You came here from the email, so you have the rule: everyday model by default, heavy mode only when the job has real reasoning inside. This page is the part I held back: the names, the numbers, and the tactics.

Everything below is checked against the companies' official pages as of July 21 2026, then compared with what real users report. Labels: OFFICIAL is the company's own words. TESTED is what I personally tested.

🛠️ the map, by name

ChatGPT

Everyday: GPT-5.5 Instant. The default, and it should be. Email, summaries, rewrites, brainstorming

Heavy: GPT-5.6 Sol. The reasoning model on paid plans, at Medium or High (Extra High and Sol Pro live on the higher tiers). Multi-step logic, code, math, one long messy document.

I default to 5.6 for almost everything: "powerful, fast, half the price of Fable, and my default personal AI"

Claude

Current lineup: Haiku 4.5, Sonnet 5, Opus 4.8, Fable 5.

Everyday: Sonnet 5. Anthropic's own model guide calls Sonnet "the daily driver," says "if you're not sure which model to pick, start here," and notes Opus "uses more of your rate limit"

Bulk and simple: Haiku 4.5. Repetitive work across many runs: reformatting, classifying, renaming, cleanup.

Heavy: Opus 4.8. The jobs where a miss costs you hours.

Pay-as-you-go heavyweight: Fable 5. Anthropic’s most capable model is no longer included with your Pro or Max usage. Every run uses separately purchased credits, so save it for the hardest, long-running projects where another model falls short. For everyday work, start with Sonnet 5.

My routine framework with Claude: Haiku → batch operations, data transformation, classification, anything repetitive across many calls. Sonnet → daily coding, feature work, code review, 90% of tasks. Opus → architecture decisions, security review, complex debugging where missing something costs hours.

Gemini

Current lineup: Gemini 3.1 Flash-Lite (efficient everyday), Gemini 3.5 Flash (balanced), Gemini 3.1 Pro (heavy). All three are available across the plan levels; the plan mostly changes how much compute and context you get, not which models you can touch.

Everyday: Flash-Lite or Flash. Heavy: Pro (plus Deep Think, which needs Pro on the top plan).

Copilot

Microsoft 365 Personal and Family include 60 AI credits a month for image generation and editing across the Office apps and Designer. That's a creative-credit pool, not a cap on every Copilot chat.

Two things just changed, so watch for them. The standalone Copilot Pro at $20 is being retired (legacy plans end August 17). The current high-usage consumer offer is Microsoft 365 Premium at $19.99/month (or $199.99/year), and Microsoft calls it "extensive" usage, not unlimited.

And the AI benefits, including those 60 credits, go to the subscription owner only. They don't split across Family members. So don't count on per-person pools.

Grok

Current model is Grok 4.5 (the old "Grok 3, 50 queries every 2 hours" numbers are stale). xAI no longer publishes a fixed free-query count, and paid usage now runs on one shared weekly pool across Chat, Imagine, Voice, and Build.

The caution flag: xAI moves quotas more than anyone. When daily video limits became one weekly pool, a reader's public breakdown put the total capacity cut at 83.8%, and one subscriber's plan went from a promised 100 videos a day, to about 30, to 7. If your quota shrinks under you, the plan changed underneath you, and there was no announcement you missed.

Tiers, for reference: Free, SuperGrok Lite, SuperGrok ($30/month), SuperGrok Heavy. Only the $30 standard price is posted publicly; the Lite and Heavy prices aren't.

⚡ why long chats eat your limit

The mechanic is documented: an LLM has no memory between turns, so the whole conversation gets sent again as input every time you reply. The longer the chat, the more each new turn costs.

❝

Message 30 costs ~30x message 1, because Claude re-reads the whole thread every turn.

Most tools now cache or compact part of that history, so the real cost varies by product. But the direction holds: "one long chat with all my context" feels smart and bills like a leak.

Five fixes, all from tested reader reports:

1. Start a new chat when the topic drifts or the thread gets long. The single most repeated limit-saver across every platform I mined.

2. Use a handoff. Before you close the old chat, run this (the practice is TESTED across multiple threads, including a reader who packaged it as a "handoff report" skill; this template wording is mine, test it once before you rely on it):

Summarize this conversation for a fresh chat: the goal, the decisions
we made, the current state of the work, and what's left to do.
Under 150 words. No filler.

Before starting the new chat, rename the old one with a clear, specific title—or copy its existing title exactly. Then paste this prompt into the new chat:

Continue this work using the handoff below. If you need missing details, refer to my past conversation titled “[exact chat name].”

[Paste the handoff summary]

Yes! ChatGPT and Claude can now read your past conversations.

3. Edit your earlier message instead of stacking follow-up 12. Editing replaces the turn. Following up adds the whole thread to the bill. (TESTED)

4. Convert files to .md or .txt before uploading. Scanned PDFs and heavy files chew tokens on reading alone. "Convert the material to .md or txt first" was a repeated fix in the file-analysis threads. I use Markitdown to convert any file to Markdown file.

5. Do first-pass reading in a free tool, then bring conclusions over. NotebookLM came up again and again as the free summarize-first step, with the paid chat reserved for the thinking (TESTED).

6. Branch before the conversation takes a detour. Go back to the last useful message and start a branch there. You keep the context before that point without dragging every later correction, failed attempt, or side question into the new path. (TESTED)

🛠️ stepping down without getting burned

The escalation rule from the threads, and the one I'd start with (TESTED ): start with the least expensive model, and move up only when the answer is inadequate. Notice the direction. The menu nudges you to start at the top. The people who stopped running out start at the bottom.

The 2-day trust routine:

  1. Pick one task type (say, email drafts).

  2. For the first 2 or 3 outputs, check the small model's work yourself before using it.

  3. If it holds up, that task type belongs to the small model now. Stop checking it.

  4. Keep checking anything with numbers, names, or legal weight, on any model, forever. That part protects you in both directions.

You're building earned trust instead of guessed trust. It takes about 2 days of light attention, and it ends the background worry that the cheap model is ruining your work behind your back.

the skeptic corner

Two camps, both quoted from the same debate:

❝

The harness is doing most of the work now, the model's almost a swappable part at this point.

❝

The model is absolutely the product.

Where the evidence in these threads lands for everyday work: the first camp is right. On email, summaries, and rewrites, model choice barely moves the answer. It starts mattering in exactly 2 places: when you run out, and when the task needs real reasoning. Which is the email's rule, arrived at from the other direction.

📨 the 30-second version

Open your app's model picker. Set the everyday model as the default. Done.

Then watch one week. If you never hit a limit, you're finished, and you saved yourself an upgrade. If you still run out, the leak is your longest chat, and the handoff template above is the fix.

Tell me in the comments which app you run and where your limit dies. I read all of them.

Sources

Official (checked July 21 2026):

Keep Reading