🧠 One Model for Everything Is Why Your AI Outputs Are Mediocre

Before the breakdown, here is our actual stack. No caveats, no ‘it depends.’

For writing anything that matters — newsletters, client reports, proposals — we use Claude. Not because it wins every benchmark, but because when we read the output back, it sounds like a person wrote it. That is harder to achieve than it looks, and it’s the thing that matters most when a real human is going to read your work.

For research on anything time-sensitive, we use Gemini’s Deep Research mode. It runs dozens of searches, synthesises the results, and cites its sources. A research task that would take 45 minutes of manual reading comes back in 15 minutes, fully cited. The free tier handles most of it. (We also use Perplexity but Gemini is better for the work we do)

For images alongside text — social posts, visual mockups, anything where we want a layout fast — we use ChatGPT. The image generation is the only reason we keep a Plus subscription, and it earns it.

We don’t pay for Grok or Copilot as standalone subscriptions. Grok comes bundled with X Premium+, which some of us have for other reasons. Copilot is bundled into Microsoft 365.

The two-model minimum we’d recommend to anyone doing serious knowledge work: Claude + Gemini. Together they cover 80% of what most people need. Add ChatGPT if you produce visual content. Add Grok if you monitor brands or track social trends professionally.

We ran the same writing prompt through Claude Sonnet 4.6, ChatGPT GPT-5.5, Gemini Pro, Grok, and Microsoft Copilot to see what actually comes out. The prompt was deliberately specific: it required a clear voice, precise constraints, and no filler.

Here it is:

Claude tends to open with a concrete observation that pulls you in immediately. The constraint about "sounding like a smart friend" is treated as the primary instruction, not a secondary note. Voice and specificity show up in the first sentence. When you read it aloud, you don't have to make adjustments.

ChatGPT produces clean, correct writing. The constraint is followed technically but the result often reads slightly more like a polished article than a conversation. It satisfies the brief without fully inhabiting it. Good for general use; less good for writing that needs a specific person's voice.

Grok is fast and can be punchy. When the constraint is factual (list, summary, data) it performs well. When the constraint is voice-based — as this one is — it more often defaults to a generic opener that technically avoids the stated prohibitions but doesn't nail the tone.

Gemini is accurate and well-structured. It tends toward slightly more formal register even when asked not to. The output is useful as a first draft but often needs one editing pass to bring it down to the conversational level the prompt asked for.

Copilot produces competent, safe writing that reads like internal comms. It's unlikely to embarrass you; it's also unlikely to make anyone lean forward.

The key insight: The gap between models is subtle when you read one output in isolation. It becomes obvious when you read all five back to back. "Writing quality" isn't a benchmark number — it's the paragraph you read and think: that's exactly how I'd say it.

Claude's output has a characteristic quality that's difficult to articulate but immediately recognizable: it doesn't pad. Most AI writing has filler — transitional phrases that exist to reach a word count, hedges that exist to cover all bases, openers that warm up before saying anything. Claude tends to cut to the thing more quickly and hold a more consistent tone across a long piece.

This matters more than it sounds. If you're writing something a client, manager, or audience will read, the margin between "fine" and "good" is felt even by people who can't name what's different.

This is how we tested this:

Claude doesn't have real-time web access in its base form (though it can use tools). For anything requiring current information — recent news, live prices, what happened this week — use Gemini or Grok first, then bring the research to Claude for the writing pass.

For image generation, Claude is not the right tool. Use ChatGPT.

For tasks where you need to paste in a very long document (100k+ tokens) and ask detailed questions about specific sections, Claude handles this well but Gemini's context window may give you more headroom on extremely long inputs.

GPT-5.4 scores 92.8% on GPQA (graduate-level factual accuracy) and 97% on SimpleQA (direct factual questions).

In practical terms, it is the most reliably accurate model for factual recall of the major group. If you ask it a question with a known, verifiable answer, it is less likely to hallucinate a confident-sounding wrong answer than most of its competitors. It also leads on LongBench v2 with a score of 95, which measures performance over very long documents. For research synthesis or legal review where you’re feeding in long source material, that’s worth knowing.

The prompt style that works best with ChatGPT

ChatGPT handles open-ended, exploratory prompts more flexibly than Claude. It’s more forgiving with loose instructions and will usually produce something useful even from an underspecified brief. For image generation specifically, the key is to describe the style, mood, and layout explicitly — vague image prompts produce generic stock-photo results.

This is how we tested this:

For detailed multi-part prompts that require strict instruction-following across a long conversation, GPT-5.4 can drift from earlier constraints. Teams doing precision work — legal drafting, compliance review, structured documentation — often notice this more than casual users. For that kind of work, Claude’s instruction-following tends to be tighter. For live social data or breaking news, Grok and Gemini are better choices.

Gemini 3.1 Pro leads with an overall composite at 93. What that score doesn’t tell you is why: Gemini’s strength is breadth, speed, and live data access — not necessarily depth of reasoning on complex problems. It returns answers fast, handles a standard 1 million-token context window, and its multilingual capability is the strongest of the group across 100+ languages.

But the real differentiator is Deep Research mode. This feature runs hundreds of searches across the web, synthesises the results, and produces a fully cited report. A topic that would take a skilled researcher 3–4 hours of reading and note-taking can come back in 15–20 minutes with sources you can verify. No other consumer facing model does this at the same depth.

For Deep Research, be explicit about time frame and source quality. Gemini will search broadly by default — if you want recent developments, say so. If you want academic or primary sources rather than blog posts, say so. The tighter the scope, the more useful the citations.

SAMPLE PROMPT: Deep Research report (enable the toggle first)

We tested this out for ourselves and Gemini of course generated a detailed report which included the latest information, detailed breakdowns, impact numbers and tables, and citations.

For final written documents, Gemini’s output is accurate but functional — it reads more like a report than a piece of writing. If the document needs to sound polished and human, take Gemini’s research output and pass it to Claude for the final draft. That two-step workflow (Gemini researches, Claude writes) is more powerful than either model used alone, and it’s how we produce most of the research-heavy content in this newsletter.

Grok 4.20’s defining specification isn’t a benchmark score — it’s the data feed. Around 68 million English-language X posts per day flow into its context in real time. On any topic where recency and social sentiment matter, Grok is answering from data that is hours old, not months old. That is a structural advantage no amount of general benchmark improvement can replicate. On pure generation speed, Grok 4.20 runs at around 235 tokens per second — roughly three times faster than GPT-5.4. For high-throughput use cases, that matters. For typical daily work, the live data feed is the more relevant advantage.

Be explicit about wanting real-time data. Ask for what people are saying ‘right now’ or ‘in the last 48 hours’. Without that framing, Grok may draw on general training knowledge rather than its live feed — which defeats the entire purpose.

SAMPLE PROMPT: Deep Research report (enable the toggle first)

When we tested this, we were aware that ChatGPT Codex released its Pets feature on 2nd May 2026 and this was Grok’s response when we queried on 4th May 2026. The results are accurate:

Grok’s real-time advantage is specific to the X ecosystem. For live web data more broadly, Gemini’s Deep Research is more comprehensive. For writing quality and complex reasoning, Claude and ChatGPT are noticeably stronger. Grok is a specialist — use it when current social sentiment or X-specific trends are the actual thing you need, not as a general-purpose substitute.

Copilot’s clearest win is Teams meeting summaries. It can join a meeting, summarize what was discussed, list action items, and give a catch-up to anyone who joined late — without leaving the Teams interface. For organisations in relentless back-to-back meetings, that is genuinely useful and saves real time.

Outside that specific use case, Copilot’s writing quality doesn’t match Claude, its reasoning doesn’t match ChatGPT, and its research capability doesn’t match Gemini. The real combined cost — $30/user/month for Copilot on top of an existing Microsoft 365 subscription — runs $50–80+ per user per month before implementation overhead.

One development worth knowing: as of Copilot Wave 3 (March 2026), Microsoft added multi-model selection to Copilot Studio, including access to Claude Opus 4.6 for demanding reasoning and writing tasks.

If your organisation is already running Copilot at enterprise scale, you may be able to route your most demanding work to Claude without leaving the Microsoft ecosystem.

Keep this open the first two weeks. The habit forms faster than you'd expect.

Use when: You need to produce a well-researched, well-written document on a topic you don't know deeply — competitor landscape, market analysis, policy brief, industry overview.

Step 1 — Gemini Deep Research

Step 2 — Claude: Turn research into a document

Step 3 (optional) — Claude: Final edit pass