AI Model Matrix Quickstart: A Five-Axis, Three-Filter Way to Pick a Model

The live Model Matrix at agenticbotsitter.com/websites/ai-model-matrix/model-matrix.html ships 1,146 models in its current snapshot. You do not need all of them. You need a way to drop 1,146 to one in under two minutes. This guide gives you that: five axes you should always look at, three filters that turn the axes into a shortlist, and one heuristic for the moment you say “good enough, ship it.” Then a worked example picks between Claude Sonnet 4.5, GPT-5.5 Pro, MiniMax M3, and Gemini 2.5 Pro.

The five axes

Every model card on the Matrix exposes the same five axes — think of them as the questions you would ask a senior engineer who has tried the model.

1. Cost — the blended per-million-token rate (input×0.25 + output×0.75, a 1:3 mix).
2. Capability — a 0–100 score plus a letter grade (A ≥ 85, B ≥ 70, C ≥ 50, D below) per category. There are 15: coding, agentic, captions, creative, science, technology, finance, legal, translation, reasoning, math, vision, longctx, speed, local. Pick the one your task actually is; do not average across them.
3. Context — the model’s published window, with chips at Any / 32K+ / 128K+ / 1M+. A 1M window on a model that loses coherence at 200K is still a 200K model in practice; check the longctx score separately.
4. Speed — tokens per second under load, or a heuristic when no measurement exists.
5. Latency — time to first token, folded into the speed score. Matters more than throughput for chat; reverse for batch.

Two of the five — speed and latency — almost always lose to cost and capability on a first pass. They become tie-breakers, not gates.

The three filters

The five axes are what you measure. The three filters are what you decide. Run them in this order.

Filter 1 — Use-case pick

Open the Rank by capability dropdown and pick the category that matches your job. If you are writing a legal brief, pick legal. If you are wiring a coding agent that reads a repo and edits files, pick agentic. Resist reasoning; that axis over-rewards math-bench models and under-rewards everything else.

This filter drops the table from 1,146 rows to roughly 60–120.

Filter 2 — Capability floor

Slide the Min performance knob until the Top Pick card stops showing models you would not trust. For most production work, set the floor at 70 (grade B or better). For agentic and coding work where a wrong answer is expensive, set it at 85 (grade A). For exploratory drafts where speed matters, 50 (grade C) is honest.

This filter usually drops the shortlist to under 20.

Filter 3 — Budget cap

Set the Max cost slider, or click a chip (≤ $0.50 / ≤ $1 / ≤ $5 / ≤ $20 / Any). The chips are the right tool for a first pass. If you have a real number — say, “this call must be under 3 cents per 1k tokens” — use the slider.

This filter drops the shortlist to under 5. That is your pick set.

The cheapest-viable heuristic

Once you have a pick set, run this three-step tie-breaker before you commit:

1. Sort by Value = score / blended_per_1m, not by score. The Matrix does this in the by-value view. A grade A model at $30 blended is often worse per dollar than a grade B model at $1 blended, and the difference is usually 50× to 100×.
2. Re-rank by your real workload. The blended rate assumes 1:3 input:output. For input-heavy calls (RAG, summarization), divide by input_per_1m; for output-heavy calls (code generation, agentic tools), divide by output_per_1m. Both numbers are exposed per row.
3. Refuse to pay for context you do not need. A 1M window on a 50K task is overhead. If your prompt fits in 32K, set the Min context chip to 32K+ and re-sort.

If two models are still tied, pick the one with the higher source_agreement and confidence. A heuristic_only score is a metadata guess, not a measurement — treat it as a placeholder.

Worked example: Sonnet 4.5 vs GPT-5.5 Pro vs MiniMax M3 vs Gemini 2.5 Pro

The four candidates, pulled from the live snapshot the Matrix was serving on 2026-07-14:

Model	Context	Blended $/1M	Input $/1M	Output $/1M
Claude Sonnet 4.5 (anthropic/claude-sonnet-4.5)	1,000,000	$12.000	$3.00	$15.00
GPT-5.5 Pro (openai/gpt-5.5-pro)	1,050,000	$142.500	$30.00	$180.00
MiniMax M3 (minimax/minimax-m3)	1,048,576	$0.975	$0.30	$1.20
Gemini 2.5 Pro (google/gemini-2.5-pro)	1,048,576	$7.812	$1.25	$10.00

Capability grades across seven categories that matter for most teams:

Model	coding	agentic	reasoning	math	longctx	speed	vision
Sonnet 4.5	C(53)	B(78)	C(63)	A(92)	A(100)	D(9)	A(100)
GPT-5.5 Pro	A(97)	A(98)	A(93)	A(92)	A(100)	D(9)	A(100)
MiniMax M3	A(92)	A(85)	A(98)	A(96)	A(100)	D(19)	C(65)
Gemini 2.5 Pro	D(35)	D(29)	C(50)	D(47)	A(100)	D(28)	?(0)

Scenario A — “I am building a coding agent that edits a real repo”

• Filter 1 (use-case): agentic.
• Filter 2 (floor): 85 (grade A). A wrong file edit costs a rollback.
• Filter 3 (budget): ≤ $20. Anything cheaper is welcome.

Pick set: GPT-5.5 Pro (A, $142.50), MiniMax M3 (A, $0.98). Sonnet 4.5 is B, out. Gemini 2.5 Pro is D, out.

Tie-break: GPT-5.5 Pro scores 98 vs M3’s 85 — a 15-point gap for 145× the price. At agentic volume (millions of tool calls per week), 15 points is real but not 145× real. Pick MiniMax M3. Re-rank by output_per_1m (agents emit more than they ingest): $1.20 vs $180.

Scenario B — “I am writing a long document with an attached PDF”

• Filter 1: longctx if the PDF is over 200K tokens; creative otherwise.
• Filter 2: 70.
• Filter 3: ≤ $20.

Pick set: Sonnet 4.5 ($12, A/longctx, B/creative), Gemini 2.5 Pro ($7.81, A/longctx, C/creative), MiniMax M3 ($0.98, A/longctx, B/creative).

Scenario C — “I need the single best answer regardless of cost”

• Filter 1: reasoning.
• Filter 2: 90.
• Filter 3: Any.

Pick set: MiniMax M3 at score 98, GPT-5.5 Pro at score 93. Tie-break does not apply — you opted out of budget. Pick MiniMax M3. It scores higher than GPT-5.5 Pro on reasoning in this snapshot, and it is cheaper. M3 sits at #9 on the reasoning by-score leaderboard at score 98; the top eight are closed frontier and most cost more.

Scenario D — “I just need a chatbot that does not embarrass me”

• Filter 1: agentic. Filter 2: 70. Filter 3: ≤ $1.

Pick set: MiniMax M3 only. Sonnet 4.5 is $12, Gemini 2.5 Pro is $7.81, GPT-5.5 Pro is $142.50. Pick MiniMax M3. That is what “cheapest viable” looks like in production: the cheapest model that clears your floor, full stop. The Matrix’s Best Value card is computed the same way — top of the by-value list after your filters are set.

When the matrix lies

Three caveats worth memorizing so you do not get burned:

• longctx and speed are mostly heuristic-only in the current build. The contributing_sources field shows metadata_heuristic for both on every frontier model. A score of 100 on longctx means “the model advertises a big window,” not “the model has been measured to use it well.” Trust the n_measured_sources and source_agreement counts more than the headline number.
• Letter grades are taken as-is from the snapshot. A grade C on Sonnet 4.5 coding is not “Sonnet 4.5 cannot code.” It is “Sonnet 4.5 has fewer measured coding benchmarks in this snapshot than the A-grade models.” Run your own eval on 20 of your real prompts before you downgrade a model on this evidence alone.
• Free and local models are filtered out by default toggles. Click Free models only or Local models only if those matter to you. Ollama-style local models pull disk, runtime, and recommended-system-RAM into the same table; a 70B model that needs 96 GB of RAM is not the same value as a 70B model that needs 48 GB.

The two-minute recipe

1. Open the Matrix.
2. Pick the use-case from the dropdown.
3. Set the capability floor.
4. Set the budget cap.
5. Read the Top Pick and Best Value cards.
6. If they disagree, trust Best Value unless your floor is near 100.
7. If they agree, ship it.

That is the whole decision: 1,146 models, seven steps, one pick.