Amazon Nova 2 Lite
Amazon Nova 2 Lite is a fast, lower-cost reasoning model from Amazon, available in Amazon Bedrock since December 2, 2025. It takes text, image, video and document input, returns text, supports extended thinking at three intensity levels and a context of up to 1 million tokens.
Released
December 2, 2025
Type
reasoning
License
proprietary
Context
1,000,000 tokens
Pricing
Input
$0.30 / 1M tokens
Output
$2.50 / 1M tokens
List price on the developer's own pages, checked . Standard tier, Global cross-Region inference, region selector US East (Ohio). Geo cross-region and in-region: $0.33 / $2.75. Priority $0.525 / $4.375; Flex and Batch $0.15 / $1.25 (global). Cache reads 75% less than on-demand input.
Capabilities
Sources
Every figure on this page was checked against these pages on . A figure we could not source is left out.
- AWS What's New: Amazon Nova 2 foundation models (Dec 2, 2025) (opens in a new tab)
- Amazon Nova 2 developer guide: What is Amazon Nova 2? (opens in a new tab)
- Amazon Nova 2 developer guide: Getting started with the API (opens in a new tab)
- Amazon Nova pricing (opens in a new tab)
Spotted something out of date? Tell us.
Links
In the news: Amazon Nova 2 Lite and Amazon
The Hacker News
Bring Your Own Agent: Your Employees' Assistants Are Already Logged In
The rollout of personal AI agents is encountering significant challenges, including privacy concerns and identification issues. Amazon has blocked Meta's Muse agent for failing to identify itself as a bot during browsing, and the Instinct agent has also raised privacy alarms. Despite these ongoing issues and accumulating negative experiences, the technology sector remains captivated by the capabilities and potential of these personal AI services. The article highlights that the personal AI agent rollout is proceeding with expected difficulties.
MIT Technology Review
Job titles of the future: Delivery drone air traffic controller
The burgeoning field of drone delivery is necessitating a new class of aviation professionals: Remote Pilots in Command (RPICs), who function as air traffic controllers for unmanned aerial vehicles. Trevor Wischnewsky, an RPIC at FlyTrex, exemplifies this emerging role. His responsibilities include overseeing fleets of hexacopter drones that transport goods within a three-mile radius to homes and businesses. Wischnewsky's role extends beyond managing FlyTrex's own drone traffic; he also coordinates with drones from competing companies such as Zipline, Amazon, and Alphabet's Wing, which are increasingly populating the skies over areas like Dallas. This coordination is crucial for ensuring the safe and efficient delivery of parcels. To qualify for such a position, significant aviation experience is required. When Wischnewsky applied to FlyTrex, he was undergoing training to become a commercial pilot and flight instructor and already held a drone pilot certification from the Federal Aviation Administration (FAA). FlyTrex mandates this FAA certification, coupled with an FAA-approved training program lasting several weeks, to equip pilots with the skills to intervene if necessary. Although Wischnewsky does not directly pilot the drones, he monitors their entire delivery process from FlyTrex's operations center. This monitoring relies on data from GPS, lasers, and various other sensors, providing a comprehensive view of the drone's status and environment. A key aspect of this role is a collaborative mindset, particularly concerning the integration of autonomous systems. Wischnewsky initially found it challenging to fully trust the automation, as traditional aviation training often advises against relying solely on automated systems. However, the Dallas-Fort Worth area serves as a critical testing ground for Unmanned Aircraft Systems Traffic Management (UTM). UTM is a sophisticated network designed to enable autonomous drones from different companies to share their real-time positional data. This shared information is vital for preventing mid-air collisions, even when drones operate beyond the visual line of sight of their controllers. Once an RPIC like Wischnewsky confirms that a drone is prepared for flight, its intended path is clear, and weather conditions are favorable, the UTM system largely manages the subsequent flight operations. The ability to remain calm under pressure is another essential attribute for RPICs, especially during peak operational periods when multiple drones are active. The increasing complexity and volume of drone traffic underscore the growing importance of these specialized air traffic control roles for the future of autonomous delivery systems.
MarkTechPost
Anthropic Releases Claude Haiku 5.5: A Small Model With 1M Context Priced at $0.10 per Million Input Tokens
Anthropic released Claude Haiku 5.5 this week, positioning it as its most affordable and rapid small-scale AI model to date. This new iteration is engineered for high-volume tasks such as text summarization, data compaction, classification, and the operation of sub-agents within larger AI systems. A key feature retained from previous versions is its substantial 1 million token context window, complemented by an output token capacity of up to 128,000 tokens. The pricing structure for Haiku 5.5 begins at $0.10 per million input tokens and $0.50 per million output tokens. Anthropic reports this represents a significant cost reduction, approximately 90% less than Claude Haiku 4.5 for prompts up to 100,000 tokens. The model is accessible via a hosted API and is generally available across multiple cloud platforms, including the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry, and the Claude Platform on AWS. Claude Haiku 5.5 introduces an adjustable "effort" setting, a novel capability for the Haiku class of models. This "adaptive thinking" feature is enabled by default, with the effort parameter set to a medium level. The model processes both text and image inputs and generates text outputs, possessing a knowledge cutoff date of June 2026. For batch processing, the system supports up to 300,000 output tokens, currently in beta. Two important technical considerations for users migrating to Haiku 5.5 are noted: attempts to use non-default values for temperature, top_p, or top_k parameters will result in a 400 error. Additionally, the model's new tokenizer processes text such that it consumes approximately 30% more tokens compared to Haiku 4.5 for the same textual content; a migration guide is available to address these changes. The pricing model for Haiku 5.5 employs a two-tier structure based on prompt token volume. For prompts containing up to 100,000 tokens, input costs are $0.10 per million tokens, and output costs are $0.50 per million tokens. Cache reads are priced at $0.01, and 5-minute cache writes are $0.125. For prompts exceeding 100,000 tokens, the rates increase to $0.50 per million for input and $2.50 per million for output. This contrasts with Haiku 4.5's pricing of $1 per million input and $5 per million output. Anthropic estimates that approximately 90% of Haiku 4.5 requests fell within the 100,000 token threshold. Factoring in the tokenizer adjustments, Anthropic projects that Haiku 5.5 will be, on average, about 75% cheaper for users. Further cost reductions of an additional 50% are achievable through batch processing. In comparison, GPT-6 Luna, another model in the market, lists identical short-context rates but initiates its higher pricing tier above 272,000 input tokens, at $0.20 for input and $0.75 for output. For a 150,000 token prompt, Luna is listed as cheaper. Anthropic has also released benchmark performance data for Haiku 5.5, though these figures are self-reported. On the OSWorld 2.1 offline subset, Haiku 5.5 achieved a score of 72.4%. This performance is significantly higher than GPT-6 Luna's reported 48.9% and Haiku 4.5's 15.7%. In the Terminal-Bench 4.0 benchmark, Haiku 5.5 scored 39.2%, compared to Luna's 16.4% and Haiku 4.5's 0.0%. These benchmarks suggest a substantial leap in performance for Haiku 5.5, particularly in tasks represented by these specific evaluations, positioning it as a competitive option for developers seeking efficient and cost-effective AI solutions for large-scale text processing and analysis.
MarkTechPost
Cohere Releases Embed 5: How It Compares to Voyage 4 Large, Gemini Embedding 2, and OpenAI
Cohere has released its new embedding model family, Embed 5, designed for enterprise search, retrieval-augmented generation (RAG), and agentic retrieval applications. The model family is offered in two distinct tiers: Embed 5 Pro, which prioritizes maximum retrieval quality, and Embed 5 Fast, which is optimized for latency and cost efficiency during live query operations. A key design innovation is that both Pro and Fast tiers share a unified embedding space, allowing users to index data using one tier and query it using the other, providing significant deployment flexibility. Both tiers are capable of processing text, images, and fused text-plus-image inputs, supporting over 100 languages and handling contexts of up to 128,000 tokens. Embed 5 is available today through the Cohere API and on platforms such as Microsoft Foundry and Amazon SageMaker, with options for private VPC or on-premise serving via vLLM. The API model identifiers are embed-v5.0-pro and embed-v5.0-fast, as detailed in Cohere's model documentation. These models can output embeddings in various dimensions, including 2048 (default), 1536, 1024, 768, 512, and 256, with embeddings returned as float, int8, or binary formats. Pricing for Embed 5 Pro is set at $0.12 per 1 million text tokens, while Embed 5 Fast costs $0.08 per 1 million text tokens. Image inputs are priced at $0.40 per 1 million tokens for both tiers. The models can embed page images directly or fuse an image with its associated metadata into a single vector, a feature particularly beneficial for scanned documents, slide decks, schematics, and charts where text extraction alone may lose critical information. Cohere conducted extensive testing across 40 development datasets, evaluating various corpus and query pairings. In these tests, an index created with Embed 5 Pro and queried with Embed 5 Fast achieved a score of 98.4, normalized against a Pro-plus-Pro setup at 100. An all-Fast configuration scored 96.6. Cohere recommends a pattern of indexing with the Pro tier for optimal quality and querying with the Fast tier for speed, provided both sides utilize the same output dimension. This split is particularly advantageous for agentic workloads, where an agent might perform dozens of searches per task, making query latency a critical factor. Cohere's internal benchmarks indicate that the Fast tier processed 377.3 documents per second, compared to 159.7 documents per second for the Pro tier. On the ViDoRe V3 benchmark, Embed 5 Pro achieved an average score of 85.8, representing an 8.8-point improvement over its predecessor, Embed 4. Embed 5 Fast averaged 84.5 on the same benchmark. For comparative context, Voyage 4 Large scored 83.7, Gemini Embedding 2 scored 83.2, and OpenAI's text-embedding-3-large scored 83.0.
The Hacker News
AI Coding Agents Exposed 13,000 Internal Images, Including Billing Records, on GitHub
AI coding agents tasked with sharing screenshots of code changes for review have inadvertently exposed over 13,000 internal images on public GitHub repositories, according to a report by security company Glow. These exposed images originated from developers across more than 300 organizations and included sensitive data such as customer billing records and previews of features that had not yet been publicly released. The security researchers at Glow discovered these exposures, noting that in the majority of instances, the images were found within the personal accounts of individual developers, suggesting a lack of centralized oversight or security protocols for AI-generated or AI-handled code-related artifacts. This incident highlights a significant blind spot in the security practices surrounding the use of AI coding assistants, which are increasingly being adopted by development teams to enhance productivity and streamline workflows. The proliferation of AI coding agents, such as GitHub Copilot, Amazon CodeWhisperer, and others, has revolutionized software development by providing real-time code suggestions, autocompletion, and even generating entire code blocks. While these tools offer substantial benefits in terms of speed and efficiency, their integration into development pipelines introduces new vectors for data leakage. The specific mechanism of exposure in this case involved developers using AI agents to generate or review code, and then inadvertently including screenshots of their development environments, which contained these sensitive images, in public repositories. The sheer volume of over 13,000 images underscores the widespread nature of this issue and the potential for significant data breaches. Glow's findings indicate that the exposed data is not limited to generic development assets but includes highly sensitive information. The presence of customer billing records is particularly concerning, as it could lead to identity theft, financial fraud, and severe reputational damage for the affected organizations. Furthermore, screenshots of unreleased features represent a breach of competitive intelligence, potentially allowing rivals to gain an advantage by learning about upcoming products or functionalities. The fact that these exposures occurred within personal developer accounts suggests that current security policies and training may not adequately address the risks associated with using AI coding tools, particularly concerning the handling of visual data generated or processed by these agents. The security implications of this incident are far-reaching. Organizations relying on AI coding agents must reassess their security postures to include the monitoring and control of AI-generated content and associated artifacts. This may involve implementing stricter policies on what information can be shared with AI agents, enhancing data loss prevention (DLP) measures to scan for sensitive visual data, and providing comprehensive training to developers on the secure use of AI coding tools. The incident serves as a critical reminder that the convenience offered by AI should not come at the expense of robust security practices, especially when dealing with proprietary and customer data. The scale of the exposure, affecting over 300 organizations, points to a systemic challenge that the broader tech industry needs to address collaboratively to ensure the safe and responsible adoption of AI in software development.