Provider Models and Limits

As of: 2026-09-05 Owner: the LLxprt Code maintainers Companion page: Provider Setup Quick Reference

This page lists model names, context windows, and recommended runtime tuning for the providers LLxprt Code supports. Model names, context windows, pricing, and provider-side limits change frequently. The values here reflect what is configured in LLxprt Code's built-in provider aliases as of the date above.

Before relying on a limit or price, verify it against the provider's own documentation. Use the links in each section to check for the most current data. LLxprt Code cannot guarantee that a provider has not changed a model name, context window, or price since this page was last reviewed.

Default models by alias

The default model each built-in alias uses when you run /provider <alias> without an explicit /model. Values reflect the alias configuration as of the date above. This is the canonical dated copy; the Provider Setup Quick Reference links here rather than duplicating it.

Alias Default model
anthropic claude-opus-5
claudecode claude-opus-5
gemini gemini-2.5-pro
openai gpt-5.5
codex gpt-5.6-sol
qwen qwen3-coder-plus
kimi kimi-for-coding
xai grok-4
deepseek deepseek-v4-flash
zai glm-5
synthetic hf:zai-org/GLM-4.7
chutes-ai zai-org/GLM-5-TEE
makora nvidia/Kimi-K2.6-NVFP4
fireworks fireworks/minimax-m3
openrouter nvidia/nemotron-nano-9b-v2
cerebras-code qwen-3-coder-480b
mistral mistral-large-latest
litellm gpt-4o
ollama-cloud kimi-k2.6
lm-studio gemma-3b-it
llama-cpp local-model

How model limits are resolved

Default model limits are data-driven and layered. Each layer takes precedence over the one below it:

  1. User override — /set context-limit <N> or a profile ephemeralSettings.context-limit. Always wins.
  2. Provider alias configuration — built-in defaults shipped with LLxprt Code, including per-model contextWindow and context-limit values.
  3. Core fallback catalog — a built-in safety net for models not covered by an alias config. The default fallback limit is 200,000 tokens.

You can override any limit at runtime:

/set context-limit 200000
/set modelparam max_tokens 4096

Note: context-limit must always exceed max_tokens / maxOutputTokens. A configured max_tokens >= context-limit is rejected as an impossible configuration. For models whose catalog context window is already 200,000, setting /set context-limit 200000 adds no headroom over the default — you may want to leave the default or increase it only if the provider's actual window is larger.

Compression and context budgeting

The automatic compression trigger fires when currentTokens >= compressionThreshold × (context-limit − completionBudget). This means:

  • Lowering context-limit increases spend superlinearly. Halving the limit does not halve the trigger point — it cuts it much more steeply, because the completion budget is subtracted from a smaller base. More compressions means more paid LLM summarization calls and more prompt-cache prefix rewrites.
  • max_tokens >= context-limit is rejected. An explicitly configured completion budget that meets or exceeds the context window leaves zero prompt budget and is now rejected with an error rather than silently triggering compression on every send.
  • compression.strategy=high-density mutates history continuously (every turn, not just at the threshold) and is hostile to prompt caching. Use middle-out (the default) when cache reuse matters.

Model geometry and budgeting

context-limit and max_tokens describe a single shared window, not two independent budgets:

  • context-limit — the total token budget for the entire request (prompt plus output). This is the ceiling the engine enforces.
  • max_tokens (also surfaced as maxOutputTokens) — the completion budget held inside context-limit, reserved for the model's response. It is subtracted from the context limit, not added to it.
  • Effective prompt budget ≈ context-limit − max_tokens − safety margin.

Because the completion budget sits inside the limit, context-limit=200000 with max_tokens=100000 is a valid configuration: it simply leaves roughly 100,000 tokens of prompt budget before the safety margin is applied. The two values never add together to claim a larger window than context-limit.

The engine automatically applies a fixed safety margin of 1,000 tokens plus a small percentage headroom (~0.5%) on top of the completion budget. You do not configure this margin; it exists to absorb overhead from tool wrappers, the system prompt, and project memory files.

Tip: If you see "would exceed the token context window" errors, lower max_tokens first (to reclaim prompt budget) or reduce the size of your project memory files.

Auth-variant note: Context windows can differ between API-key access and OAuth/subscription access for the same model. When in doubt, start lower and increase until you hit a provider limit error.

Anthropic (Claude)

Anthropic documentation · Anthropic models

Anthropic API key (anthropic alias)

Configured context-limit for current-generation models (claude-opus-5, claude-opus-4-8, claude-fable-5-1, claude-fable-5, claude-sonnet-4-6, claude-sonnet-5): 1,000,000 tokens.

Max output tokens configured: 128,000.

Reasoning is enabled by default for claude-(opus|sonnet|haiku|fable) models, with reasoning.effort set to high only for claude-opus-5, claude-opus-4-8, claude-fable-5-1, claude-fable-5, claude-sonnet-4-6, and claude-sonnet-5. Other matching models (for example, claude-haiku-4-5) get reasoning enabled without a default effort. Temperature, top_p, and top_k are disallowed for all claude-(opus|sonnet|haiku|fable) models (reasoning models manage sampling internally).

Common models: claude-opus-5, claude-sonnet-5, claude-sonnet-4-6, claude-haiku-4-5.

Note: The 1,000,000-token context is the value configured in the alias. Anthropic may gate large context windows by plan or credit tier. Check Anthropic's documentation for your account's actual limits.

Recommended settings:

/set context-limit 200000
/set modelparam max_tokens 4096

Profile JSON (API key):

{
  "version": 1,
  "provider": "anthropic",
  "model": "claude-opus-5",
  "modelParams": { "max_tokens": 4096 },
  "ephemeralSettings": { "context-limit": 200000 }
}

Environment variable: export ANTHROPIC_API_KEY=sk-ant-...

Summarized thinking (not a streaming bug): Anthropic returns a summary of the model's reasoning, produced by a different model than the one that generated the raw thinking. Billing reflects the full raw thinking, so the visible thinking text is typically far shorter than the billed completion tokens (a median of roughly 4 visible characters per billed token). A summary occasionally ends mid-sentence at a clean word boundary — this is a summarizer artefact, not a truncation or streaming bug. LLxprt records and displays exactly what the API returns.

Claude Code OAuth (claudecode alias)

Same context and output limits as the anthropic alias for current-generation models. Available static models include: claude-opus-5, claude-fable-5-1, claude-fable-5, claude-opus-4-8, claude-opus-4-7, claude-opus-4-6, claude-sonnet-5, claude-sonnet-4-6, and earlier versions.

See Provider Setup Quick Reference for OAuth setup instructions.

Pricing

Pricing for Anthropic models is not tracked on this page. Check Anthropic's pricing page for current rates.

OpenAI (API)

OpenAI documentation · OpenAI models

OpenAI API key (openai alias)

For gpt-5.6 models, the configured context-limit is 1,050,000 tokens with max output tokens of 128,000. Reasoning is enabled by default with reasoning.effort set to high.

GPT-5.x reasoning models do not support temperature, top_p, top_k, frequency_penalty, or presence_penalty. Use /set reasoning.effort instead.

Reasoning effort values: minimal, low, medium, high, xhigh, max.

Common models: gpt-5.6, gpt-5.5.

Note: The actual context window for your model depends on your OpenAI API plan. The configured limit is a default; check OpenAI's documentation for your model's actual window.

Recommended settings:

/set context-limit 400000    # adjust to your model's actual window
/set modelparam max_tokens 8192
/set reasoning.effort high

OpenAI Codex OAuth (codex alias)

The codex alias uses the ChatGPT subscription backend. Its provider-level context-limit is 262,144 tokens, while gpt-6-astra has a per-model OAuth context-limit of 872,000 tokens.

Model OAuth context limit Reasoning effort values
gpt-6-astra 872,000 low, medium, high, xhigh, max
gpt-5.6-sol (default) 262,144 minimal, low, medium, high, xhigh, max
gpt-5.6-terra, gpt-5.6-luna 262,144 minimal, low, medium, high, xhigh, max
gpt-5.3-codex-spark 131,072 Provider-dependent

Common Codex models: gpt-6-astra, gpt-5.6-sol (default), gpt-5.6-terra, gpt-5.6-luna, gpt-5.5, gpt-5.3-codex-spark.

See Provider Setup Quick Reference for OAuth setup instructions.

Pricing

Pricing for OpenAI models is not tracked on this page. Check OpenAI's pricing page for current rates.

Google Gemini

Google AI documentation · Gemini models

Gemini API key (gemini alias)

The core fallback catalog lists these context windows for Gemini 2.5 models:

  • gemini-2.5-pro: 1,048,576 tokens
  • gemini-2.5-flash: 1,048,576 tokens
  • gemini-2.5-flash-lite: 1,048,576 tokens

Note: The maximum context window for Gemini models may depend on your API plan and tier. Check Google's documentation for your account's actual limits.

Common models: gemini-2.5-pro, gemini-2.5-flash, gemini-2.5-flash-lite.

Recommended settings:

/set context-limit 1048576
/set modelparam max_tokens 4096

Profile JSON:

{
  "version": 1,
  "provider": "gemini",
  "model": "gemini-2.5-pro",
  "modelParams": { "temperature": 0.7, "max_tokens": 4096 },
  "ephemeralSettings": { "context-limit": 1048576 }
}

Environment variable: export GEMINI_API_KEY=...

Pricing

Pricing for Gemini models is not tracked on this page. Check Google AI's pricing page for current rates.

Qwen (DashScope)

DashScope documentation

Qwen API key (qwen alias)

Configured context-limit for qwen3-coder-plus: 1,000,000 tokens. Max output tokens: 65,536.

Common models: qwen3-coder-plus.

Note: The qwen alias is for Qwen's own DashScope service. It is not used for Cerebras.

Recommended settings:

/set context-limit 200000
/set modelparam max_tokens 4096

Profile JSON:

{
  "version": 1,
  "provider": "qwen",
  "model": "qwen3-coder-plus",
  "modelParams": { "temperature": 0.7, "max_tokens": 4096 },
  "ephemeralSettings": { "context-limit": 200000 }
}

Environment variable: export DASHSCOPE_API_KEY=...

Pricing

Pricing for Qwen models is not tracked on this page. Check Alibaba Cloud's DashScope documentation for current rates.

Kimi (Moonshot AI)

Moonshot AI documentation

The kimi alias supports two paths:

  • kimi-for-coding — the subscription-served model. Context window: 262,144 tokens. Max output: 32,768 tokens. This is the alias default.
  • kimi-k3 — the pay-per-token model on the raw Moonshot API. Context window: 1,048,576 tokens. Max output: 131,072 tokens.

To use kimi-k3, point the alias at the raw Moonshot endpoint:

/provider kimi
/baseurl https://api.moonshot.ai/v1
/keyfile ~/.moonshot_key
/model kimi-k3

Recommended settings (kimi-for-coding)

/set context-limit 262144
/set modelparam max_tokens 32768
/set reasoning.effort medium

Recommended settings (kimi-k3)

/set context-limit 1000000
/set modelparam max_tokens 131072
/set reasoning.effort max

Profile JSON (pay-per-token kimi-k3):

{
  "version": 1,
  "provider": "kimi",
  "model": "kimi-k3",
  "modelParams": { "max_tokens": 131072 },
  "ephemeralSettings": {
    "context-limit": 1000000,
    "base-url": "https://api.moonshot.ai/v1",
    "reasoning.effort": "max",
    "reasoning.enabled": true,
    "reasoning.includeInResponse": true
  }
}

Multimodal support (Kimi)

The kimi alias declares media capabilities:

Capability Status (per alias config) Notes
Inline images Enabled Images in user messages and tool responses flow as image_url content parts.
PDF/document upload Enabled Base64 PDFs are uploaded with purpose: file-extract and referenced by file id, keeping large documents out of the token budget.
Video Experimental (off by default) Base64 videos are uploaded with purpose: video, then sent as video_url content with an ms://<file-id> URL on Kimi/Moonshot endpoints.

To enable experimental video forwarding:

/set kimi.experimental-video true

This is gated behind both the kimi.experimental-video setting and the alias's mediaSupport.videoSupport capability flag. Third-party aliases (Chutes, Synthetic) do not declare this capability.

Pricing

Pricing for Kimi/Moonshot models is not tracked on this page. Check Moonshot AI's pricing page for current rates.

Other API-key providers

The following providers use the OpenAI-compatible protocol. Some ship a configured context window for their default model; others do not. Where a window is not configured, the core fallback catalog provides a default limit of 200,000 tokens, and you should verify the true window against the provider's documentation.

Provider Alias Default model Configured context window (as of the date above)
xAI (Grok) xai grok-4 Not configured — 200,000 fallback applies. Check xAI's docs.
OpenRouter openrouter nvidia/nemotron-nano-9b-v2 Not configured — 200,000 fallback applies. Check OpenRouter's model docs.
Fireworks fireworks fireworks/minimax-m3 minimax-m3: 1,000,000 tokens, reasoning high.
Cerebras Code cerebras-code qwen-3-coder-480b Not configured — 200,000 fallback applies. Check Cerebras's docs.
DeepSeek deepseek deepseek-v4-flash deepseek-v4*: 1,000,000 tokens, reasoning high.
Z.AI zai glm-5 glm-5.2: 1,000,000 tokens, reasoning high. The default glm-5 has reasoning high but no explicit window — it falls back to 200,000.
Makora makora nvidia/Kimi-K2.6-NVFP4 Not configured — 200,000 fallback applies.
Synthetic synthetic hf:zai-org/GLM-4.7 Not configured — 200,000 fallback applies.
Chutes AI chutes-ai zai-org/GLM-5-TEE Not configured — 200,000 fallback applies.
Mistral mistral mistral-large-latest Not configured — 200,000 fallback applies. Check Mistral's docs.

For providers without a configured window, start with a conservative limit and adjust based on the provider's documentation:

/set context-limit 200000
/set modelparam max_tokens 4096

The core fallback catalog provides a default limit of 200,000 tokens for models not explicitly listed.

Local models

Context windows for local models (LM Studio, llama.cpp, Ollama) depend entirely on your local runtime and model build. Start small and increase:

/set context-limit 32000
/set modelparam max_tokens 2048

See Using Local Models for complete guidance.

Related