### Adaptive Thinking Architecture Claude Fable 5 removes manual thinking budget configuration entirely. Adaptive thinking is the only supported mode: the model allocates reasoning tokens based on task complexity, and developers adjust depth through the effort parameter. Raw chain-of-thought output is never returned; only summarized or omitted thinking blocks are accessible. This architecture reduces per-request overhead on simple tasks while maintaining depth on complex multi-step reasoning without developer tuning. ### Context Window and Agentic Workflow Support The 1 million token context window operates at standard per-token pricing with no long-context surcharge. Combined with 128k output tokens per request, the model supports end-to-end workflows that previously required chunking or retrieval-augmentation. Features like compaction and context editing beta reduce token accumulation costs across long agentic sessions, making it practical for sustained multi-turn autonomous workflows. ### Refusal and Fallback Infrastructure Claude Fable 5 introduces structured refusal handling at the API level. Declined requests return HTTP 200 with a machine-readable stop_reason rather than an error, enabling deterministic programmatic routing. Server-side fallback via the fallbacks parameter and SDK middleware across five languages allow teams to retry declined requests against lower-tier models automatically, with fallback credit offsetting prompt cache costs on the retry.
### Computer Use and Agentic Capabilities GPT-5.4 represents a shift from passive text generation to active computer operation. The model processes screenshots and emits keyboard and mouse commands, achieving 75 percent on OSWorld Verified benchmarks, surpassing human performance at 72.4 percent. This enables autonomous agents that navigate desktop environments, operate browsers via Playwright, and execute multi step workflows across applications without manual intervention. ### Tool Search and Token Efficiency The introduction of tool search architecture addresses the cost implications of large tool ecosystems. Rather than consuming thousands of tokens per request to define all available tools, GPT-5.4 retrieves specific tool definitions on demand. This reduces token usage by 47 percent when using MCP servers, directly lowering API costs while enabling integration with extensive tool libraries that were previously prohibitively expensive to maintain in context. ### Professional Knowledge Work Performance On GDPval benchmarks spanning 44 occupations, GPT-5.4 achieves 83 percent professional equivalence, with specific strength in spreadsheet modeling at 87.3 percent accuracy. The model demonstrates particular utility for investment banking analysis, presentation generation, and document editing tasks that require maintaining context across extended workflows.
### Adaptive Intelligence Architecture Gemini 3.1 Flash-Lite introduces a practical approach to reasoning through its adjustable thinking levels. This feature allows developers to specify exactly how much computational depth the model applies to each request. For high frequency tasks like content moderation or translation, users can minimize latency. For complex UI generation or simulation tasks, deeper reasoning mode delivers precision comparable to larger tier models without the proportional cost increase. ### Speed and Efficiency Metrics The model achieves 363 tokens per second output speed while maintaining competitive benchmark scores including 86.9% on GPQA Diamond. This performance profile positions it specifically for real time applications where responsiveness directly impacts user experience. The architecture leverages Google's TPU infrastructure to deliver these speeds at $0.25 per million input tokens, creating a distinct value proposition for throughput intensive operations. ### Production Readiness Currently in preview status, Flash-Lite integrates directly into existing Google AI Studio and Vertex AI workflows. The 1 million token context window enables processing extensive documentation or video content in single passes. However organizations should evaluate the preview status against their stability requirements for mission critical deployments.
### Efficient Edge Deployment Architecture LFM2.5-1.2B-Thinking delivers data center reasoning capabilities to mobile environments through specialized quantization and NPU optimization. The model maintains 52 tok/s decoding speed at 16K context on AMD Ryzen NPUs using FastFlowLM runtime. Memory efficiency peaks at 720MB on Snapdragon 8 Elite devices while supporting 32K token contexts. ### Curriculum RL Training Methodology The training pipeline employs parallel domain specific tracks rather than simultaneous multi domain training. This approach uses iterative model merging to combine specialized checkpoints for math, reasoning, and tool use without capability interference. The doom loop mitigation strategy reduces repetitive generation from 15.74% to 0.36% through asymmetric ratio clipping and dynamic filtering. ### Hardware Ecosystem Integration Day zero support spans Qualcomm Hexagon NPUs, AMD XDNA, and Apple Neural Engine through partnerships with Nexa AI and FastFlowLM. The open weight distribution includes native integration with llama.cpp, MLX, and vLLM frameworks. Developers can deploy across smartphones, IoT devices, and embedded systems without cloud dependencies. ```
### Diffusion Architecture Advantage Mercury 2 represents a fundamental departure from autoregressive transformer architectures. By employing parallel refinement across the entire sequence rather than sequential token prediction, the model eliminates the linear latency growth typical of traditional LLMs. This diffusion approach generates multiple tokens simultaneously and converges through iterative refinement, fundamentally altering the speed quality curve that has constrained production AI systems. ### Production Latency Optimization The model specifically targets compounding latency in agentic workflows where inference calls chain across dozens of steps. Traditional reasoning models increase latency proportional to test time compute, making multi step agents impractical for real time applications. Mercury 2 maintains reasoning grade quality within strict latency budgets, enabling complex agent loops that previously required sacrificing either intelligence or responsiveness. The 128K context window further supports stateful agent operations without frequent context window resets.
### Infrastructure Architecture GPT-5.3-Codex-Spark represents OpenAI's first production deployment on non GPU inference infrastructure, utilizing Cerebras' Wafer Scale Engine 3. This 4 trillion transistor processor eliminates the memory bandwidth bottlenecks inherent in discrete GPU architectures, enabling the 1000+ tokens/second throughput. The WebSocket based persistent connection architecture reduces roundtrip overhead by 80%, fundamentally changing the latency profile for interactive development tools. ### Workflow Differentiation Unlike GPT-5.3-Codex, which optimizes for autonomous execution over extended durations, Spark is explicitly tuned for collaborative iteration. The model's "lightweight" editing philosophy, minimal targeted changes without automatic test execution, prioritizes responsiveness over comprehensiveness. This creates a distinct use case: Spark excels at exploration and rapid prototyping where developer direction changes frequently, while standard Codex handles substantial refactoring requiring sustained autonomous operation. ### Performance Benchmarking On SWE-Bench Pro and Terminal Bench 2.0, Spark demonstrates that reduced latency need not sacrifice capability. The model completes software engineering tasks in a fraction of GPT-5.3-Codex's time while maintaining competitive accuracy metrics. This performance profile makes Spark particularly effective as a sub-agent in multi agent workflows, handling read heavy exploration and summarization tasks that feed into main agents running deeper reasoning models.
### Long Horizon Planning Capabilities Claude Sonnet 4.6 demonstrates sophisticated multi step strategic thinking evident in the Vending Bench Arena evaluation. The model developed an autonomous strategy of heavy capacity investment during initial simulation phases followed by a sharp profitability pivot. This temporal reasoning capability translates to real world business applications where models must balance immediate execution against long term objectives without human micromanagement. ### Computer Use Implementation The 72.5% OSWorld Verified score represents substantial progress in GUI automation since October 2024's initial 14.9% baseline. Sonnet 4.6 processes screen states as visual inputs rather than requiring structured API access, enabling integration with legacy systems lacking modern interfaces. Security considerations include enhanced prompt injection resistance compared to version 4.5, though enterprises should implement additional safeguards when deploying autonomous browser agents on untrusted domains. ### Context Window Utilization The 1M token capacity supports workflows previously requiring chunked processing or retrieval augmentation. In software engineering contexts, this enables holistic codebase comprehension where the model maintains architectural consistency across thousands of lines. API users should note this capability remains beta restricted, requiring specific implementation patterns for production deployment.
### Benchmark Performance and Reasoning Architecture Gemini 3.1 Pro establishes a new baseline for reasoning centric AI models, distinguished by its 77.1% verification on ARC-AGI-2. This benchmark specifically tests adaptability to novel logic patterns rather than memorized knowledge, indicating genuine advancement in core cognitive architecture. The 2x improvement over Gemini 3 Pro suggests significant refinements in the model's chain of thought capabilities and abstract pattern recognition. ### Agentic Workflow Integration The model's rollout across Google's development stack—including Antigravity and Android Studio—positions it as infrastructure for autonomous agentic systems. Its ability to generate functional code assets like animated SVGs and interactive 3D simulations demonstrates practical utility beyond text generation. These capabilities enable developers to prototype sensory rich interfaces and complex data visualizations without manual coding of graphics pipelines. ### Multi-Modal Output Capabilities Unlike text-only LLMs, 3.1 Pro generates executable code outputs that render as visual and interactive experiences. The synthesis of hand tracking integration with generative audio in the starling murmuration example reveals sophisticated cross modal reasoning. This technical architecture supports iterative creative workflows where models don't just suggest designs but produce deployable interactive assets.
The introduction of tool search architecture addresses the cost implications of large tool ecosystems. Rather than consuming thousands of tokens per request to define all available tools, GPT-5.4 retrieves specific tool definitions on demand. This reduces token usage by 47 percent when using MCP servers, directly lowering API costs while enabling integration with extensive tool libraries that were previously prohibitively expensive to maintain in context.
On GDPval benchmarks spanning 44 occupations, GPT-5.4 achieves 83 percent professional equivalence, with specific strength in spreadsheet modeling at 87.3 percent accuracy. The model demonstrates particular utility for investment banking analysis, presentation generation, and document editing tasks that require maintaining context across extended workflows.
The model achieves 363 tokens per second output speed while maintaining competitive benchmark scores including 86.9% on GPQA Diamond. This performance profile positions it specifically for real time applications where responsiveness directly impacts user experience. The architecture leverages Google's TPU infrastructure to deliver these speeds at $0.25 per million input tokens, creating a distinct value proposition for throughput intensive operations.
Currently in preview status, Flash-Lite integrates directly into existing Google AI Studio and Vertex AI workflows. The 1 million token context window enables processing extensive documentation or video content in single passes. However organizations should evaluate the preview status against their stability requirements for mission critical deployments.
The training pipeline employs parallel domain specific tracks rather than simultaneous multi domain training. This approach uses iterative model merging to combine specialized checkpoints for math, reasoning, and tool use without capability interference. The doom loop mitigation strategy reduces repetitive generation from 15.74% to 0.36% through asymmetric ratio clipping and dynamic filtering.
Day zero support spans Qualcomm Hexagon NPUs, AMD XDNA, and Apple Neural Engine through partnerships with Nexa AI and FastFlowLM. The open weight distribution includes native integration with llama.cpp, MLX, and vLLM frameworks. Developers can deploy across smartphones, IoT devices, and embedded systems without cloud dependencies.
A Human Edited Software Directory
The model specifically targets compounding latency in agentic workflows where inference calls chain across dozens of steps. Traditional reasoning models increase latency proportional to test time compute, making multi step agents impractical for real time applications. Mercury 2 maintains reasoning grade quality within strict latency budgets, enabling complex agent loops that previously required sacrificing either intelligence or responsiveness. The 128K context window further supports stateful agent operations without frequent context window resets.
GPT-5.3-Codex-Spark represents OpenAI's first production deployment on non GPU inference infrastructure, utilizing Cerebras' Wafer Scale Engine 3. This 4 trillion transistor processor eliminates the memory bandwidth bottlenecks inherent in discrete GPU architectures, enabling the 1000+ tokens/second throughput. The WebSocket based persistent connection architecture reduces roundtrip overhead by 80%, fundamentally changing the latency profile for interactive development tools.
Unlike GPT-5.3-Codex, which optimizes for autonomous execution over extended durations, Spark is explicitly tuned for collaborative iteration. The model's "lightweight" editing philosophy, minimal targeted changes without automatic test execution, prioritizes responsiveness over comprehensiveness. This creates a distinct use case: Spark excels at exploration and rapid prototyping where developer direction changes frequently, while standard Codex handles substantial refactoring requiring sustained autonomous operation.
On SWE-Bench Pro and Terminal Bench 2.0, Spark demonstrates that reduced latency need not sacrifice capability. The model completes software engineering tasks in a fraction of GPT-5.3-Codex's time while maintaining competitive accuracy metrics. This performance profile makes Spark particularly effective as a sub-agent in multi agent workflows, handling read heavy exploration and summarization tasks that feed into main agents running deeper reasoning models.
Claude Sonnet 4.6 demonstrates sophisticated multi step strategic thinking evident in the Vending Bench Arena evaluation. The model developed an autonomous strategy of heavy capacity investment during initial simulation phases followed by a sharp profitability pivot. This temporal reasoning capability translates to real world business applications where models must balance immediate execution against long term objectives without human micromanagement.
The 72.5% OSWorld Verified score represents substantial progress in GUI automation since October 2024's initial 14.9% baseline. Sonnet 4.6 processes screen states as visual inputs rather than requiring structured API access, enabling integration with legacy systems lacking modern interfaces. Security considerations include enhanced prompt injection resistance compared to version 4.5, though enterprises should implement additional safeguards when deploying autonomous browser agents on untrusted domains.
The 1M token capacity supports workflows previously requiring chunked processing or retrieval augmentation. In software engineering contexts, this enables holistic codebase comprehension where the model maintains architectural consistency across thousands of lines. API users should note this capability remains beta restricted, requiring specific implementation patterns for production deployment.
Gemini 3.1 Pro establishes a new baseline for reasoning centric AI models, distinguished by its 77.1% verification on ARC-AGI-2. This benchmark specifically tests adaptability to novel logic patterns rather than memorized knowledge, indicating genuine advancement in core cognitive architecture. The 2x improvement over Gemini 3 Pro suggests significant refinements in the model's chain of thought capabilities and abstract pattern recognition.
The model's rollout across Google's development stack—including Antigravity and Android Studio—positions it as infrastructure for autonomous agentic systems. Its ability to generate functional code assets like animated SVGs and interactive 3D simulations demonstrates practical utility beyond text generation. These capabilities enable developers to prototype sensory rich interfaces and complex data visualizations without manual coding of graphics pipelines.
Unlike text-only LLMs, 3.1 Pro generates executable code outputs that render as visual and interactive experiences. The synthesis of hand tracking integration with generative audio in the starling murmuration example reveals sophisticated cross modal reasoning. This technical architecture supports iterative creative workflows where models don't just suggest designs but produce deployable interactive assets.