Annie Tsai

I’ve learned over the past two years that when you dive deep into building agentic systems to solve real business problems, the LLM model(s) you choose can make or break how successful that agentic system is in achieving its goals relative to the price it costs to perform that work.

The model you chose shapes everything downstream, from the quality of your outputs and the cost of running them, to the ceiling on what you can build next and, most critically, the depth of context your AI accumulates about you, your business, and how you think. This context created over hundreds of conversations is one of the most valuable assets in your toolkit, and every time you switch tools chasing the newest release, you reset it to zero.

This week, I want to walk through how to think about choosing an LLM as a business decision.

Context Is the Moat, Not the Model

Harvard Business Review published an article just this week titled “When Every Company Can Use the Same AI Models, Context Becomes a Competitive Advantage.” Read On –> The thesis: when everyone has access to the same foundational models, the differentiator is no longer which model you picked. It’s how well you’ve taught it to understand your business — your data, your workflows, your decision-making patterns, and your institutional knowledge.

The agentic systems I’ve built at Interact became powerful because I invested the time to refine their contexts in a systematic and step-functioned manner — our customer lifecycle, our pricing logic, the institutional knowledge locked in raw data, our escalation patterns, our edge cases. The model choice does matter from the standpoint of cost-to-performance (which we will discuss further down), but the context I built around it mattered a lot more. That context compounds, and it compounds faster when you commit to a primary tool and go deep instead of jumping between platforms every time a benchmark leaderboard shuffles.

Your first LLM decision matters more than you think. You can always switch later, but the depth of relationship you build with your primary tool is where the real returns show up.

Phase 1: Pick Your Primary and Go Deep

If you’ve been jumping between ChatGPT, Claude, Gemini, and Grok every few weeks based on whatever dropped on X that morning, I’m going to give you permission to stop.

Ethan Mollick has been saying this consistently throughout 2025. In both his June and October 2025 guides on his Substack One Useful Thing Read On –>, he makes the case that for most people who want to use AI seriously, the question has shifted from “which model is best” to “which system is best for how you work.” His recommendation: pick one of three platforms (ChatGPT, Claude, or Gemini), pay for the $20/month tier, and use it seriously.

In a November interview with Insight Partners Read On –>, he put it even more directly: stop overanalyzing model choices, pick one advanced model, apply it to real work, and learn from the results.

I agree. Here are the considerations I’d think about for choosing that first primary tool:

What’s your primary daily use case?

If you spend most of your time writing, communicating, and synthesizing, Claude has consistently been the strongest at producing natural, high-quality written output and extended thinking-partner work. If you’re analyzing data, building code, or working across multiple programming languages, ChatGPT and Claude both perform well, with Claude currently leading on real-world software engineering benchmarks and ChatGPT offering broader IDE integrations through Copilot and Cursor. If deep research is your primary mode — pulling together information from multiple sources into comprehensive analysis — all three major platforms now offer deep research features, with Gemini having a slight edge given its native access to Google Search. If you’re heavy on X/Twitter and want real-time social data baked into your workflows, Grok has a native advantage there, plus strong math and reasoning capabilities through its Think mode. Your daily driver should match where you spend the most hours.

Where does your work already live?

If your organization runs on Google Workspace, Gemini has native integration that reduces friction. If you’re in the Microsoft ecosystem, Copilot and OpenAI’s tools are already in the furniture. If you need flexibility across environments or you’re building things, Claude gives you a strong general-purpose partner with deep reasoning and writing capability. If your business relies heavily on X for distribution, community, or real-time market intelligence, Grok’s native platform integration gives you a direct line into that data. Ecosystem fit matters because friction kills adoption.

Context window and memory

Can the model hold enough of your work — your documents, your conversation history, your project context — to be genuinely useful across a session or a multi-day project? This is one of the most underrated differentiators in practice and one of the least discussed in reviews. As of early 2026, both Gemini and Claude now offer 1 million token context windows — roughly 750,000 words or 1,500 pages of text in a single session. Claude’s newest flagship, Opus 4.6, launched February 5th with 1M tokens in beta, and its Sonnet models have supported 1M since mid-2025. The critical difference isn’t just window size, though — it’s how well the model actually retains and uses information across that full context. On the MRCR v2 benchmark, which tests whether models can find specific information buried in massive documents, Opus 4.6 scored 76% at 1 million tokens while earlier models scored below 20% on the same test. ChatGPT’s GPT-5.2 supports up to 400K tokens, and Grok’s latest models offer up to 1M as well, with xAI’s Grok 4.1 fast pushing to 2M tokens for extended agentic workflows. If your work involves synthesizing across large bodies of information, this matters more than almost any benchmark score.

Privacy and data handling

Where is your data going, and is it being used to train the model? For any business use, this can’t be an afterthought — especially if you’re feeding it customer data, financial information, or proprietary strategy. Read the terms and understand what you’re opting into. Anthropic’s Claude does not use your conversations to train models by default, and their enterprise tier adds additional data isolation. OpenAI’s ChatGPT Team and Enterprise plans offer similar protections, but the free and Plus tiers do use your data for training unless you opt out. Google’s Gemini for Workspace includes enterprise-grade data governance for organizations already on Google Cloud. If your compliance requirements are strict enough that data cannot leave your infrastructure at all, open-source models like Meta’s Llama that you can self-host are worth exploring. I have my own instance of Llama training on a NVIDIA DGX Spark in my home office as we speak. Grok deserves a specific note here: it’s tightly integrated with the X platform, and xAI’s data handling policies are less transparent than those of Anthropic, OpenAI, or Google. This is something worth scrutinizing if you’re feeding it anything proprietary.

How does it handle being wrong?

Some models hallucinate confidently and never flag uncertainty, while others hedge constantly. Depending on whether you’re using AI for customer-facing work, internal analysis, or creative ideation, that behavioral difference changes your risk profile significantly. In my experience, Claude tends to acknowledge the limits of what it knows and will tell you when it’s uncertain, which is valuable for high-stakes or compliance-sensitive work. ChatGPT can be more confident in its outputs, which works well for brainstorming and ideation where momentum matters more than precision. Gemini’s grounding with Google Search gives it an advantage when factual accuracy and recency are the priority. Grok leans confident and fast, which suits brainstorming and real-time analysis, but its “unfiltered” branding means you’ll want to double-check outputs before using them in anything customer-facing or compliance-sensitive.

Does the interaction feel right?

I don’t see this on enough evaluation lists. The interface, the rhythm of conversation, the tone, and the way it responds to pushback do make a difference in both adoption and outcomes of that adoption. You’re going to spend a lot of time here, and if the interaction feels clunky or the tone is off, you won’t use it enough to get the compounding benefits. Claude’s conversational style tends to feel the most natural for extended writing and thinking-partner work. ChatGPT has the most versatile interface with voice mode, image generation, and a broad plugin ecosystem. Gemini feels most seamless if you already live in Google’s products. Grok has a distinct personality which some people love and others find distracting for serious work. Spend a week with each before committing — the one that matches your working rhythm is the one you’ll actually use.

The goal here is to find your best match — and then go deep enough that the tool starts working with you instead of just for you.

Phase 2: As Business Problems Get Specific, Expand Deliberately

Once you’ve built real fluency with your primary model — once you understand its strengths, its blind spots, and how it handles the kinds of problems you throw at it — that’s when the second question emerges: is this the right model for this specific workflow?

This is the maturity curve, and you are on it right now. Back in June, Andreessen Horowitz surveyed 100 CIOs across 15 industries in mid-2025 Read On –> and found that 37 percent are now using five or more models in production, up from 29 percent the year before. The primary driver was that model differentiation by use case has become more pronounced, and different models genuinely perform differently depending on the task. Imagine what those numbers are now.

Their findings echo what I’ve seen in practice: one model might be stronger for fine-grained code completion while another is better at system-level architecture. One might be stronger at writing and content generation while another handles complex question-answering more effectively. These differences are meaningful, and they only become visible once you’ve spent enough time with one model to recognize the gaps.

The questions to ask in Phase 2 are different from Phase 1:

Match the reasoning depth to the task

Not every business problem requires your most powerful model. A customer support triage system needs speed and consistency, a strategic analysis needs deep reasoning, and a data enrichment pipeline needs something lightweight that can run thousands of calls cheaply. Using a frontier model for everything is like assigning your most senior person to every task in the company, and it is both expensive and a waste of their strengths.

Factor in latency

For anything customer-facing or real-time, response speed matters. For batch processing, back-office analysis, or overnight data runs, you can trade speed for depth. The best model for the job is often the one that meets the speed and quality requirements at the right price, which may or may not be the most powerful option available.

Understand integration requirements

Does this workflow need structured outputs? Tool calling? Multi-modal input? API reliability at scale? These technical requirements narrow your options quickly and often matter more than which model scored highest on a benchmark.

Use models against each other

This is something I do constantly when building agentic systems, and I think it’s one of the most underutilized strategies available right now. You can use one model to generate output and a second model to evaluate, challenge, or refine it — creating a feedback loop that produces significantly more sophisticated and reliable results than any single model working alone. Think of it as building a team where each member has a different perspective: one drafts, one pressure-tests, one synthesizes. The cost of running two mid-tier models in this kind of adversarial or collaborative loop is often less than running one frontier model, and the output quality can be materially better.

Name multi-model orchestration as the horizon

For those of you further along in the journey, the endgame is routing different tasks to different models automatically. You don’t need to build this today, but knowing that’s the direction helps you make architecture decisions now that won’t box you in later.

The ROI of Right-Sizing: Price to Return

This is the part that matters most at scale and IMO gets the least airtime in model comparison conversations.

The pricing spread across LLMs is enormous, and most people have no idea. At the API level, the cost per million tokens ranges from fractions of a penny for lightweight models to $15 or more for the most powerful frontier models. At the consumer subscription level, you’ll pay roughly $20/month for ChatGPT, Claude, and Gemini’s standard paid tiers, while Grok requires X Premium+ at $40/month or a separate SuperGrok subscription. But what you get for that $20 (and what it costs when you scale beyond personal use) varies dramatically.

Right-size the model to the task. If 70 percent of your workflows can be handled by a mid-tier model and you only need the frontier model for the other 30 percent, build model specification per sub-task so that the mix delivers better ROI than running everything through the top tier. This is the same operational discipline you likely already apply to headcount planning, vendor management, and infrastructure decisions.

If you want a single resource to bookmark, artificialanalysis.ai is a clean independent dashboard for comparing models across intelligence, speed, latency, and price. It’s independent data with no vendor affiliation. Use it to see where your current model sits relative to the alternatives and whether you’re overpaying for capability you don’t need or underinvesting in capability you do.

The operators who get this right will be the ones who treat model selection the same way they treat any other resource allocation decision: what’s the outcome, what’s the cost, and what’s the return?

Your Starting Audit

Here’s your practical starting point:

Pick three workflows you touch every week. For each one, write down the input, the desired output, the acceptable speed, and how much it matters if the AI gets it wrong.

Then ask yourself two questions:

1. Is my current model actually the right fit for these workflows, or have I just been defaulting?
2. Am I getting compounding returns from the context I’ve built, or am I resetting that context every time I chase the latest release?

If the answer to the first question is “I’ve been defaulting” — go back to Phase 1 and be intentional about your primary.

If the answer to the second is “I keep switching” — stop this practice and go deep with one or two models for their relative areas of best fit. The returns will show up faster than you expect.