Amazon Nova Context Window: Token Limits for Micro, Lite, Pro, and Premier

As of September 2026, official Amazon Web Services documentation specifies Amazon Nova context windows as 128,000 tokens for Nova Micro, 300,000 tokens for Nova Lite and Nova Pro, and 1,000,000 tokens for Nova Premier. Amazon Nova context windows range from 128,000 tokens for Nova Micro, 300,000 tokens for Nova Lite and Pro, up to 1,000,000 tokens for Nova Premier, with multimodal video support up to 90 minutes.

These four foundation models constitute the understanding tier of the Amazon Nova model family on Amazon Bedrock. Rather than applying a single uniform token capacity across every release, AWS engineered distinct context window sizes aligned with specific latency, cost, and reasoning profiles.

Token Capacity Across Understanding Models

The Amazon Nova understanding models address different operational needs across the sequence length spectrum:

• Amazon Nova Micro (128,000 tokens): Nova Micro is a text-only model engineered for the lowest response latency at minimal cost. Its 128,000-token context window handles roughly 96,000 English words. It fits high-throughput text classification, query transformation, language translation, and light conversational routing. Because Nova Micro omits multimodal processing and document parsing layers, its memory footprint remains exceptionally compact.
• Amazon Nova Lite (300,000 tokens): Nova Lite provides a 300,000-token context window capable of ingesting text, images, and video. It is built for fast, cost-sensitive processing across diverse document types and streaming media. A 300,000-token capacity accommodates approximately 225,000 English words, equivalent to several dense books or hundreds of document pages in a single prompt.
• Amazon Nova Pro (300,000 tokens): Nova Pro matches the 300,000-token context window of Nova Lite while delivering higher reasoning accuracy and deeper tool-calling capabilities. It serves as AWS's balanced model for code analysis, structured data extraction, mathematical problem solving, and multi-turn autonomous agent coordination.
• Amazon Nova Premier (1,000,000 tokens): Nova Premier is AWS's flagship multimodal model for complex reasoning tasks. Its 1,000,000-token context buffer processes 750,000 words of text, hundreds of document pages, 500 images, or 90 minutes of continuous video. It targets large-scale enterprise analysis, comprehensive codebase refactoring, and multi-document synthesis across disparate formats.

The table below outlines context window sizes, output token limits, input modalities, and document support across the four Amazon Nova understanding models:

Output Token Caps and Input Headroom

Every Amazon Nova understanding model enforces a maximum output limit of 10,000 tokens per individual request. This output capacity is separate from the input buffer. However, the total tokens in any single call (input prompt tokens plus generated output tokens) must not exceed the model's total context window.

For example, if you send an input prompt containing 295,000 tokens to Amazon Nova Pro, the model can generate at most 5,000 output tokens before hitting the 300,000-token total ceiling. If your request specifies a max tokens parameter of 10,000, generation terminates early once the total token count reaches 300,000. Maintaining sufficient headroom between input size and maximum context window prevents truncated completions in production environments.

• AWS Bedrock Context Window: Token Limits, Model Capacities, and Memory ArchitectureThe AWS Bedrock context window defines the maximum sequence of input and output tokens a hosted foundation model...
• DeepSeek V3 Context Window: Token Limits, Output Caps, and ArchitectureThe DeepSeek V3 context window provides 128,000 tokens for API endpoints and developer integrations, with a maximum...
• GPT-4o Context Window: Token Limits, Architecture, and MCP SearchThe GPT-4o context window is 128,000 tokens, supporting up to 16,384 completion tokens per API request. While 128,000...
• OpenAI o1 Context Window: Token Limits, Reasoning Overhead & SolutionsThe OpenAI o1 context window is 200,000 tokens with a maximum completion limit of 100,000 tokens, where internal...
• Gemini 2.0 Flash Context Window: 1M Architecture and RAG Best PracticesThe Gemini 2.0 Flash context window spans 1,048,576 input tokens and an 8,192-token output ceiling. While ingesting...
• Cohere Context Window: Command R Token Limits and Enterprise SearchThe Cohere context window provides 128,000 tokens of sequence capacity on Command R and Command R+, and 256,000 on the...

More on this subject: Agent Memory and Storage (248 guides)

A common trap when deploying multimodal models is assuming that file sizes in megabytes translate linearly into token counts. With Amazon Nova, token consumption varies depending on whether the input consists of raw text, formatted office documents, high-resolution images, or temporal video streams.

Understanding how Bedrock tokenizes non-text modalities allows teams to forecast token burn and prevent unexpected context overflow.

Document and Image Ingestion Mechanics

Amazon Nova Lite, Pro, and Premier support direct document ingestion via the Amazon Bedrock Converse API and InvokeModel API. Supported formats include PDF, DOC, DOCX, XLS, XLSX, CSV, HTML, TXT, and Markdown files. Amazon Nova Micro does not support document attachment blocks and accepts only raw text strings.

When a document is passed to Amazon Nova Lite, Pro, or Premier, Bedrock processes both the textual content and the structural layout:

• Standard Text Documents: For standard English text files, token conversion closely follows standard byte-pair encoding rules, averaging approximately 750 words per 1,000 tokens. A 30-page text document of 15,000 words consumes roughly 20,000 tokens.
• Complex Spreadsheets and Tables: CSV, XLS, and XLSX files consume tokens rapidly because row delimiters, column boundaries, numerical figures, and cell coordinates each require distinct tokens. A financial spreadsheet with 2,000 rows and 15 columns can consume 40,000 to 60,000 tokens.
• Rendered Images and Visual Pages: Nova processes images by converting them into visual tokens based on image dimensions and aspect ratios. A high-resolution document page or diagram rendered at standard 150 to 300 DPI consumes between 1,000 and 1,600 tokens per image. A 50-page scanned PDF converted into page images consumes between 50,000 and 80,000 tokens before prompt instructions or system prompts are added.

Video Frame Sampling and Duration Ceilings

The most demanding input modality for Amazon Nova is video. Nova Lite, Pro, and Premier process video by extracting and encoding individual frames along with associated audio channels.

Amazon Nova Premier supports video inputs of 90 minutes in duration, which is the longest native video processing window available in Bedrock. Nova Lite and Pro process video for shorter clips, recommended between 15 and 30 minutes to maintain optimal response times.

Video token utilization depends on the frame sampling rate and the visual motion complexity of the footage:

• Frame Extraction Rates: AWS Bedrock samples video frames at regular intervals, typically 1 frame per second for low-motion presentations or screencasts, and up to higher sampling frequencies for dynamic content.
• Token Conversion Per Second: At standard sampling resolutions, video input consumes between 150 and 250 tokens per second of footage. A 10-minute video clip (600 seconds) translates to roughly 90,000 to 150,000 tokens.
• 90-Minute Maximum on Premier: A full 90-minute video recording consumes virtually the entire 1,000,000-token context window of Amazon Nova Premier. Submitting a 90-minute video leaves minimal headroom for detailed system prompts, extensive reasoning scratchpads, or multi-turn conversational follow-ups.

The following table provides typical token conversion benchmarks across document and video formats:

When processing long video files or extensive document collections, engineering teams face significant architectural tradeoffs between raw token capacity, inference latency, and operational cost.

While having 300,000 or 1,000,000 tokens available in Amazon Bedrock is technically impressive, filling those buffers creates severe performance and operational penalties. Treating large context windows as unbounded scratchpads introduces three critical engineering challenges: prefill latency, API timeout failures, and compounding token costs.

Prefill Latency and Read Timeout Constraints

Transformer inference consists of two computational stages: the prompt prefill phase and the token generation phase. During prefill, the model must process the entire input prompt simultaneously, calculating key-value attention representations for every token.

As input sequence lengths grow, time-to-first-token increases substantially:

• Short Prompts (below 4,000 tokens): Prefill completes in milliseconds. Nova Micro and Lite respond almost instantaneously.
• Medium Prompts (50,000 to 100,000 tokens): Prefill latency stretches to several seconds on Nova Pro and Premier as the attention matrix scales across tens of thousands of tokens.
• Full Context Prompts (250,000 to 900,000 tokens): Prefill latency can take 20 to 50 seconds or longer depending on Bedrock cluster utilization and regional traffic.

This prefill delay creates a serious vulnerability with standard SDK configurations. Default AWS SDK clients (including boto3 in Python and the JavaScript SDK in Node.js) enforce a default socket read timeout of 60 seconds.

When a client submits a 250,000-token prompt or a 45-minute video to Amazon Nova Pro or Premier, the prefill computation and initial token generation frequently exceed 60 seconds. When this happens, the client socket closes with a read timeout error, terminating the connection while Bedrock continues processing in the background.

To prevent client-side timeouts during large-context inference, developers must explicitly adjust their HTTP client settings:

import boto3
from botocore.config import Config

#-- configure extended read timeouts for large-context Amazon Nova inference
bedrock_config = Config(
    read_timeout=3600,
    connect_timeout=15,
    retries={"max_attempts": 3, "mode": "adaptive"}
)

bedrock_client = boto3.client(
    service_name="bedrock-runtime",
    region_name="us-east-1",
    config=bedrock_config
)

Setting read timeout to 3,600 seconds provides up to 60 minutes of client tolerance, matching AWS Bedrock's internal maximum inference execution timeout.

Token Cost Accumulation in Multi-Turn Agent Loops

Inference pricing on Amazon Bedrock scales with model size and token volume. Nova Micro is priced for low-cost utility tasks, while Nova Lite provides high-throughput multimodal processing. Nova Pro costs roughly an order of magnitude more than Lite, reflecting its advanced reasoning and code generation abilities. Nova Premier sits at the top tier, reflecting the high compute demands of its 1,000,000-token attention space.

In autonomous agent workflows, models do not execute in a single shot. An agent inspects a goal, invokes a tool, receives tool output, updates its scratchpad, and iterates through multiple turns.

If an agent loads a 200,000-token documentation corpus into its context window, that entire 200,000-token payload is resent to Bedrock on every single iteration:

• Iteration 1 (Initial Analysis): The model ingests 200,000 input tokens.
• Iteration 2 (First Tool Execution): The model ingests the initial 200,000 tokens plus tool definitions and execution output, totaling 205,000 input tokens.
• Iteration 3 (Second Tool Execution): The prompt grows to 210,000 input tokens.
• Iteration 4 (Final Synthesis): The prompt reaches 215,000 input tokens.

Over four turns, the agent processes 830,000 cumulative input tokens for a single user task. When an agent workflow runs thousands of times per day across an organization, repeatedly resending massive context buffers creates substantial cloud inference bills.

Explicit Prompt Caching Capabilities and Boundaries

To mitigate the cost and latency of repetitive prompts, Amazon Bedrock supports explicit prompt caching on Amazon Nova models. Prompt caching allows developers to store static prompt prefixes in memory, skipping redundant prefill computation on subsequent requests.

Key specifications for Amazon Nova prompt caching include:

• Minimum Checkpoint Size: Each cache checkpoint requires at least 1,000 tokens.
• Maximum Checkpoint Capacity: Nova models enforce a cache threshold of 20,000 tokens per checkpoint.
• Checkpoints Per Request: The Converse API supports 4 distinct cache checkpoints across system prompts and tool definitions.
• Time-to-Live (TTL): Checkpoints remain active in cache for 5 minutes, resetting with every successful cache hit.
• Pricing Benefit: Tokens read from the cache receive a substantial discount compared to standard input token rates.

While prompt caching is effective for reusable tool schemas and system instructions, its 20,000-token ceiling reveals a fundamental constraint: you cannot cache a 200,000-token document collection. The 20,000-token limit covers only a small fraction of a 300,000-token context window and an even smaller portion of a 1,000,000-token window. Any dynamic document content exceeding 20,000 tokens must be re-ingested and calculated at standard on-demand rates on every API call.

Extend Amazon Nova Beyond Raw Context Limits

Connect Amazon Nova agents to persistent Fast.io workspaces over MCP. Query indexed documents with citations instead of exhausting your Bedrock context window. Starts with a 14-day free trial.

Start 14-Day Trial

The presence of a 300,000 or 1,000,000-token context window tempts developers to treat the LLM as both reasoning engine and database. In practice, stuffing large file collections directly into model context produces severe architectural flaws that go beyond latency and cost.

Context windows represent active working memory, not persistent enterprise storage.

The Limits of Brute-Force Context Ingestion

Relying entirely on giant context windows introduces three primary operational points of failure:

1. Attention Degradation (The Needle-in-a-Haystack Problem): Although modern foundation models can process 300,000 or 1,000,000 tokens, factual retrieval accuracy is not uniform across the entire prompt. Models demonstrate higher retrieval accuracy on facts located near the beginning or end of long sequences, while information buried in the middle of a massive prompt suffers from measurable recall degradation.
2. Context Isolation and Data Duplication: When documents live only inside an API prompt, they exist in isolation. Other agents, automated scripts, and human collaborators cannot see, verify, or update those files. If two separate agents work on the same legal matter or technical project, each must download, format, and re-upload duplicate copies into Bedrock.
3. Loss of Version Control and Audit Trails: Passing raw documents into an API call provides zero version tracking. If a teammate updates a specification or a client modifies an agreement, the agent continues running against obsolete context unless a brittle local file-synchronization pipeline is maintained.

Automatic Workspace Indexing and Metadata Extraction

The production alternative to brute-force context packing is coupling Amazon Nova with an intelligent cloud workspace platform like Fast.io.

Rather than shoving multi-megabyte PDFs, transcripts, and codebases into Amazon Bedrock, organizations store their master files in persistent Fast.io workspaces. When files land in a workspace, Fast.io handles the indexing and data extraction layers:

• Intelligence Mode and Hybrid Search: Enabling Intelligence Mode on a Fast.io workspace automatically indexes all files for hybrid retrieval, combining full-text lexical keyword search with high-dimensional semantic vector search. An agent does not need to manage an external vector database, configure chunking parameters, or maintain embedding pipelines.
• Structured Data with Metadata Views: For structured extraction tasks across contracts, receipts, or engineering specs, Metadata Views convert unstructured documents into queryable tables. Users specify desired fields in plain language (such as renewal dates, counterparty names, total amounts, or compliance statuses), and Fast.io extracts typed schemas without manual OCR templates.
• Cloud Sync and Imports: Fast.io supports Cloud Sync for Dropbox, Box, and OneDrive on a schedule or on demand (one-way or two-way). Google Drive files can be imported directly via URL import, with automated sync coming soon. Files updated in external storage are automatically re-indexed and made accessible to AI agents.

By offloading document persistence to an intelligent workspace, the Amazon Nova context window remains pristine. The model dedicates its token budget to system instructions, complex multi-step reasoning, and tool evaluation, pulling only precise factual excerpts into active memory when required.

Connecting Amazon Nova models to Fast.io workspaces is accomplished using the Model Context Protocol (MCP). Fast.io provides a remote, hosted MCP server accessible over Streamable HTTP and SSE, eliminating the need to manage local subprocesses or custom container runtimes.

Through this integration, Nova models running on Amazon Bedrock gain dynamic access to a consolidated MCP toolset for file searching, content reading, and workspace management.

Configuring the Fast.io Remote MCP Endpoint

The Fast.io MCP server operates remotely at https://mcp.fast.io/mcp/tools.

When configuring an AI assistant, developer environment, or agent framework, the connection configuration specifies the remote URL and your workspace API key:

{
  "mcpServers": {
    "fastio": {
      "url": "https://mcp.fast.io/mcp/tools",
      "headers": {
        "Authorization": "Bearer YOUR_FASTIO_API_KEY"
      }
    }
  }
}

This configuration connects the agent directly to your Fast.io organization without requiring local npm packages, Docker containers, or file system mounts.

Query Lifecycle and Context Preservation

When an Amazon Nova agent receives a complex user inquiry involving large document collections, it follows a structured retrieval lifecycle that preserves context headroom:

1. User Request Received: The user submits a targeted inquiry: "What are the exact termination notice requirements in the 2026 enterprise vendor contract?"
2. Context Budget Assessment: The agent evaluates its prompt state. With 1,000 tokens for system guidelines, 2,000 tokens for tool definitions, and 100 tokens for the prompt, Nova retains 296,000 tokens of headroom.
3. Targeted MCP Tool Call: Rather than ingesting the full 80-page contract, the Nova model generates a tool call to the Fast.io storage tool with action="search" and query="termination notice requirements 2026 enterprise vendor contract".
4. Hybrid Index Execution: The Fast.io MCP server executes a hybrid search across the workspace, evaluating full-text keywords and semantic embeddings across all indexed PDF, Word, and text documents.
5. Excerpt Delivery with Citations: Fast.io returns the two exact contractual paragraphs that govern termination notice, along with file IDs and line citations, consuming 600 tokens.
6. Synthesized Grounded Response: Amazon Nova synthesizes the final response, citing the specific contract and clause.

The entire interaction consumes 4,000 total tokens on Amazon Bedrock, rather than the 60,000 tokens required to load the complete agreement. This leaves ample room for subsequent reasoning steps, additional tool executions, and multi-turn conversational history.

Agent-to-Human Handoff and Subscription Details

Autonomous agents running on Fast.io do not operate in a vacuum. Once an agent gathers documents, generates reports, or processes contract data, it can organize those assets into shareable workspaces:

• Ownership Transfer: An agent can create an organization, build workspaces, ingest source files, and transfer full ownership of the organization to a human stakeholder while retaining administrative access.
• Granular Access and Scoped Shares: Teams can distribute deliverables using branded Send, Receive, and Exchange shares. Share links can be configured with password protection or expiration dates. External clients and stakeholders view or download shared files directly in their browser without consuming a seat license or creating an account.
• Per-File Version History: Every document write, update, or overwrite preserves full version history. If an agent updates a file or two parallel agents process the same directory, previous file versions remain fully restorable through the version history and the append-only audit log.
• Integrated E-Signature: When finalized documents require formal authorization, workspaces support built-in e-signature capabilities. Users can dispatch agreements for signature directly from the workspace, route signers in sequence or parallel, enforce one-time passcodes, generate tamper-evident audit certificates, and automatically file executed agreements back into workspace storage.

Fast.io operates on a transparent subscription model with usage-based credits, detailed on the Fast.io pricing page:

Monthly plans start with a 14-day free trial that requires a credit card and concludes early if trial credits are exhausted. Annual plans start immediately as paid subscriptions without a trial period. Storage and seat allowances are included directly with each plan, while credits meter AI operations such as workspace search and intelligent indexing.