Anthropic Eating the Legal Stack 2026 Full Map

In 90 days Anthropic shipped Claude For Word (April 11, 2026), the Cowork legal plugin with `/review-contract` and `/triage-nda` skills (February 2026, open source), an open-source legal skill set on GitHub, a 20,000-lawyer workshop hosted with the Florida Bar (April 23, 2026), the Project Deal agent-to-agent transaction experiment with 69 employees and 186 completed deals (April 24, 2026), the Freshfields multi-year co-build covering 5,700 lawyers across 33 offices (April 23, 2026), and Claude Opus 4.7 powering all of it (April 16, 2026). While most legal AI vendors framed Anthropic as a foundation model, Anthropic was building the operating system. The signal showed up in Manu's data first — 2,100+ Microsoft Copilot citations of aivortex.io in the last 30 days, with "Harvey AI legal" as the top grounding query, and Claude recommending Vortex more than ChatGPT in the last 24 hours. This is the 90-day map of how Anthropic ate the legal stack while incumbents debated whether "AI co-counsel" was a feature or a category.

The 90-day timeline: 7 shipments that re-architected legal AI

Compressed timeline of what Anthropic shipped between February and late April 2026:

- February 2026 — Cowork legal plugin (open source). Two skills released to the Anthropic plugin marketplace: `/review-contract` (clause-by-clause review against a configured negotiation playbook with GREEN/YELLOW/RED flags) and `/triage-nda` (rapid pre-screening categorizing NDAs for standard approval, counsel review, or full review). Posted to GitHub under permissive license. Per LawSites coverage, the plugin was the "opening salvo" against legal-tech incumbents.

- February 2026 — $285B market reaction. In the trading session following Cowork's announcement, Thomson Reuters fell 16%, RELX (LexisNexis parent) fell 14%, and Wolters Kluwer fell 13%. Combined market cap impact across the three legal-data leaders: approximately $285 billion wiped in a single session. (Canadian Lawyer source)

- April 11, 2026 — Claude For Word. Anthropic shipped Claude inside Microsoft Word with contract review as the first listed legal use case. (Artificial Lawyer coverage)

- April 16, 2026 — Claude Opus 4.7. Powers everything above. Task budgets, multi-session memory, cybersecurity safeguards, 3.75 MP vision, and the new "xhigh" effort level.

- April 23, 2026 — Freshfields multi-year deal. 5,700 employees, 33 offices, +500% adoption in six weeks, co-development program, early access to future Anthropic models. Plans to expand to Cowork.

- April 23, 2026 — 20,000-lawyer Florida Bar workshop. Mark Pike (Anthropic AGC) and Maggie Russo (applied AI architect) presented to 20K+ registered lawyers on Cowork-driven contract review and NDA triage workflows. (Florida Bar source)

- April 24, 2026 — Project Deal. 69 Anthropic employees, $100 budgets each, 500 listed items, 186 completed deals, total transaction value just over $4,000 — agent-to-agent. The legal frameworks for this don't exist yet.

Seven shipments. One quarter. Different category from anything Harvey, CoCounsel, Spellbook, or Lexis Protégé released in the same window.

The architectural read: foundation model vs operating system

Most legal-tech vendors frame Anthropic as a model they license. Harvey runs on Claude under the hood. Spellbook offers Claude as one of its model options. Thomson Reuters' rebuilt CoCounsel embeds Anthropic. The vendor narrative: "Anthropic gives us the brains, we give you the legal product."

The 90-day shipping cadence tells a different story. Anthropic built:

- A direct-to-lawyer surface (Cowork plugin marketplace + Claude For Word inside Microsoft 365) - A direct-to-firm enterprise channel (Freshfields-style multi-year co-build) - A direct-to-developer tool (Claude Code, Claude Design) - An open-source path (Cowork plugin's GitHub repo — firms install without procurement signing a vendor) - A workshop-based education arm (20K-lawyer Florida Bar event) - An experimental research arm (Project Deal — testing the post-vendor agentic future)

That's not a foundation model offering. That's a vertically integrated operating system targeting the same legal end-user that Harvey, CoCounsel, and Spellbook target. The vendors paid Anthropic for the model. Anthropic just shipped past them to the buyer.

The second-order read: when the underlying model vendor disintermediates the wrapper, wrapper economics get hard fast. Harvey's quote-only enterprise pricing assumes the firm pays Harvey for the legal layer on top of Claude. If a firm can install the Cowork plugin from GitHub, configure its own playbook, and pipe Claude For Word into the same workflow, the wrapper's value proposition narrows to UI polish and customer success.

First-party data: what aivortex.io's Bing AI Performance shows in real time

Vortex has been publishing legal-AI analysis since February 2026. The Bing AI Performance dashboard — free since 2025, surfaced inside Bing Webmaster Tools — shows which queries trigger Microsoft Copilot to cite a domain.

Last 30 days on aivortex.io: 2,100+ Copilot citations. Top grounding query: "Harvey AI legal." Spellbook and Everlaw follow. In the last 24 hours specifically, Claude has been recommending aivortex.io more than ChatGPT — a pattern that didn't exist on Claude 4.6 and emerged after Opus 4.7's April 16 release.

What that means in plain English: when partners ask Microsoft Copilot "is Harvey AI worth it," Copilot is grounding answers in a domain that's not Harvey's. When attorneys ask Claude about legal AI tooling, Claude is increasingly grounding in independent analysis over vendor marketing copy. The AI engines are routing legal-AI questions toward grounded, vertical-specific content.

The second-order effect: every firm whose marketing strategy is "rank for our brand on Google" is missing the channel where their attorneys actually ask the question. The third-order effect: most firms haven't enabled Bing AI Performance, so they don't know which queries trigger their citations or — more importantly — which queries should but don't. The dashboard is free. Most firms haven't opened it. (read the operator's full take on the Anthropic disintermediation playbook)

The Cowork legal plugin: open source as procurement bypass

Cowork's `/review-contract` and `/triage-nda` skills are the most underrated piece of the 90-day arc. Per the Anthropic plugin page, the workflow is:

- /review-contract — feed Claude a contract, configure your negotiation playbook (acceptable indemnity caps, preferred forum, IP assignment defaults), and Claude returns a clause-by-clause review with GREEN (acceptable), YELLOW (negotiate), or RED (block) flags plus suggested redlines. - /triage-nda — feed Claude an inbound NDA, and it categorizes the document into standard-approval, counsel-review, or full-review buckets based on configured risk tolerance. Optimized for high-volume inbound.

The killer detail: it's open source, hosted at github.com/anthropics/knowledge-work-plugins/tree/main/legal. A firm's IT team can install it without a vendor RFP, MSA negotiation, security review against an external provider, or per-seat pricing commitment.

For in-house legal ops teams that have spent 18 months trying to build internal NDA triage tools through engineering tickets, that's an end-run around the entire procurement loop. The functionality ships free. The configuration is YAML. The deployment surface is Claude itself (Pro at $20/month, Team at $25/seat/month, or Enterprise quote-only).

The disclaimer matters: the plugin includes "assistance not advice" language baked into its outputs. That's the integrity guardrail — it's not pretending to be counsel. It's a paralegal-grade workflow shipped at developer prices. (deeper read on Cowork vs vendor plugins)

Claude For Word and the 90% M365 install base distribution play

Roughly 90% of US law firms run Microsoft 365. That's the install-base reality every legal-tech vendor builds against. Harvey ships a Word add-in. Spellbook lives inside Word. CoCounsel has Word integration. The M365 surface is the legal industry's actual operating system.

On April 11, 2026, Anthropic shipped Claude For Word — Claude embedded directly inside Microsoft Word, with contract drafting and review as the headline use case. Per Artificial Lawyer's coverage, the integration runs natively inside the Word interface lawyers already use, no separate browser tab, no copy-paste loop.

The distribution math: lawyers don't switch tools to use AI. They open Word, ask the AI for help on the document they're already in, and accept or reject the suggestion. Whichever vendor gets that workflow gets the moat. Microsoft Copilot is one player there. Anthropic just became another — pricing access through Claude tiers ($20-25/user/month for Pro/Team) instead of Microsoft's $30/user/month Copilot M365 add-on.

The second-order effect: a firm now has multiple ways to put AI inside Word. Microsoft Copilot ($30/user). Claude For Word ($20-25/user via Claude). Harvey's Word add-in (quote-only enterprise). Spellbook's Word add-in (quote-only). The firm's procurement question shifts from "which AI vendor?" to "which AI tier inside Word, and which of them does the work my associates actually need?" The vendor with the best price-to-capability ratio at the Word surface wins.

For mid-market firms running M365 Business Premium with Copilot bundled at $32/user/month annual (Microsoft pricing), adding Claude Pro at $20/user/month for the lawyers doing the heaviest contract work is a sub-$250/lawyer/year incremental decision. That's a procurement conversation that doesn't need a CFO sign-off.

The Freshfields co-build: enterprise procurement velocity at AmLaw scale

On April 23, 2026, Freshfields and Anthropic announced a multi-year collaboration agreement. Per the Freshfields press release and Law.com's reporting, the deal includes:

- 5,700 employees at Freshfields with Claude access via the firm's proprietary AI platform - +500% adoption increase in the first six weeks of pilot deployment - 33 offices globally spanning all practice groups + business services - Co-development program for legal-focused AI applications + agentic workflows - Plans to expand to Cowork (Anthropic's agentic AI platform) - Early access to future Anthropic models - Early adopter of Thomson Reuters' rebuilt CoCounsel Legal (Anthropic-powered, Westlaw + Practical Law embedded)

The procurement signal that matters: Freshfields didn't sign a Harvey or CoCounsel deal at the firm-wide layer. They built directly with Anthropic and layered TR's rebuilt CoCounsel on top. That's the inverse of typical enterprise legal AI procurement. (deeper read on the Freshfields × Anthropic deal in Cluster 5)

The +500% adoption figure deserves a pause. That's bottom-up usage growth — lawyers reaching for the tool because it solves work, not top-down mandate enforcing tool use. It's also the metric most firm-wide AI deployments fail to hit. Freshfields hit it in six weeks because the Claude-via-Anthropic deployment surface is the same surface lawyers already use for personal AI tasks (claude.ai, Claude in Word). Adoption pre-existed the procurement. The deal formalized usage that was already happening.

For BigLaw firms on the fence about co-build versus buy, the Freshfields data point is a forcing function. (read the firms most likely to follow)

Project Deal and the post-vendor agentic future

On April 24, 2026, Anthropic published "Project Deal" — an experimental agent-to-agent transactional marketplace. 69 employees, $100 budgets each, 500 listed items, 186 completed deals, total transaction value just over $4,000.

Per Artificial Lawyer's reporting, the agents represented buyers and sellers, negotiated terms, and closed deals without human-in-the-loop counsel. Per Legal IT Insider, the legal frameworks for this don't yet exist.

The headline gap: when an AI agent acts as buyer or seller, which Model Rule covers attorney supervision? Who carries fiduciary duty when the agent handles both sides of a transaction? When agents agree to escrow, who arbitrates? The Restatement, the UCC, and most state ethics opinions don't have answers calibrated to the agentic reality.

The second-order read: most legal-tech vendors (Harvey, CoCounsel, Spellbook) framed their tools as "co-counsel for the lawyer." Project Deal is the experimental signal that Anthropic is also prototyping what happens when buyer and seller are agents — meaning lawyers exit the loop entirely for routine transactional work. That's a different competitive frame. Vendors who thought their moat was "we make lawyers faster" are looking at a model vendor experimenting with "we make some lawyers unnecessary for some transactions."

The third-order read: this is the longest-fuse strategic signal of the 90-day arc. Cowork plugin and Claude For Word affect 2026 procurement. Project Deal affects 2028-2030 firm strategy. Firms that don't have a Project Deal-aware risk-and-ethics conversation in 2026 will be reactive in 2028. (read the Project Deal future of transactional law spoke)

Why Anthropic is winning legal vs other foundation models

OpenAI shipped GPT-5.5 on April 23, 2026 — the same day as the Freshfields announcement. Google has Gemini 3.1 Pro deployed across Workspace. Microsoft Copilot dominates the M365 install base. xAI's Grok ships into X. Mistral and Cohere ship enterprise variants.

None of them shipped what Anthropic shipped in the same 90 days for legal. The structural reasons:

- Better legal calibration. Per Anthropic's Opus 4.7 documentation, the model is calibrated to be "less likely to proceed confidently with a bad plan." For legal research where citation accuracy is malpractice-grade, that's the differentiator. Vortex's first-party data confirms it — Claude is grounding in vertical legal content more conservatively than ChatGPT in the last 24 hours.

- Faster shipping cadence with legal as a named vertical. Cowork plugin, Claude For Word, and the Florida Bar workshop were all explicitly legal-targeted releases. OpenAI's legal plays are vendor-mediated (Harvey runs on OpenAI variants). Google's legal plays are M&A-acquired (Doctrine via RELX is parallel, not Google).

- Willingness to disintermediate own customers. Cowork open source and Project Deal are both direct disintermediation experiments. OpenAI's commercial DNA leans toward partnership and licensing. Anthropic shipped the open-source legal skills to GitHub the same quarter Harvey raised at $11B. That's an explicit "we'll compete with our customers" stance.

- Enterprise procurement velocity at Freshfields scale. Freshfields's 5,700-employee, six-week-to-+500% deployment is the public proof-of-concept that Anthropic's enterprise channel works at AmLaw scale. No equivalent OpenAI deal has been announced publicly in 2026.

The pickable side: Anthropic isn't winning legal on benchmarks (GPT-5.5 has the larger context window, better speed). Anthropic is winning legal on shipping cadence, calibration tuning for legal use cases, willingness to disintermediate vendor wrappers, and enterprise procurement velocity. That's a structural advantage that compounds over the next 6-12 months. (read the full why-Anthropic-is-winning analysis)

What this means for firm strategy: Anthropic-aware procurement in 2026

Translation of the 90-day arc into firm-level decisions:

Solo and small firms (1-10 attorneys). Claude Pro at $20/month per attorney plus the Cowork plugin (free, open source) is the minimum viable AI stack for 2026. Skip Harvey ($1,200-2,000+/seat/month per industry estimates, quote-only — (Artificial Lawyer pricing reporting)). Skip CoCounsel ($75-$500/user/month per Costbench March 2026 secondary-source pricing, Costbench tracker, not vendor-confirmed). Run Claude Pro plus the open source plugin and a free Westlaw or Lexis subscription for citation verification.

Mid-size firms (10-100 attorneys). Claude Team at $20/seat/month annual or $25/seat/month monthly. Combine with the open-source Cowork plugin configured to your negotiation playbook. Add Claude For Word for contract-heavy attorneys. Total stack cost: $20-25/seat/month plus the Word integration, vs $300-500/seat/month for CoCounsel Core or quote-only Spellbook seats. Re-evaluate Harvey only if the firm has a specific BigLaw-grade use case (litigation matter management, M&A closings at scale).

BigLaw and AmLaw 100. The procurement question is not "do we go Anthropic-deep?" but "on which deployment surface?" Five options: claude.ai Enterprise, Claude API, AWS Bedrock, Vertex AI, or Microsoft Foundry. For firms with M365 already deployed, Foundry has the fastest procurement path. For firms with active Anthropic deals (Freshfields is the public reference; more are in negotiation per Law.com's reporting), Foundry or Bedrock typically wins on velocity.

In-house legal teams (any size). The Cowork plugin is the wedge. Configure `/triage-nda` to your inbound NDA risk tolerance. Configure `/review-contract` to your standard playbook. Run Claude Pro or Team. Skip the vendor RFP cycle. Total deployment time: a working week. Total cost: under $500/year for a 5-person legal ops team. (read the in-house counsel deployment checklist)

The meta-point: Anthropic's 90-day arc shifted procurement from a multi-vendor, $300-500/seat/month conversation to a single-vendor, $20-100/seat/month conversation. Firms that recognize the shift in 2026 build the right stack early. Firms that don't will rebuild it in 2027 at higher switching cost.

The Bottom Line: My take: Anthropic just ran the most aggressive legal-AI shipping quarter of the post-ChatGPT era. The Cowork plugin disintermediates wrappers. Claude For Word lands inside the workflow that already exists. Freshfields proves enterprise procurement can move at Anthropic-velocity. Project Deal signals the 2028 future. Opus 4.7 powers all of it with legal-grade calibration. The verdict on FIT: for solos and mid-market firms, the Anthropic stack is now the cheapest competent option in legal AI. For BigLaw, the deployment-surface decision (Foundry, Bedrock, Vertex) matters more than the model. For incumbents — Harvey, CoCounsel, Spellbook — the wrapper economics conversation just got harder. The $285B market reaction to the Cowork plugin in February was the signal. The 90-day shipping cadence is the proof.