Claude Opus 4.7 Anthropic Legal Context Window Task Budgets

Claude Opus 4.7 — released April 16, 2026 per Anthropic's announcement — is the model layer powering every other piece of Anthropic's 2026 legal stack. The dev coverage led with 87.6% on SWE-bench Verified and 94.2% on GPQA Diamond. The legal-relevant changes are different and underreported: task budgets that cap token spend deterministically across an agentic loop, multi-session memory persisting context across sessions via a scratchpad/notes file, 3.75-megapixel vision input (up from 4.6's 1.15 MP), the new "xhigh" effort level between high and max, and cybersecurity safeguards shipping by default for the first time. Pricing is unchanged at $5 per million input tokens and $25 per million output tokens per Anthropic's pricing page — but the new tokenizer increases token counts by 1.0 to 1.35x depending on content, so consumption-based bills went up materially this month at the same usage. Vortex's first-party data shows Claude has been recommending aivortex.io more than ChatGPT in the last 24 hours since Opus 4.7's release. This is the operator's read on how Opus 4.7's specific capabilities matter inside the Anthropic legal stack.

Task budgets: deterministic spend per matter inside Cowork plugin runs

Discovery document review has been the wedge use case for legal AI since 2023. The economic problem has always been the same: how do you tell a partner what the AI portion of a matter will cost when token consumption is unpredictable?

Opus 4.7's task budgets answer that. Per the Opus 4.7 documentation, the model can target a token budget across an entire agentic loop — say, 2 million tokens for a first-pass relevance review on a 50,000-document production. Claude tracks against the cap with a running countdown and prioritizes the highest-signal documents first. When it hits the cap, it stops gracefully and reports what it covered.

For a $5/M input + $25/M output Opus 4.7 deployment, a 2M-token budget on a typical 70/30 input/output split is roughly $22 per agent run. A partner can put that line item in a budget memo and defend it. That's a different conversation than "Claude usage is somewhere between $300 and $4,000 this month."

Applied to the Anthropic legal stack: when the Cowork plugin's `/review-contract` skill runs against a 50-clause MSA, task budgets cap the per-document review cost. When `/triage-nda` runs across 100 inbound NDAs in a batch, the budget governs the batch. When Claude For Word does multi-document analysis inside an open document, the budget caps the per-task spend.

The second-order effect: discovery vendors who priced themselves on "per-document fixed fee" assumptions are now competing against an in-house Cowork-on-Claude workflow with provable per-matter economics. The third-order effect: insurance carriers writing AI deployment policies will start asking firms whether they use models with task budgets, because predictability lowers operational risk. (read the /review-contract clause-by-clause guide)

Multi-session memory: long-running matters hold context across days

M&A diligence runs 5-15 days. Multi-day depositions span weeks of prep. White-collar matters can hold context for months. Until 4.7, every Claude session started cold — the lawyer re-loaded the matter, re-explained the parties, re-grounded the model in the facts. That's the context-loss tax. It made AI useful for one-shot tasks and frustrating for long-horizon work.

Opus 4.7's scratchpad/notes file persistence changes that math. Claude writes structured notes mid-session, the firm saves the file with the matter, and the next session reads it back. Claude resumes where it left off — same parties, same facts, same line of analysis. The model behaves more like an associate who took notes than a chatbot you re-introduce every morning.

For a 12-day M&A diligence engagement, that's roughly 12 hours of analyst re-priming time saved across the matter. At a $400/hr blended rate, that's $4,800 per matter recovered. For multi-matter practices running 6-10 simultaneous diligence engagements, the recovered hours compound.

Applied to the Anthropic legal stack: the Cowork plugin's `/review-contract` skill can now run across multiple sessions on the same M&A matter, holding context about the deal structure, party preferences, and previously-flagged clauses. Claude For Word inside Microsoft Word maintains the same context across days. Claude Design can hold prototype context across sessions when iterating on internal tools.

The operational caveat: scratchpad files are now firm-data assets. They contain matter-specific reasoning, party identities, and analysis pathways. Storage policy, retention, and access control belong in the AI use policy from week one. (firm AI policy template spoke)

Vision improvements: 3.75 MP for evidence review and OCR

Opus 4.7's vision input bumped to 2,576 pixels / 3.75 megapixels — up from 4.6's 1.15 MP. That's a 3.26x improvement on image fidelity. For evidence review, OCR on scanned discovery, and whiteboard capture from depositions, this is the difference between "we'll need a paralegal to retype" and "the model reads it cleanly the first pass."

Applied to the Anthropic legal stack: lawyers feeding Claude scanned discovery documents (often poor-quality PDFs from older case files) get materially better OCR accuracy. Deposition exhibits photographed by a paralegal can be analyzed directly without a re-typing step. Whiteboard photographs from M&A working sessions can be transcribed accurately into Claude For Word for follow-up drafting.

For litigation teams, that's a measurable workflow improvement. A typical 500-page document production with mixed scan quality previously required 20-40 hours of paralegal OCR cleanup before AI analysis was viable. With Opus 4.7's improved vision, that's compressed to 4-8 hours of spot-checking the model's OCR output rather than full retyping.

The second-order effect: evidence prep workflows that previously required dedicated OCR vendors (Adobe Acrobat Pro, ABBYY FineReader at enterprise pricing) can run inside Claude with no separate vendor relationship. The third-order effect: small firms that previously couldn't afford OCR vendor licenses can now do high-quality evidence review at Claude Pro pricing ($17-$20/user/month). (Claude For Word contract drafting spoke)

Tokenizer cost reality: pricing is same, bills go up

Per Anthropic's pricing page, Opus 4.7 is $5 per million input tokens and $25 per million output tokens — identical to 4.6 sticker pricing. Procurement teams reading the pricing page see no change.

What the pricing page doesn't say: the new tokenizer increases token counts on the same content by 1.0 to 1.35x depending on content type. Code-heavy content compresses better; legal prose with case citations and Latin phrases sits closer to the 1.35x ceiling. Per-task spend went up materially this month at the same usage.

A firm running $8,000/month of Claude consumption on 4.6 contract-review workflows is now running $8,800-$10,800 on the same workflow with 4.7. Most procurement teams haven't caught it because the $5/M input rate is unchanged. The bill arrives in May.

Applied to the Anthropic legal stack: the Cowork plugin's `/review-contract` skill, Claude For Word document analysis, Claude Design prototype generation, and Claude Code deployment all consume more tokens per task than they did on 4.6. Firms on Claude Pro ($17-$20/user/month flat per user) and Team ($20-$25/seat/month) are partially insulated by usage caps. The pain lands hardest on Enterprise consumption deals where pricing scales with token consumption.

The operational move: audit your enterprise contract for any consumption-based scaling clauses, and re-price your internal AI cost-recovery rates if you bill matter spend back to clients. The math compounds across multi-matter practices. (read the cost analysis comparison spoke from Cluster 1)

Cybersecurity safeguards: privileged context risk mitigated at model layer

*United States v. Heppner* (SDNY, February 17, 2026) ruled that written exchanges between criminal defendant Bradley Heppner and consumer Claude were not protected by attorney-client privilege or work-product doctrine. The court reasoned that Claude isn't an attorney, so privilege doesn't attach, and Heppner generated the materials independently of counsel direction. (Heppner explainer)

That ruling created a clear operational rule: keep privileged context out of consumer AI. The harder question — what about associates jailbreaking enterprise Claude for a privileged use case the firm explicitly didn't authorize? — sat in the org-policy layer. Until 4.7.

Opus 4.7 ships with automated detection and blocking for prohibited cybersecurity uses by default. It's the first Claude where the firm's "what if associates use it for X" risk is mitigated at the model layer, not just the policy layer. The model itself refuses or flags certain categories of misuse without relying on monitoring downstream.

For managing partners writing AI deployment policies, this is the first time the model layer carries some of the compliance weight. It doesn't replace policy, training, or audit logs. But it reduces the surface area where a single rogue prompt creates a privilege defense problem.

Applied to the Anthropic legal stack: the Cowork plugin's outputs (which carry "assistance not advice" disclaimer language as the integrity guardrail), Claude For Word's drafting outputs, and Claude Code's deployment scripts all run against the cybersecurity safeguards by default. For BigLaw firms whose risk-and-ethics committees stalled enterprise AI rollouts pending model-layer guarantees, 4.7 unlocks a procurement conversation that was frozen on 4.6.

The second-order effect: the next BigLaw RFP cycle will explicitly ask AI vendors which model versions ship cybersecurity safeguards by default. Vendors running on older models or third-party APIs without these safeguards lose RFP points.

The Bottom Line: My take: Opus 4.7 isn't a benchmark story for legal — it's a procurement story. Task budgets give partners predictable per-matter AI spend. Multi-session memory unlocks long-horizon work that 4.6 made painful. Vision improvements compress OCR workflows. Cybersecurity safeguards reduce a real malpractice surface. The tokenizer change quietly raises consumption bills by up to 35% — audit your enterprise contract this week. Inside the Anthropic legal stack, every other component (Cowork plugin, Claude For Word, Claude Design, Claude Code) gets better with Opus 4.7 underneath. For solos and mid-size firms on Pro or Team plans, the upgrade pays for itself in calibration improvements alone. For BigLaw on enterprise consumption deals, run the per-matter math against your specific workflow profile.