GPT-5.5 in Codex and Codex CLI is the legal-tech engineering story behind the legal research story. Per OpenAI's launch announcement on April 23, 2026, GPT-5.5 rolled out simultaneously to ChatGPT and to Codex — OpenAI's developer-focused product line — with the new model accessible via Codex CLI for command-line and IDE workflows. For law firms running their own legal-tech engineering (in-house developers, legal-ops leads building agents, IT teams maintaining custom integrations), this is the model that makes the build-your-own track competitive against vendor wrappers. For BigLaw firms with established legal-tech teams, this is the upgrade path for the agents and tools already in production.

Codex CLI is OpenAI's command-line interface for code-focused workflows — installation, code generation, debugging, refactoring, and agent orchestration directly from the developer's terminal. With GPT-5.5 as the underlying model as of April 23, the CLI inherits the calibration, 1M context, and tool-call improvements covered in the GPT-5.5 anchor.

For legal-tech teams, the practical use cases are three. First, building research agents that integrate with firm databases (Westlaw, Lexis, prior matter repositories, conflict-check systems). Second, generating one-off scripts for matter-specific tasks (parse a 600-document discovery production, extract entities from deposition transcripts, normalize contract clauses across a portfolio). Third, maintaining and debugging existing legal-tech infrastructure as it scales.

The operational unlock with 5.5: the model's improved tool-call coherence (covered in the tool calls and legal research coherence spoke) means agents built with Codex CLI are more reliable in production. Pre-5.5, brittle tool-call behavior meant every legal-tech build needed a robust error-handling shim. Post-5.5, the model handles much of the error recovery natively, which means smaller teams can build and maintain more agent infrastructure than was viable in 2025.

Legal-tech build economics have been gated on three resources: developer time, model reliability, and integration complexity. GPT-5.5 in Codex CLI moves all three.

Developer time: Codex CLI's code generation lets a competent developer prototype a research agent in 4-8 hours that would have taken 40-80 hours of bespoke work pre-5.5. The model handles boilerplate — API client wrappers, prompt templates, error-handling skeletons — and the developer focuses on legal-specific logic.

Model reliability: as covered in the tool-call coherence spoke, agentic workflows that weren't reliable enough for production pre-5.5 are reliable enough now. That changes the build-vs-buy decision for mid-market firms.

Integration complexity: Codex CLI's reasoning over codebases means it can read your firm's existing tech stack (matter management, document management, conflict checks, time entry) and propose integrations that fit. Pre-5.5, integration design needed senior developer judgment. Post-5.5, the model handles the design draft and the developer reviews.

The practical result: a mid-market firm with one in-house legal-tech developer can now maintain agent infrastructure that previously required two or three. The Pro vs standard upgrade decision covers when to push complex agentic builds to Pro.

Vertical legal AI vendors (Harvey, Spellbook, CoCounsel) have justified premium pricing partly on the orchestration layer above the foundation model — prompt engineering, tool integrations, workflow templates, vendor-managed updates. Per the GPT-5.5 vs Harvey AI / CoCounsel vendor decision spoke, the calculus shifts when foundation models handle orchestration competently with developer-friendly tooling.

For a firm with even modest legal-tech engineering capacity (one developer or a competent legal-ops lead), building a research agent on GPT-5.5 with custom tool integrations is feasible in 1-2 weeks of work. Per-query API costs run $0.15-$0.60 depending on complexity. Maintenance burden is moderate — model updates, prompt tuning, integration drift — but lower than pre-5.5.

For firms without engineering capacity, vendor wrappers still make sense. Harvey's value proposition for AmLaw 100 firms still includes vendor-managed compliance, vendor-managed support, and vendor-managed updates that some firms value. The break-even depends on firm size and engineering investment.

The second-order angle: vendor pricing pressure. As foundation-model orchestration improves, vendor wrappers need to differentiate on more than "we built the agent for you." The vendors that survive will compete on workflow templates, regulatory compliance, and industry-specific data — not on the orchestration layer alone. The Anthropic eating the legal stack analysis covers this broader vendor-vs-foundation tension.

Five Codex CLI workflows that legal-tech teams are running in production as of April 2026:

- Discovery production parsers — Codex generates matter-specific scripts that ingest a discovery production (often delivered as a multi-format archive: PDFs, native files, image OCR), normalize the documents, and pipe them into the firm's review platform. With 5.5's 1M context, the parser can also generate first-pass relevance summaries. - Contract clause extractors — given a portfolio of agreements, generate scripts that extract specific clause types (indemnification, limitation of liability, governing law) into a structured database for portfolio analysis. - Citation verification pipelines — automate the verification step covered in the citation verification protocol spoke. Take an associate's draft brief, extract citations, query Westlaw or Lexis for each, and flag any that don't verify cleanly. - Conflict-check enhancers — augment the firm's existing conflict-check tooling with name-normalization and entity-resolution logic that catches conflicts the rule-based system misses. - Time entry assistants — generate time entries from associate work logs, calendar entries, and document-modification logs.

Each of these is a 1-3 week build for a competent developer with Codex CLI assistance. The same builds were 3-12 weeks of work pre-5.5. That's the productivity unlock.

Cost structure and procurement for Codex CLI deployments

Codex CLI billing runs on the same OpenAI API rates as direct GPT-5.5 use: $5/M input + $30/M output for standard, $30/M input + $180/M output for Pro per OpenAI's API pricing. Cached input drops to $0.50/M (90% off) for repeated context. Batch API runs at 50% off.

For a legal-tech team running development workflows (code generation, debugging, refactoring), monthly Codex CLI spend lands in the $300-$1,500 range per developer depending on usage intensity. That's well within the legal-tech engineering budget of any firm with 10+ attorneys.

For production agent deployments serving the firm's attorney base, costs scale with attorney query volume. A 25-attorney firm running 200 queries per attorney per month (5,000 queries firm-wide) lands around $625/month in API costs at standard rates — see the API pricing firm cost analysis for full modeling.

Procurement-wise, OpenAI API access requires a separate procurement track from ChatGPT Business or Enterprise. Legal-tech teams typically run the API on a firm-managed key with usage logging and per-team budget controls. For firms running both ChatGPT Business (for attorney chat use) and OpenAI API (for legal-tech engineering), the two procurement tracks are independent — same vendor, different contracts.

The Bottom Line: My take: GPT-5.5 in Codex CLI is the change that makes in-house legal-tech engineering competitive against vendor wrappers for mid-market firms. The build-vs-buy threshold dropped meaningfully on April 23. For firms with one in-house developer and a willingness to maintain custom infrastructure, the per-matter economics now beat most vendor wrapper pricing. For firms without engineering capacity, vendor wrappers still make sense — but the gap narrowed.

AI-Assisted Research. This piece was researched and written with AI assistance, reviewed and edited by Manu Ayala. For deeper takes and the perspective behind the research, follow me on LinkedIn or email me directly.