AI Models

TeXRA supports a variety of language models from different providers, allowing you to choose the best fit for your task's complexity, required speed, and budget (think of it as choosing your research assistant's brain!). This guide provides an overview of the models available by default.

Model Providers Overview

TeXRA primarily integrates with models from:

1. Anthropic (Claude family)
2. OpenAI (GPT and O-series families)
3. Google (Gemini family)
4. Other Providers (via OpenRouter, including Grok, DeepSeek)

You can select the desired model from the dropdown list in the TeXRA UI. Hovering over an option shows its provider, context window, and estimated cost.

Default Model Selection

Here's a quick comparison of the models available by default in TeXRA:

Anthropic Models

Known for strong instruction following and context handling.

Model ID	Key Strength / Use Case	Relative Cost	Relative Speed	Notes
opus41T	Latest Opus with explicit reasoning steps	$$$$	Slow	Claude 4.1 Opus with thinking
opus41	Latest high quality, complex tasks	$$$$	Slow	Claude 4.1 Opus
opus4T	Opus 4 with explicit reasoning steps	$$$$	Slow	Claude 4 Opus with thinking
opus4	Opus 4 high quality, complex tasks	$$$$	Slow	Claude 4 Opus
sonnet45T	Latest Sonnet with explicit reasoning steps	$$$	Medium	Claude 4.5 Sonnet with thinking
sonnet45	Latest strong all-rounder	$$$	Medium	Claude 4.5 Sonnet
sonnet4T	Sonnet 4 with explicit reasoning steps	$$$	Medium	Claude 4 Sonnet with thinking
sonnet4	Sonnet 4 strong all-rounder	$$$	Medium	Claude 4 Sonnet
sonnet37T	sonnet37 with explicit reasoning steps	$$$	Medium	Good for math, complex logic
sonnet37	Strong all-rounder, good context	$$$	Medium
sonnet35	Good balance of quality/cost (older Sonnet)	$$$	Medium
haiku45T	Fast Claude 4.5 with explicit reasoning	$$	Fast	Claude 4.5 Haiku with thinking
haiku45	Fast Claude 4.5 responses	$$	Fast	Claude 4.5 Haiku

Sonnet 4 / 4.5 1M Context (Beta)

To experiment with Anthropic's 1M-token context window for Sonnet 4 or 4.5, enable "texra.model.useAnthropic1MBeta": true in VS Code settings. The extension attaches the context-1m-2025-08-07 beta header for these requests. Only Sonnet 4-family models support this beta, and TeXRA still enforces the tier‑4 limit of 200 K tokens.

OpenAI Models

Known for strong reasoning and creative capabilities.

Model ID	Key Strength / Use Case	Relative Cost	Relative Speed	Notes
o1	Advanced reasoning, math, figures	$$$$	Slow	Explicit reasoning
gpt45	High quality, vision (Preview)	$$$$	Medium
gpt5pro	Premium reasoning & coding	$$$$	Slow	400k ctx, 272k max output
gpt5	Flagship reasoning & coding	$$$	Medium	400k context
gpt5-	Flagship mini, fast	$$	Fast	400k context, mini
gpt5--	Flagship nano, fastest	$	Very Fast	400k context, nano
gpt41	Long-context vision, powerful	$$$	Medium	1M tokens context
gpt41-	Long-context vision, cost-effective	$$	Medium	1M tokens context, mini
gpt41--	Long-context vision, cheapest	$	Medium	1M tokens context, nano
gpt4o	Strong all-rounder, vision	$$$	Medium	Good default choice
gpt4ol	Latest gpt4o, potentially better	$$$	Medium
o3	Coding, tool calling	$$$	Medium
o3pro	Reliable answers, heavy compute	$$$$	Slow	o3-pro
o3-	Fast reasoning	$$$	Fast	o3-mini
o1-	Fast reasoning (smaller o1)	$$$	Fast	o1-mini
gptoss	Open-weight reasoning, large context	$$	Medium	gpt-oss-120b (OpenRouter only)
gptoss-	Open-weight reasoning, cost-effective	$	Fast	gpt-oss-20b (OpenRouter only)

Note: GPT-5 and GPT-5 Pro reasoning summaries require additional account verification. TeXRA disables them by default—enable "texra.model.gpt5ReasoningSummary": true if your account supports this feature.

Google Models

Known for large context windows, multimodality, and speed/cost efficiency.

Model ID	Key Strength / Use Case	Relative Cost	Relative Speed	Notes
gemini25p	Strong reasoning, vision, large context	$$$	Medium	Latest Pro model
gemini2p	Good reasoning, vision, very large context	$$	Medium
gemini25f	Fast reasoning, large context	$$	Fast	Latest Flash model
gemini2fT	gemini2f with explicit reasoning steps	$	Fast
gemini2f	Fastest, most cost-effective, vision	$	Very Fast	Good for simple tasks, native PDF/audio

DeepSeek Models

Strong technical and coding performance, cost-effective. DeepSeek's API now supports function calling so agents can use external tools during a run.

Model ID	Key Strength / Use Case	Relative Cost	Relative Speed	Notes
dsv3	Good coding & general tasks	$	Fast	DeepSeek V3 Chat
dsr1	Advanced reasoning	$$	Medium	DeepSeek R1

Moonshot Kimi Models

High context models from Moonshot, suitable for complex reasoning and large documents.

Kimi K2 is Moonshot's open-source 1T‑parameter MoE model (32B active). It excels at coding and agentic tasks but currently lacks multimodal and thought-mode support. The 0905 preview offers a 256k context window, and a high-speed turbo variant is also available.

Model ID	Key Strength / Use Case	Relative Cost	Relative Speed	Notes
kimit	Detailed reasoning with vision	$$$	Medium	Kimi Thinking Preview
kimi	Large context, general tasks	$$	Medium	128k context
kimiv	Vision-enabled variant	$$	Medium	128k context, vision
kimi2	Agent tasks, 256k context	$$$	Medium	Kimi K2 0905 Preview (moonshotai/kimi-k2-0905)
kimi2turbo	Fast agent tasks	$$$$	Very Fast	Kimi K2 Turbo Preview (moonshotai/kimi-k2-turbo)

The earlier Kimi K2 0711 model remains available on OpenRouter as moonshotai/kimi-k2.

Additional resources: API – $0.15/million input tokens (cache hit), $0.60/million input tokens (cache miss), $2.50/million output tokens. Tech blog, Weights & code, GitHub.

DashScope Qwen Models

Cost-effective models from Alibaba with strong multilingual capabilities.

Model ID	Key Strength / Use Case	Relative Cost	Relative Speed	Notes
qwen3max	Flagship coding agent, 262k ctx	$$$	Medium	Qwen3-Max latest, no thinking mode
qwenplus	Hybrid thinking, 1M ctx + tool use	$$	Medium	Qwen Plus latest, enable_thinking
qwenturbo	Fast responses with optional thinking mode	$	Fast	Qwen Turbo, enable_thinking

Deep thinking models first stream their reasoning before the final answer. qwenplus and qwenturbo support this mode. Pass enable_thinking: true in the request body to turn it on; commercial tiers disable it by default. qwen3max always runs in non-thinking mode.

from openai import OpenAI
import os

client = OpenAI(
    api_key=os.getenv("DASHSCOPE_API_KEY"),
    base_url="https://dashscope-intl.aliyuncs.com/compatible-mode/v1",
)

resp = client.chat.completions.create(
    model="qwen-plus-latest",
    messages=[{"role": "user", "content": "Who are you?"}],
    extra_body={"enable_thinking": True},
)
print(resp.choices[0].message.reasoning_content)
print(resp.choices[0].message.content)

Use the thinking_budget parameter to cap how many tokens the reasoning step can consume.

Copilot Models

GitHub Copilot models are available through VS Code's built-in Language Model API. These models require user consent and sign in to GitHub Copilot.

Model ID	Key Strength / Use Case	Relative Cost	Relative Speed	Notes
copilot4o	Strong all-rounder, vision	$$	Medium	Uses GPT-4o backend

Grok / xAI Models

Large context models from xAI.

Model ID	Key Strength / Use Case	Relative Cost	Relative Speed	Notes
grok4	Very large context, strong reasoning	$$$	Medium	xAI Grok 4
grok3	Large context, alternative reasoning	$$$	Medium	xAI Grok 3
grok3-	Faster Grok 3 (mini)	$$	Fast	xAI Grok 3 Mini

Other Models (Available Primarily via OpenRouter)

These models are generally accessed by enabling OpenRouter in settings.

Model ID	Key Strength / Use Case	Provider	Relative Cost	Relative Speed
llama31	Strong open model, large context	Meta	$$$	Medium
qvq-72b	Strong multi-lingual	Qwen/Alibaba	$$	Medium

Relative Cost/Speed are estimates: $ = Low/Fast, $$$$ = High/Slow.

Choosing the Right Model

Consider these factors:

• Task Complexity: Simple corrections might only need a $/Fast model (gemini2f), while complex paper transformations benefit from $$$$/Slow models (opus, o1).
• Budget: Use cost indicators ($ - $$$$) to guide selection.
• Speed: If quick turnaround is needed, prefer Fast/Very Fast models.
• Special Capabilities: Do you need explicit reasoning (sonnet45T, sonnet37T, gemini2fT, o1, o3-, o1-, gptoss, gptoss-, dsr1), vision (gpt5, gpt5pro, gpt4o, gemini*), native PDF/audio (gemini*), or very large context (gemini*, gpt41, gpt5, gpt5pro)?

Experimentation is often key to finding the best model for your specific needs and writing style.

Model Configuration

You can customize which models appear in the TeXRA dropdown list via VS Code Settings (Ctrl+,). Search for texra.models and edit the JSON array. Here are the defaults:

Model Availability

The specific models available by default and their identifiers (sonnet45, gpt5, gpt5pro, etc.) are maintained by the TeXRA developers and may change in future updates based on new releases and performance evaluations.

"texra.models": [
  "gemini25p",
  "gemini25f",
  "opus41T",
  "sonnet45T",
  "sonnet4T",
  "gpt5",
  "gpt41",
  "deepseek",
  "deepseekT",
  "kimi2",
  "qwen3max",
  "grok4"
]

Instruction Polishing Model

TeXRA also uses a dedicated setting for polishing instruction text before an agent run. Set "texra.model.instructionPolishModel" to any short name from the enum (same as texra.models) to pick the model that handles this formatting step when Copilot is disabled. The default is sonnet45.

"texra.model.instructionPolishModel": "sonnet45"

This setting is independent from the dropdown list—use it to lock polishing to a stable model while you experiment with other agents. The setting accepts any model from the same enum as texra.models, ensuring you can only select valid models.

Using OpenRouter

To access models not directly integrated (like Llama or Qwen), find alternative pricing, or ensure access if a direct API key isn't available, you can use OpenRouter.

1. Get an OpenRouter API key.
2. Add the key using the TeXRA: Set API Key command (select OpenRouter).
3. Enable OpenRouter in VS Code Settings: "texra.model.useOpenRouter": true.

When enabled, TeXRA will route API calls for all compatible models (including Anthropic, OpenAI, Google, DeepSeek, Grok, etc., if supported by OpenRouter) through OpenRouter instead of their direct APIs.

Streaming Support

For long responses or reasoning-heavy models, you can enable streaming to see incremental results. This is often more robust for complex tasks.

Configure streaming in VS Code Settings:

// General streaming toggle (applies if specific model type toggle isn't set)
"texra.model.useStreaming": false,

// Specific toggle for Anthropic reasoning models
"texra.model.useStreamingAnthropicReasoning": false,

// Specific toggle for OpenAI reasoning models
"texra.model.useStreamingOpenAIReasoning": false,

// Specific toggle for Google models
"texra.model.useStreamingGoogle": false

// Similar configuration exists for DeepSeek and OpenRouter models

Next Steps

• Built-in Agents: See which agents work well with different models.
• Configuration: Learn about other model-related settings like streaming.
• OpenAI Responses API: Overview of the new API used when useOpenAIResponsesAPI is enabled.