Gemini vs. Perplexity vs. ChatGPT vs. a Donkey: the "2026 best legal AI" is all of them

Why Model Lock-In Fails

The creative destruction of weekly new AI releases makes this the best of times and also the worst of times in tooling.

Larry Summers used the language of "stagflation" to describe a mixed reality where consumers experience inflation without corresponding growth in employment and productivity in real terms. We see many of our pro users experiencing "technical stagflation," where there is a proliferation of useful tools and at the same time, more than ever, users do not have a primary tool they want to purchase. On the one hand, every tool is unique, powerful, and cool. On the other hand, there will be something more powerful tomorrow. Ben Anderson identifies the related phenomenon of "technical deflation," where the ease and falling cost of building software creates incentives to delay—not because tools are unstable, but because tomorrow's tools will be even easier to use.[9]

Simultaneously, every AI company claims to have made the “one model to rule them all”. And many of them even succeed, for days or weeks at a time. Suddenly Gemini can see. Overnight, Claude can read 4 times more context. The pace of progress in AI is unfathomable. Every day, a new release, and with it, an entire family of models that become immediately obsolescent.

Releases are meaningful for legal tasks: a new benchmark upset and a new architecture that didn’t exist a week ago mean that your entire stack is replaced by something faster, bigger, or more specialized.

The pace is so relentless that even Sergey Brin, who stepped away from Google's daily operations years ago, is now back in the office almost every morning, drawn in by the rapid advances of the current AI wave.[1] When Google's co-founder returns to the frontline because the field is moving too fast to follow from afar, it reveals everything about the moment we're in: this era rewards adaptability, not loyalty to any one model.

For lawyers, this turbulence is not abstract. Every new release can meaningfully change the quality of drafting, summarization, extraction, statute mapping, or the ability to read a massive case file in one shot. A model gaining a larger context window or improved reasoning isn’t a novelty; it directly impacts the accuracy and thoroughness of the output a lawyer relies on. That means a workflow that felt “best-in-class” on Monday can quietly become second-rate by the end of the week, not because the lawyer did anything wrong, but because the underlying model landscape shifted again.

Technical stagflation means the tools get better faster than legal teams can standardize around them. The result: lawyers need systems that evolve at the speed of the models, not systems that freeze them to yesterday’s frontier.

Every frontier-model organization is shipping new versions in fast, sequential, iterative cycles, and this is the shape of technical stagflation: a world where innovation accelerates while stability collapses, where capability expands but workflows degrade unless they evolve with the frontier. It does not just explain the chaos; it makes one conclusion unavoidable: a model-agnostic future is not optional. It is the only way forward.

The Case for Model Routing

Technical stagflation forces the question into the open, but the truth is we needed a model router long before the model ecosystem became this chaotic. Models have always had different strengths, weaknesses, and blind spots. Some are exceptional at reasoning but slow on large files. Some are fast but brittle. Some can process massive context windows. Others fall apart when you feed them anything longer than a few pages. Even today we routinely see situations where GPT-5 performs better than GPT-4.1 on complex reasoning, but a smaller or cheaper model handles summarization or extraction faster and with fewer hallucinations.

For lawyers the stakes have always been higher. When your work depends on accuracy, completeness, and fast turnaround, letting a platform quietly select a model on your behalf simply is not good enough.

Legal practice requires intentionality. You need to know whether the model handling your brief can reason deeply, whether it can handle long files, whether it is prone to omissions, and whether another model would be a better match for the task. Blind trust in a default model is not a practice. It is, and will always be, a structural risk.

At jhana, we have seen firsthand how quickly legal users gravitate toward tools that let them choose the best model for the job. Lawyers deserve what software engineers take for granted: transparency, model choice, and full control over their stack. Not magic. Control.

This is exactly why we rebuilt our entire stack at jhana to be fully model-agnostic. The trend is clear: model diversity is not a temporary phase; it is the environment we live in. The system can route tasks across different models, adopt new releases the day they appear, and incorporate entirely new providers without disruption. This is not a patchwork or a temporary fix. It is a full LLM gateway architecture designed to stay current even as the frontier resets every week. A single-model platform will always fall behind. A model-agnostic platform stays alive by design.

Upload Limits & Throttles: ChatGPT vs Claude vs Perplexity vs jhana

Legal work is not general. It is structured, procedural, and rooted in a methodology that demands precision at every step. Every plaint, every written statement, every affidavit, every annexure follows a sequence and a standard. There is no room for guesswork, and there is very little tolerance for deviation. Most general-purpose AI platforms were not built for this reality. ChatGPT, Gemini, and Perplexity Pro all have their strengths, but they still impose hard physical limits that make serious legal work impossible.

File size caps, page limits, rate ceilings, token walls, and the absence of auto-continue show up at the worst possible time: usually when a lawyer is trying to upload a lengthy plaint with annexures, a combined case bundle stitched for listing, or a multi-hundred-page petition set with affidavits, exhibits, and typed sets.

ChatGPT caps individual files at 512 MB per file, with additional organisation-level quotas and token-processing ceilings.[2] Claude accepts roughly 30 MB per file in chat and warns that PDFs over 100 pages may require splitting or will be downgraded to text-only extraction.[3] Perplexity allows up to 50 MB per file, with up to 30 files per upload and 500 files in an enterprise repository.[4] Many of these limits apply even on paid and enterprise tiers, and consumer tiers are often stricter.

And processing limits matter just as much as upload limits: even if a large file uploads, the model may only read a fixed token slice. OpenAI community documentation notes a practical processing window of around 110,000 tokens, roughly 200 pages for many formats.[5] These constraints appear exactly when lawyers attempt to load thousand-page filings, stitched bundles, or agreements with extensive annexures. The platform technically accepts the file but cannot actually process it, creating a mismatch between legal work and general-purpose AI systems.

These platforms also provide no real methodological support. There is no guided workflow for preparing a plaint, drafting a written submission, reviewing a case bundle, analyzing complaints, or checking statutory compliance. Lawyers are forced to improvise inside a generic chat interface, which produces inconsistent results across teams and across matters. And none of these tools offer legal-grade prompt management. A professional handling sensitive case files needs structure, versioning, visibility, and control; not a blank prompt box that resets with every new chat.

The pricing makes the disconnect even harder to justify. For the same subscription cost, users are effectively locked into one model, even when that model is not the best fit for the task. There is no routing, no way to compare model strengths, and no ability to adopt the latest releases until the platform eventually decides to roll them out.

By contrast, jhana costs a fraction of that, yet gives lawyers multiple frontier models, unlimited routing, legal memory, and structured review. On general tools, your quota buys one model and a set of constraints. On jhana, the same spend buys model diversity, stability, and actual legal capability. The price is higher on their side. The value is higher on ours.

What is a GPT Wrapper, and should I be scared of them?

Legal work is rooted in doctrine, procedure, and documentation. It requires understanding, not mere prediction, and far more than just “attention”. And this is where domain infrastructure actually matters.

Users have always told us the same story: ChatGPT confidently cites a case that never existed, blends two unrelated judgments into one, and insists an order is reported when it is not. None of this is surprising. Frontier models were never trained on the structure, hierarchy, or chronology of Indian case law. At jhana, we built the parts of the system that frontier models simply do not provide.

Their own cut-off dates make this clear: GPT-4.1 stops in June 2024,[6] GPT-5 stops in September 2024,[7] and Claude Opus 4.5 stops in May 2025.[8] None of them carry the evolving, jurisdiction-specific record lawyers actually rely on.

The foundation of jhana is the National Legal Archive, which solves this by replacing model cut-off dates with structured, continuously updated legal memory. The archive gives the system recall, structure, and continuity, not whatever happens to be floating around the open internet, giving the model a map of the legal universe instead of forcing it to guess. At jhana, this is the principle we design for.

When a lawyer asks for a document review, the core question is not stylistic polish but completeness. Can the system work through a 1000 page filing with its verifications and annexures? Can it map definitions and follow obligations across an SHA to give the lawyer a full overview without missing a detail? General-purpose models cannot explain what they processed or what they skipped, because they cannot reliably ingest the entire document in the first place.

Many users come to us after discovering that Perplexity quietly ignored pages in a typed set or glossed over a clause that would materially change an agreement. This is why we built Suit. It provides a systematic, auditable read in the same way a registry officer or diligent lawyer would, with a record of what was checked, what was flagged, and why. Suit ingests full filings, stitched bundles, handwritten indic records, and scanned annexures as a single structured unit instead of breaking on file-size caps, token ceilings, or partial reads. Suit does not just read documents. It holds them together.

These pieces are not add-ons. They are the difference between an AI that produces text and an AI that assists with legal work. Frontier models can reason, summarize, and translate beautifully, but they cannot deliver legal reliability without the scaffolding that surrounds them. That scaffolding is the real work: the archive, the document structure, the auditing logic, the procedural understanding. It turns AI from just predicting the next token or word into a legal tool.

General-purpose systems will always treat legal work as another form of text. jhana treats it as a discipline. That is why the proprietary infrastructure matters. It is not a layer or a “wrapper” on top of the models; it is the foundation that makes the models useful for law in the first place.

Buyer's Remorse: Avoiding betting on a losing LLM

Models do not simply improve; they also regress, disappear, change behaviour, or get replaced without warning. Lawyers experience this directly. A model that handled large PDFs smoothly last month might suddenly refuse to load them after a silent update. Another might lose its ability to follow a clause chain after an internal tuning change. Some users tell us they wake up to find that their preferred model has been rate-limited, repriced, or removed entirely. None of this is theoretical. It is the daily reality of depending on third-party models that you do not control.

A model-agnostic system protects against this volatility. When one model breaks, another can take its place. When a new release appears, it can be adopted immediately. When a provider changes pricing or removes a feature, the workflow does not collapse. The system absorbs the shock because it was built to route around single-point failures. This is not just a technical benefit. It is a workflow benefit. It means that your drafting pipeline, your document review process, your research workflow, or your internal SOPs do not hinge on the stability of one model.

Model Behaviour on Legal Tasks: GPT-5 vs Claude 4.5 vs Gemini 3 Benchmarks

Most lawyers use whatever model their platform defaults to. They never see what the others can do. That is a blind spot. A big one.

Different models behave differently. Radically differently. Some reason better. Some extract better. Some read messy scans better. Some track definitions better. Some summarise better. Some run faster. Some hold context better. Some hallucinate less. Some map obligations across an SHA with precision. Some spot defects in a plaint with almost clerk-level discipline.

You cannot learn this from a marketing page. You only learn it by trying them side by side.

We see this at jhana every day. A model that struggles with a plaint’s annexures becomes brilliant at cross-referencing an agreement. Another that is slow for research becomes unbeatable for extraction. A model that fails on a typed set performs flawlessly on a scanned document. Users discover these differences within minutes, and after that, the idea of relying on a single model feels primitive.

Expanding your awareness is not experimentation for fun. It is professional literacy. Knowing your models is becoming part of knowing your craft. The lawyer who understands the strengths and weaknesses of multiple models is not “technical.” They are simply better equipped. And in a field that updates every week, being better equipped is now a competitive edge.

The Full-Stack Lawyer is a Technical Professional

There is a familiar idea in recovery programs: identity drives behaviour. People say “I am not a smoker” or “I am not a drinker” because naming the identity changes the decisions that follow. This is not spiritual guidance or manifestation. It is a practical philosophy of work. And it applies, unexpectedly, to the future of law firms.

A shift is underway. Lawyers are beginning to identify not just as practitioners of law, but as professionals who understand the systems beneath their work. Not engineers, not product managers, not prompt hobbyists, but something in between. At jhana, we have seen the birth of a new kind of professional, neither lawyer nor engineer, neither manager nor advisor, but someone who can interpret model behaviour, compare outputs, route work intelligently, and maintain stability even as the frontier moves.

Identifying as a technical personality is not about writing code. It is about recognising that modern legal work now rests on top of fast-moving platforms and choosing to stay aware, adaptive, and in control of those platforms rather than being shaped by them.

Why Model-Agnostic is the Only Way

Frontier AI is fast, unstable, and constantly shifting. Models rise and fall in weeks. Capabilities regress without warning. File-handling changes. Pricing shifts. APIs break. In this environment, single-model platforms cannot keep up. They impose limits, regress unpredictably, and turn every workflow into a gamble. No serious legal practice can anchor itself to one model in a landscape that refuses to stand still.

jhana was built for this environment. It brings multiple models into one system, chosen not only for the task at hand but also made available the moment they launch so lawyers can use them, compare them, and see the differences for themselves. It adds structured legal memory so the system recalls instead of guessing. It enforces document completeness so nothing is missed, whether it is a clause, an annexure, or an obligation running across an SHA. And it offers resilience by design, so workflows survive even when the ecosystem shifts overnight. It is a system built for lawyers, not a repurposed chat interface.