talk directly with the team about capacity, networking, and deployment.

FULL AGENTOS PRIVATE DEDICATED ENVIRONMENT > TEMPORARY SANDBOX

Every Plan Includes a Full Private Dedicated NVMe VPS — Not an Ephemeral Sandbox

Skip throwaway 24-hour test sandboxes. Starting at $30/mo (Go), your subscription provisions a dedicated bare-metal isolated Windows Server 2022 or Ubuntu LTS NVMe VPS preloaded with OpenWebUI, DeepSeek Harness, Hermes, OpenClaw, persistent project memory, and a 24/7 unmetered qwen3.8-27b lane.

Deploy Private AgentOS ($30/mo) → Inspect All Models