TechPrufer
Get started
Harvey
Harvey logoHarvey · Feature trail

Multi-Model AI Platform

9 posts · Apr 2025 – May 2026

What this trail covers

Harvey's platform continues to evolve its AI agent capabilities, shifting the paradigm from prompt-based assistance to task-oriented execution. The system now empowers agents to autonomously plan, research, execute, and deliver finished work products, with lawyers focusing on goal setting and final review. This evolution spans transactional work, litigation, compliance, and in-house operations, aiming to increase leverage by automating repeatable, high-volume tasks. The platform now integrates multiple leading foundation models (Anthropic, Google, OpenAI) and provides users with auto-routing to the best model for a given task, or the option to select a specific model. This multi-model approach allows for optimization of individual model strengths for specific sub-tasks within complex agents and workflows. The platform also emphasizes transparent evaluation through benchmarks like BigLaw Bench and expert reviews to guide model selection and demonstrate system performance.

All Harvey trails
  1. Claude Opus 4.8, Now Live in HarveyThis post announces the integration of Claude Opus 4.8 into Harvey's multi-model AI platform. It highlights Opus 4.8's performance improvements on the Legal Agent Benchmark (LAB) and BigLaw Bench, noting its ability to break 10% on LAB's all-pass scoring and achieve 91.1% on BigLaw Bench. The post also mentions qualitative feedback regarding the model's self-correction and editing behavior, while noting areas for improvement in handling ambiguity and surfacing judgment calls. The rollout strategy for different regions is also detailed.
  2. Claude Opus 4.7, Now Live in HarveyThis post announces the integration of Claude Opus 4.7 into Harvey's multi-model AI platform. It highlights Opus 4.7's improved reasoning calibration, stronger performance on ambiguous tasks, and increased substantive accuracy, evidenced by a 90.9% score on the BigLaw Bench evaluation. The post also notes qualitative feedback regarding minor polish gaps compared to Opus 4.6.
  3. Why Harvey is Multi-Model by DesignThis post details the architectural commitment to a multi-model AI platform, emphasizing the strategic decision to integrate multiple frontier models from different providers (Anthropic, Google DeepMind, OpenAI). It explains the rationale behind this choice, focusing on three pillars: quality (matching models to specific tasks based on their strengths), reliability (ensuring service continuity by routing work to alternative models during provider disruptions or outages), and choice (allowing customers to control which models are available and to select preferred models). The post also touches on operational aspects like regional availability, data sovereignty, and enforcing security/compliance standards across providers. It mentions continuous evaluation through BigLaw Bench and expert preference testing as methods for model selection.
  4. Sonnet 4.6, Now Live in HarveyThis post announces the integration of Claude Sonnet 4.6 into Harvey's multi-model AI platform. It details Sonnet 4.6's performance on the BigLaw Bench evaluation suite (87.6% score, 35% perfect scores) and highlights its strengths in numerical precision, thoroughness on ideation tasks, and root cause analysis in code. It also notes a behavioral observation: Sonnet 4.6 is precise with explicit instructions but less so with implicit ones, especially for multi-document synthesis.
  5. Opus 4.6, Now Live in HarveyThis post announces the integration of Claude Opus 4.6 into Harvey's multi-model AI platform. It highlights Opus 4.6's performance on the BigLaw Bench evaluation suite (90.2% score), its strengths in substantive accuracy, source grounding (over 120 inline citations), and multi-step legal reasoning. A specific failure mode observed is the model's tendency to over-contextualize responses when uncertain, leading to unnecessary tokens. The post also mentions ongoing work to refine how extended thinking is deployed for legal workflows requiring succinct output. Rollout is planned for US and EU clients, with AU availability to follow.
  7. Building in Active Guardrails With GPT-5.2This post details the integration of GPT-5.2 into Harvey's platform, highlighting its enhanced 'capability awareness' which allows it to reason about its own limitations and proactively communicate constraints to users. This moves the system from passive execution to active guardrails, reducing hallucinations and silent failures. Key improvements include better reasoning about impossible tasks, improved citation generation for transparent sourcing, and strict adherence to formatting standards for structural integrity. This enhances the AI's ability to act as a transparent partner and a more reliable coworker.
  8. Building a Legal Coworker with GPT-5This post announces the integration of GPT-5 into the Harvey platform, highlighting its enhanced reasoning, tool use, and drafting capabilities. It details how GPT-5 enables more complex, multi-stage workflows that can branch, loop, and self-correct, moving Harvey towards an 'intelligent coworker' paradigm. The post quantifies GPT-5's performance improvement on the BigLaw Bench evaluation suite (5.09% over previous frontier models) and provides an example of an orchestrated agent task for regulatory gap analysis. It emphasizes the shift from prompt-based execution to project-based management and proactive collaboration, leveraging GPT-5's ability to transition between in-depth reasoning and high-throughput tasks.
  9. Integrating Deep Research into HarveyThis post details the rapid integration of OpenAI's Deep Research API into Harvey's platform within 12 hours of its release. It highlights the role of Harvey's AI-native architecture and iterative development culture in this speed. Key technical contributions include: 1. Integration of Deep Research's URL citations into Harvey's citation system for traceability and verification. 2. Leveraging the Workflow Engine's 'AI building blocks' and 'thinking states' to seamlessly integrate the new model's reasoning and actions, allowing for user evaluation of intermediate results and intervention. 3. Utilizing AI-assisted development practices, including AI coding assistants and Streamlit for rapid UI prototyping, to overcome challenges with new and undocumented APIs. The post also mentions open-sourcing starter code for the Deep Research integration.
  10. Expanding Harvey's Model OfferingsThis post details the integration of multiple leading foundation models (Anthropic, Google) alongside existing OpenAI models into the Harvey platform. It introduces an auto-routing mechanism to select the best model for a given legal task and a model selector for user-driven choice. The technical rationale for this multi-model approach is explained, highlighting how different models have unique strengths and weaknesses for specific legal tasks (e.g., Gemini 2.5 Pro for drafting vs. trial preparation). The post also discusses the impact on agents and workflows, enabling more precise model selection for individual steps within complex processes. Evaluation methodologies, including the BigLaw Bench and expert reviews, are emphasized as critical for guiding model selection and demonstrating performance improvements.
  11. Resilient AI Infrastructure | HarveyThis post details the infrastructure supporting Harvey's multi-model AI platform, focusing on reliability and scalability. Key contributions include a centralized Python client library abstracting model interactions, a model health tracker, weighted model endpoint selection based on SLIs, a feature-aware rate limiting system using a Redis-backed token bucket algorithm, a model proxy for secure and instrumented access, and granular observability with burn rate alerts and token accounting. It also describes how new models are evaluated and integrated.