LLM Models

How models, their prices and their capabilities work, and how to choose a model for a job.

Overview

LLM Models are the specific AI models available from the providers you have connected. Once a provider is connected, every approved model it offers can be chosen for a job. You do not add models or enter prices yourself: each model's price and capabilities come from Outermind's model catalog, so cost tracking works from the first run. New to the idea of a model? Start with What Is an AI Model?

Modern LLM providers charge for more than just input and output tokens. Cache reads, cache writes, and long context requests each have their own pricing dimensions. The catalog carries all of them, so the Usage Dashboard reflects your actual provider billing.

Understanding Models

Model Hierarchy

Provider (e.g., Anthropic)
└── Model (e.g., claude-sonnet-4-6)
    ├── Input pricing (per million tokens)
    ├── Output pricing (per million tokens)
    ├── Cache read discount (fraction of input price)
    ├── Cache write premium (multiplier on input price)
    └── Pricing tiers (long context surcharge configuration)

Model Capabilities

Beyond pricing, models carry capability flags that control which agent tools are available. The current flags are:

Flag	Description
Supports Tool Calling	Model can invoke functions/tools during a conversation
Supports Streaming	Model streams tokens as they are generated
Supports Vision	Model can process images in the prompt (e.g., screenshots, attachments)
Supports Image Generation	Model can generate images (e.g., gpt-image-2, gpt-image-1, DALL-E 3). Enables the `images.create` action for agents.

These flags come from the model catalog. A model with Supports Image Generation can be chosen for the Creates images job, which makes the images.create action work for agents that have the images tool. Image generation is supported by OpenAI and Microsoft Foundry (Azure) models such as gpt-image-2.

Different models offer different speed/quality/cost tradeoffs:

Model Tier	Speed	Intelligence	Cost	Best For
Small (Haiku, GPT-4o-mini)	Fast	Good	Low	Simple tasks, high volume
Medium (Sonnet, GPT-4o)	Moderate	Excellent	Medium	General use, balanced
Large (Opus, GPT-5.2)	Slower	Best	High	Complex reasoning

Context Windows and Output Limits

Context window is the maximum number of tokens a model can process in a single request (input + output combined). Max output is the ceiling for generated response tokens.

Provider	Model	Context Window	Max Output
Anthropic	Claude Opus 4.6	1,000,000	128,000
Anthropic	Claude Sonnet 4.6	1,000,000	128,000
Anthropic	Claude Opus/Sonnet 4.5	200,000	64,000
Anthropic	Claude Haiku 4.5	200,000	64,000
OpenAI	GPT-5.2	400,000	128,000
OpenAI	GPT-5 Mini	400,000	128,000
OpenAI	GPT-4o / GPT-4o Mini	128,000	16,384
xAI	Grok 4	256,000	100,000
xAI	Grok 4.1 Fast	2,000,000	100,000
Gemini	Gemini 2.5 Pro / Flash	1,000,000	65,536

Context window sources (verified 2026-03-15):

Viewing Models

Where to See Models

Go to Workforce > Intelligence > AI Models and click the model on any job. This opens the model picker, which lists every model you can use for that job, across all your connected providers.

Model Information

Each row in the picker shows the model's name, its provider, the list price for input and output per million tokens, and its context window. Select a row to see more on the right:

Model ID - The technical identifier used by the provider
Fit for this job - How well the model scored for this job in Outermind's tests, or "Not tested"
List price - Cost per million input and output tokens, with a comparison to the model you use now
Context window and longest reply
Tool calling and Reads images - What the model can do
Your data goes to - Where the model runs
Released - When the model came out

Adding Models

You do not add models by hand. Connecting a provider makes every approved model it offers available in the picker, with its price and capabilities already filled in from the model catalog. See LLM Providers for how to connect one.

The page where you used to add, edit, test and remove models one at a time, including models with a custom name or endpoint, has been removed. Models you added there before still work and still appear in the picker.

How Pricing Works

Accurate pricing enables:

Cost tracking per execution
Cost analysis by agent
Budget monitoring and alerts
ROI calculations including cache efficiency

Each model's prices are part of its catalog entry: the input price, the output price, the cached input discount, the cache write premium and any long context surcharge. You do not enter or edit them. The sections below explain what each one means so you can read your costs.

Cache Read Discount

The Cached Input Discount is a multiplier applied to the input price for cache read tokens. Values less than 1.0 represent a discount:

Value	Meaning	Effective cost
`0.10`	90% discount	Cache reads cost 10% of standard input price
`0.25`	75% discount	Cache reads cost 25% of standard input price
`0.50`	50% discount	Cache reads cost 50% of standard input price
`1.00`	No discount	Cache reads charged at full input rate

Cache Write Premium

The Cache Write Premium is a multiplier applied to the input price for cache write (cache creation) tokens. Values greater than 1.0 represent a premium:

Value	Meaning	Effective cost
`1.00`	No premium	Cache writes charged at standard input rate
`1.25`	25% premium	Cache writes cost 125% of standard input price

Pricing Tiers

A model's pricing tiers describe its long context surcharge. They are stored in this shape (a model with no surcharge has none):

{
  "longContext": {
    "thresholdTokens": 200000,
    "inputMultiplier": 2.0,
    "outputMultiplier": 1.5
  }
}

thresholdTokens - Input token count above which the surcharge activates
inputMultiplier - Input price multiplier when above threshold (2.0 = double price)
outputMultiplier - Output price multiplier when above threshold (1.5 = 50% more)

Current Pricing Reference

Prices as of March 2026 (verify with provider):

Anthropic

Model	Input (per 1M)	Output (per 1M)	Cache Read Discount	Cache Write Premium	Long Context
Claude Opus 4.6	$5.00	$25.00	0.10 (90% off)	1.25 (25% premium)	None (flat rate to 1M)
Claude Sonnet 4.6	$3.00	$15.00	0.10 (90% off)	1.25 (25% premium)	`>`200k: $6/$22.50
Claude Opus 4.5	$5.00	$25.00	0.10 (90% off)	1.25 (25% premium)	`>`200k: 2x/1.5x
Claude Sonnet 4.5	$3.00	$15.00	0.10 (90% off)	1.25 (25% premium)	`>`200k: 2x/1.5x
Claude Haiku 4.5	$1.00	$5.00	0.10 (90% off)	1.25 (25% premium)	`>`200k: 2x/1.5x

Most Anthropic models have a long context surcharge at 200,000 tokens (2x input, 1.5x output). The exception is Claude Opus 4.6, which charges a flat rate up to 1M tokens with no surcharge.

OpenAI

Model	Input (per 1M)	Output (per 1M)	Cache Read Discount	Cache Write Premium	Long Context
GPT-5.2	$1.75	$14.00	0.10 (90% off)	1.00 (none)	`>`200k: 2x/1.5x
GPT-5 Mini	$0.25	$2.00	0.10 (90% off)	1.00 (none)	`>`200k: 2x/1.5x
GPT-4o	$2.50	$10.00	0.50 (50% off)	1.00 (none)	`>`200k: 2x/1.5x
GPT-4o Mini	$0.15	$0.60	0.50 (50% off)	1.00 (none)	`>`200k: 2x/1.5x

OpenAI caching is automatic with no write premium. GPT-5 family models get a 90% cache discount; GPT-4o family gets 50%. Legacy models (GPT-4, GPT-3.5 Turbo) do not have long context surcharges.

xAI (Grok)

Model	Input (per 1M)	Output (per 1M)	Cache Read Discount	Cache Write Premium
Grok 4	$3.00	$15.00	0.25 (75% off)	1.00 (none)
Grok 4.1 Fast	$0.20	$0.50	0.25 (75% off)	1.00 (none)

xAI models do not have a long context surcharge. Caching is automatic.

Google Gemini

Model	Input (per 1M)	Output (per 1M)	Cache Read Discount	Cache Write Premium	Long Context
Gemini 2.5 Pro	$1.25	$10.00	0.10 (90% off)	1.00 (none)	`>`200k: $2.50/$15
Gemini 2.5 Flash	$0.30	$2.50	0.10 (90% off)	1.00 (none)	`>`200k: 2x/1.5x

Gemini models have a long context surcharge at 200,000 tokens: 2x input, 1.5x output.

Pricing Sources (verified 2026-03-15):

warning

Prices change frequently. The prices above were last verified on March 15, 2026, so check your provider's own price list for current rates. The model picker shows the list price Outermind uses for each model.

Understanding Cache Pricing

What Are Cache Reads and Cache Writes?

LLM providers cache portions of prompts (typically system prompts and tool definitions) to avoid reprocessing identical content on every request. When a request reuses cached content:

Cache read - The provider retrieves previously processed prompt content from cache. Because the provider already processed this content, it charges a steep discount (typically 75-90% off standard input price).
Cache write - The first time content is processed and stored in cache. Anthropic charges a small premium for this initial storage (25% above input price). Other providers like OpenAI handle caching automatically with no explicit write charge.

How Different Providers Handle Caching

Provider implementations differ in important ways:

Provider	Cache writes	Cache reads	Write premium	Read discount
Anthropic	Explicit - you can see `cache_creation_input_tokens`	Explicit - `cache_read_input_tokens`	1.25x	0.10x (90% off)
OpenAI / Azure	Automatic - no separate field reported	Reported in `cached_tokens` (subset of prompt_tokens)	None (1.00x)	0.50x (50% off)
xAI (Grok)	Automatic	Tracked internally	None (1.00x)	0.25x (75% off)
Google Gemini	Explicit via Content Caching API	`cached_content_token_count`	None (1.00x)	0.10x (90% off)

Because providers report tokens differently, Control Bridge applies a provider-aware formula when calculating costs. Anthropic's input_tokens field counts only the uncached tokens; cache writes and cache reads are reported separately and added on top. OpenAI's prompt_tokens field includes cache read tokens as a subset. The cost engine handles these differences automatically.

Why Cache Pricing Matters for Cost Accuracy

Without accurate cache pricing, cost reporting can be significantly skewed:

Cache reads overcharged - If cache read tokens are billed at the full input rate instead of the discounted rate, costs appear higher than actual provider billing.
Cache writes undercharged - If Anthropic cache write tokens are not tracked with the 1.25x premium, costs appear lower than actual billing.
Net effect varies by workload - For agents with large, stable system prompts that benefit from caching, the reads discount typically outweighs the write premium. Configuring both correctly is essential for accurate cost tracking.

Cost Calculation Formula

Control Bridge uses a multi-dimensional cost formula for each LLM call:

totalCost =
  (standardInputTokens / 1M) * inputPrice * inputMultiplier
  + (cacheReadTokens / 1M) * inputPrice * cachedInputDiscount * inputMultiplier
  + (cacheWriteTokens / 1M) * inputPrice * cacheWritePremium * inputMultiplier
  + (outputTokens / 1M) * outputPrice * outputMultiplier

Where:

standardInputTokens - Non-cached, non-write input tokens (provider-specific calculation)
cachedInputDiscount - The Cached Input Discount configured on the model (e.g., 0.10)
cacheWritePremium - The Cache Write Premium configured on the model (e.g., 1.25)
inputMultiplier and outputMultiplier - 1.0 normally; elevated to 2.0/1.5 when long context surcharge applies

Long Context Surcharges

What Triggers a Surcharge

When the total input tokens in a request exceed 200,000 tokens, most major providers charge a higher rate for that entire request. This threshold applies to the full context window including system prompts, conversation history, and document content.

How the Multipliers Work

When a request crosses the threshold, all input-category costs (standard input, cache reads, and cache writes) are multiplied by the inputMultiplier, and output costs by the outputMultiplier. The most common configuration across providers is:

Input multiplier: 2.0 (double the standard input rate)
Output multiplier: 1.5 (50% more than standard output rate)

This applies to the entire request - not just the tokens above the threshold.

Which Providers and Models Have Surcharges

Provider	Threshold	Input multiplier	Output multiplier	Notes
Anthropic	200,000 tokens	2.0x	1.5x	All Claude models
OpenAI	200,000 tokens	2.0x	1.5x	Modern models (GPT-4o, GPT-5). Legacy GPT-4 and GPT-3.5 are exempt
Azure AI Foundry	200,000 tokens	2.0x	1.5x	Modern models (gpt-4o, gpt-4o-mini, gpt-4-turbo). Legacy gpt-4, gpt-4-32k, gpt-35-turbo are exempt
Google Gemini	200,000 tokens	2.0x	1.5x	All Gemini models
xAI (Grok)	N/A	N/A	N/A	No long context surcharge

info

A long context surcharge is part of the model's pricing tiers in the catalog. Models without one have no pricing tiers.

Model Selection

The AI Models Page

Go to Workforce > Intelligence > AI Models to control which model handles each internal work type (called a "job" on this page). The page is covered in full in The AI Models Page; this section is a quick reference for the jobs themselves.

Work Type	What it governs
Reasoning	Thinking: scheduled deep work, planning, and the project runs your agents do on their own
Chat	Conversation: Boardroom chat, Teams and Slack replies, and agent chat windows
Composition	Writing: email replies, drafts, and the daily brief
Classification	Sorting: deciding which emails matter and what they are about, plus other routine labelling
Safety	Screening: checking content for PII and unsafe instructions, before your agents act and before anything is sent
Summarization	Condensing: shortening documents, memories, and activity into recaps
Image	Pictures: image generation for the `images.create` agent tool action. On the page this job is called Creates images

For example, if email replies read poorly, change Composition. If scheduled planning runs feel shallow, change Reasoning.

In Your setup, each job row shows the model it currently resolves to and a follow chip (Default, Your pick: stays current, or Your pick: pinned). Most jobs fall back to the Operations model when no assignment exists yet; that row reads "Uses Operations: <model>" with no follow chip until you choose a model.

Image is the exception. Its model picker only lists models with Supports Image Generation enabled (see Model Capabilities above); other active models never appear as options for this job. Image also has no fallback: leaving it unassigned means the images.create agent tool action returns an error to agents instead of silently using another model.

Until you choose one, the Creates images row reads "Off. Choose an image model." Click it to open the picker. If none of your connected providers has an image model, the picker lists the ones you could use under Needs a connection, such as GPT Image 2 from OpenAI or from Microsoft Foundry. Select the OpenAI one and paste your OpenAI key right there to connect and use it; the Foundry one needs your Azure setup.

The tenant setup wizard's Model Assignments step includes the same Image row, filtered the same way. It's optional: if your tenant has no image-capable model yet, or you skip the row, the wizard completes normally with no Image assignment. When an image-capable model is available, the wizard defaults the selection to gpt-image-2 if present.

Choosing the Right Model

Consider these factors when selecting models for agents:

Task Complexity

Simple: Classification, routing, simple Q&A - Haiku / GPT-4o-mini
Moderate: Customer support, content generation - Sonnet / GPT-4o
Complex: Analysis, multi-step reasoning - Opus / GPT-4 Turbo

Response Time

Smaller models respond faster
Consider user expectations
Batch processing can use larger models

Cost vs Quality

Calculate expected costs using the full multi-dimensional formula:

Daily cost = (
  avg_standard_input_tokens * inputPrice
  + avg_cache_read_tokens * inputPrice * cachedInputDiscount
  + avg_cache_write_tokens * inputPrice * cacheWritePremium
  + avg_output_tokens * outputPrice
) * daily_executions / 1,000,000

For agents with large, stable system prompts, cache reads can significantly reduce the effective input cost. Use the Usage Dashboard's cache token data to understand your actual cache hit rates.

The Model Picker (per-agent, SME, and DreamState overrides)

Every place you choose a model uses the same picker: a job on the AI Models page, a model override on an existing agent's edit page, a Specialist Agent (SME), or DreamState's dream-processing model. The picker opens as a window titled with the job you're setting (for example, "Choose a model for Writes"). The list is on the left and the details of the selected model are on the right:

Search - type part of a model or provider name; the words "cheap", "azure", "vision" and "new" also work as shortcuts.
Filter chips - narrow by provider or by capability. A chip that is switched on is filled in; one that is off is outlined. A provider you have not connected has a dashed outline. A capability the job requires (such as tool calling) shows as a locked chip and cannot be removed.
Sort - choose Best for this job, Lowest price, or Newest.
Groups - results are grouped into From your providers, Needs a connection (providers you haven't connected yet, collapsed until you click Show them), and Cannot do followed by the job's name, such as Cannot do Reasoning (hidden by default until you click Show why).
Hide older versions - on by default, this collapses superseded model versions out of the list.
Detail pane - select a row to see its fit for the job, list price per 1M tokens (shown as "Nx the price of now" relative to your current model), release date, context window, and whether it calls tools and reads images.
Compare - tick the box on up to 3 rows to see them side by side; the picker refuses a fourth. The tick boxes are for comparing only. To choose a model, click its row, then Use [model].

If you select a model from a provider you haven't connected yet, the details on the right show a field for that provider's API key. Paste the key and click Check key. When the key works, the provider is connected, the model stays selected, and the button at the bottom changes to Use [model]. For Azure AI Foundry the picker shows Set up your Azure instead, which takes you to the Azure setup page.

To return an override to your tenant's defaults, choose Use default - this clears the override, and the field goes back to resolving the model from your AI Models page jobs.

Model Recommendations by Use Case

Use Case	Recommended Model	Reasoning
Email triage	Claude Haiku	Fast, cost-effective for classification
Customer support	Claude Sonnet	Good balance of quality and cost
Technical support	Claude Sonnet	Handles complexity well
Executive summaries	Claude Opus	Highest quality output
Data extraction	Claude Haiku	Structured tasks do not need large models

Token Counting

What Are Tokens?

Tokens are the units LLMs use to process text:

~4 characters = 1 token (English)
~3/4 words = 1 token
Code and special characters may use more tokens

Token Types

Each LLM call can produce several categories of tokens:

Token type	Description	Billing
Input tokens	New, non-cached prompt content	Standard input price
Cache read tokens	Prompt content retrieved from provider cache	Discounted input price
Cache write tokens	Prompt content written to provider cache for the first time	Premium input price (Anthropic only)
Output tokens	Generated response content	Output price

Estimating Token Usage

For a typical email interaction:

System prompt: 200-500 tokens
Email content: 100-1,000 tokens
Response: 100-500 tokens

Average execution: ~1,000-2,000 total tokens. Agents with large system prompts benefit more from caching because the system prompt content is reused across calls.

Viewing Token Usage

Check token usage in Governance > Activity History > Agent Runs:

Open a run and select a model-call step in its trace
Open the step inspector's Tokens tab
See input, output, cache read, and cache write token counts for that call

Aggregate cache token totals by agent and day are available in Governance > Activity History > Platform Usage.

Cost Tracking

Execution-Level Costs

Each execution records:

Standard input tokens used
Cache read tokens (from provider cache)
Cache write tokens (cache creation, Anthropic only)
Output tokens generated
Calculated cost using the full multi-dimensional formula

Aggregate Views

In Governance > Activity History > Platform Usage, view:

Total costs by time period
Costs by provider
Costs by agent
Cost trends
Cache token totals per agent per day

Setting Budgets

While Control Bridge does not enforce budgets automatically, you can:

Monitor costs in the Usage dashboard
Set up alerts for unusual spending
Review high-cost executions
Adjust model selection for cost optimization

Best Practices

Cost Optimization

Start with smaller models - Upgrade only if quality is insufficient
Test model changes - Compare quality before switching
Monitor outliers - Investigate unexpectedly expensive executions
Optimize prompts - Shorter prompts mean fewer input tokens
Leverage caching - Stable system prompts benefit from cache reads at a steep discount

Quality Assurance

Review sample outputs - Regularly check agent responses
A/B test models - Compare different models on same inputs
User feedback - Track response quality ratings
Adjust as needed - Upgrade models for struggling agents

Pricing Updates

Nothing to update - List prices live in Outermind's model catalog, not in your settings
Watch for changes - Provider pricing changes frequently; the picker shows the list price in use
Report a wrong price - Tell support if a list price does not match your provider
Recalculate budgets - After significant price changes

Troubleshooting

Model Not Available

Symptoms: Model shows as unavailable or missing

Solutions:

Verify provider API key has access to the model
Check if model requires special access (waitlist)
Confirm model ID is correct

Incorrect Cost Calculations

Symptoms: Costs do not match provider billing

Solutions:

Open the model picker and compare the list price shown for the model with your provider's price list
Remember that cache reads are cheaper than standard input (for example 0.10x for Anthropic, 0.50x for OpenAI GPT-4o models) and that Anthropic cache writes cost 1.25x
Check whether the requests are over 200,000 tokens, where most providers add a long context surcharge
Compare the token breakdown in Governance > Activity History > Agent Runs with your provider dashboard to confirm token counts match
If a list price in the picker is wrong, contact support. Prices are maintained in the model catalog and cannot be edited on the page

Model Performance Issues

Symptoms: Slow responses or timeouts

Solutions:

Check provider status for outages
Consider switching to a faster model
Optimize prompt length
Review concurrent request limits

Overview​

Understanding Models​

Model Hierarchy​

Model Capabilities​

Context Windows and Output Limits​

Viewing Models​

Where to See Models​

Model Information​

Adding Models​

How Pricing Works​

Cache Read Discount​

Cache Write Premium​

Pricing Tiers​

Current Pricing Reference​

Anthropic​

OpenAI​

xAI (Grok)​

Google Gemini​

Understanding Cache Pricing​

What Are Cache Reads and Cache Writes?​

How Different Providers Handle Caching​

Why Cache Pricing Matters for Cost Accuracy​

Cost Calculation Formula​

Long Context Surcharges​

What Triggers a Surcharge​

How the Multipliers Work​

Which Providers and Models Have Surcharges​

Model Selection​

The AI Models Page​

Choosing the Right Model​

Task Complexity​

Response Time​

Cost vs Quality​

The Model Picker (per-agent, SME, and DreamState overrides)​

Model Recommendations by Use Case​

Token Counting​

What Are Tokens?​

Token Types​

Estimating Token Usage​

Viewing Token Usage​

Cost Tracking​

Execution-Level Costs​

Aggregate Views​

Setting Budgets​

Best Practices​

Cost Optimization​

Quality Assurance​

Pricing Updates​

Troubleshooting​

Model Not Available​

Incorrect Cost Calculations​

Model Performance Issues​

Related Topics​

Overview

Understanding Models

Model Hierarchy

Model Capabilities

Context Windows and Output Limits

Viewing Models

Where to See Models

Model Information

Adding Models

How Pricing Works

Cache Read Discount

Cache Write Premium

Pricing Tiers

Current Pricing Reference

Anthropic

OpenAI

xAI (Grok)

Google Gemini

Understanding Cache Pricing

What Are Cache Reads and Cache Writes?

How Different Providers Handle Caching

Why Cache Pricing Matters for Cost Accuracy

Cost Calculation Formula

Long Context Surcharges

What Triggers a Surcharge

How the Multipliers Work

Which Providers and Models Have Surcharges

Model Selection

The AI Models Page

Choosing the Right Model

Task Complexity

Response Time

Cost vs Quality

The Model Picker (per-agent, SME, and DreamState overrides)

Model Recommendations by Use Case

Token Counting

What Are Tokens?

Token Types

Estimating Token Usage

Viewing Token Usage

Cost Tracking

Execution-Level Costs

Aggregate Views

Setting Budgets

Best Practices

Cost Optimization

Quality Assurance

Pricing Updates

Troubleshooting

Model Not Available

Incorrect Cost Calculations

Model Performance Issues

Related Topics