Pick one model per job, test them side by side on your own tasks, and keep a six-slot stack that fits how you work.
You buildYour model stack
A one-page stack file with six slots (main chat, cheap and fast, open or local, image, video, coding agent), each filled from your own blind test, not from a headline.
Run one task through two or three models and judge the results blind
Fill six stack slots and write down why each model earned its place
Read a comparison table with its check date in mind
A map, not a ranking
Dozens of AI products compete for the same job titles. You do not need all of them. You need a short stack: one tool you open first, and a few specialists you call when the task demands it.
ProductCurrent modelStrong atPrice
ClaudeAnthropicFable 5.2 · Opus 5.5Long agent sessions, deep research, finished documents and code$8 / $40 · Opus $4 / $20 ↗
ChatGPTOpenAIGPT-6 AstraComplex reasoning and coding; the widest consumer app$10 / $50 ↗
GeminiGoogleGemini 4 CarbonFirst on the LMArena text board on 11 Nov 2026 (preliminary)$2 / $10 to 30 Nov, then $4 / $20 ↗
GrokxAIGrok 4.7Long prompts: up to 500k tokens of context$2 / $6 ↗
DeepSeekDeepSeek · Open weights: yesV4-ProLow price, 1M-token context, weights on Hugging Face$0.66 / $1.98 off-peak ↗
MistralMistral AI · Open weights: yesMistral Large 4Open-weight multimodal model from a European labLe Chat Pro $14.99 / month ↗
Meta AIMetaMuse Spark 1.1Muse replaced the Llama line; Muse Wonder ranks third on LMArena text— ↗
Checked on 2026-11-16. Drag to tilt
Image and video
ProductCurrent modelStrong atPrice
GPT ImageOpenAIGPT-Image-2.6Detailed compositions, precise edits, SVG and animated GIF output$30 per 1M output image tokens ↗
Nano BananaGoogleNano Banana 2.1Cheap fast images inside Gemini and AI Studio$0.034 per 1K image ↗
MidjourneyMidjourneyV8.2Aesthetics, image quality, personal stylefrom $10 / month ↗
FLUXBlack Forest LabsFLUX 3 ImagePhotoreal images by API, a light Klein model for volume$0.048 per 1K image ↗
RunwayRunwayGen-4.5A studio in one place: hosts Kling and Seedance toofrom $15 / month ↗
LumaLuma AIRay3.2HDR and EXR output, video-to-videofrom $30 / month ↗
SoraOpenAI—OpenAI's model list no longer shows a video model— ↗
Checked on 2026-11-16. Drag to tilt
ProductCurrent modelStrong atPrice
Claude CodeAnthropicClaude CodeTerminal, IDE, desktop and web; long multi-step changesin Claude Pro, $20 / month ↗
CodexOpenAICodex CLI · cloudIn every ChatGPT plan; light use on the free planfrom $8 / month (Go) ↗
AntigravityGoogleAntigravity CLIReplaced Gemini CLI for free users on 18 Jun 2026Individuals plan: free ↗
GitHub CopilotGitHubCopilot coding agentTakes an issue and opens a pull requestPro $10 / month ↗
CursorAnysphereCursorAn editor built around the agentfree plan; paid on cursor.com ↗
Checked on 2026-11-16. Drag to tilt
The method: choose by task
Write down the job in one sentence, then give the same job, with the same prompt, to two or three models. Hide which answer came from which. Pick the one you would send to a client, then reveal the names.
Asking which model to use
Weak
Which AI model is the best?
Strong
I write {{task_type}} for {{audience}} about {{weekly_volume}} times a week. I care about {{quality_priority}} and I never paste {{forbidden_data}}. From the models in my table, suggest two or three to test side by side, say what each should be good at for this job, and give me a pass or fail checklist for judging the outputs.
Why it works: The weak question has no answer without a task. The strong one names the job, the limits and the judging rule, so the reply is a test plan.
Blind side-by-side test
Run this exact prompt in {{model_count}} different models, each in a fresh chat with my passport pasted first.
TASK
{{task_text}}
INPUT
{{task_input}}
RULES
- Answer the task only. No preamble, no closing offer.
- Do not name yourself or your vendor anywhere.
- If you lack information, say what is missing instead of guessing.
I will paste the answers back to a judge chat labelled A, B and C in random order.
A fresh chat and the same passport remove the two biggest hidden differences between models: leftover context and your own setup.
Blind judge
You are a strict judge. Below are {{model_count}} answers to the same task, labelled A, B, C. You do not know which model wrote which.
TASK
{{task_text}}
MY CRITERIA
{{criteria_list}}
ANSWERS
A: {{answer_a}}
B: {{answer_b}}
C: {{answer_c}}
For each answer give a pass or fail on every criterion with one short reason. Flag any invented fact, quote or link. Then rank the answers and say in two sentences what would make the weakest one acceptable.
Criteria written before you see the answers keep you from rewarding whichever one sounds most confident.
Route my task to the right model
Here is my current model table:
{{my_model_table}}
Here is a list of tasks I do in a normal week:
{{task_list}}
For each task, name the model slot it belongs to: main chat, cheap and fast, open or local, image, video or coding agent. Give one sentence of reasoning based only on the table. Mark any task where the table does not give enough information, and tell me what to test. Do not invent features that are not in the table.
Feeding the table back in keeps the answer tied to facts you checked, not to whatever the chat remembers.
From tasks to stack
My tasksBlind testSix slotsStack file
Assemble the stack
List ten real tasks
Take them from your last two weeks: emails, summaries, a picture, a script. Real work beats invented examples.
Pick three tasks that differ
One writing task, one reasoning task, one with a strict format. Different shapes reveal different strengths.
Blind test two or three models per task
Use the side-by-side prompt, then the judge prompt. Write your criteria first.
Fill the six slots
Main chat model, cheap and fast model, open or local model, image tool, video tool, coding agent. A slot can stay empty if you have no such task yet.
Add a review date
Put a date four to six weeks ahead in the file. That is when you rerun one test.