My Model Rankings and Favorites, From Gauntlet Receipts

This is where I pick favorites. Every pick comes from a gauntlet receipt. A run log, a pass check, a cost line.

No benchmark scores in this post. If you want the numbers, the full board is at wikiwayne.com/wikibench.

How my 7 game gauntlet works

Same 7 game prompts for every model. NES Racer, City Chomp, Hogger Boss Fight, Pirate Ship, Brick NYC, Silly String Madness and Vegas Snowman.

One shot each. No edits to the output.

Pass means the game loaded and ran. Finish was clean, the HTML was complete, the scripts parsed, no page errors on load, and the first frame wasn't blank. Pass does not mean the game is good. I say that every time.

Then I record the games on my Mac mini so you can see them. The machines are in The AI Hardware I Own: ASUS Ascent GX10, a Mac mini and a MacBook Pro.

My favorite for cheap, clean first tries: Muse Spark 1.3

Meta Muse Spark 1.3 (meta/muse-spark-1.3) is my favorite on this gauntlet.

In my Oct 12 run it went 7 of 7 on the first try. It was the only model in that round to do it. It also used the fewest input tokens and cost the least, $0.44 for all seven.

Same round, for context:

• Muse Wonder (meta/muse-wonder): 6 of 7 first try, 7 of 7 with retries, $1.93.
• Grok 4.7 (x-ai/grok-4.7): 6 of 7 first try, 7 of 7 with retries, $3.27.
• Claude Opus 5.5 (anthropic/claude-opus-5.5): 4 of 7 first try, 5 of 7 with retries, $6.14. In that round it never passed City Chomp or Silly String Madness. City Chomp hit the provider's content filter on all 4 tries. Later I reworded the City Chomp and Silly String prompts to drop the harm wording, and Opus passed both on a softened rerun. That got it to 7 of 7, but not in one clean pass.

GPT-6.1 Terra matched it

GPT-6.1 Terra ran the same 7 prompts through Nous Portal. 7 of 7 on the first try, no retries. 41 cents total, about 6.8 minutes wall clock, three at a time (about 15 minutes of generation summed).

Per game it ran about 5 to 7.5 cents. Cheapest was Hogger Boss Fight. Priciest was Brick NYC.

That's a hair under Muse Spark 1.3 on cost. Both are clean sweeps on the first try.

Gemini 4 Argon: clean sweep at a higher price

Gemini 4 Argon (google/gemini-4-argon) also went 7 of 7 on the first try through Nous Portal. No retries needed. $1.3359 total.

Three clean sweeps. Different prices. If cost matters to you, that spread is the story.

Earlier runs for context

• Mistral Large 4: 7 of 7.
• MISSINGNO (a stealth model): 4 of 7 on the first try. Its 7 of 7 is its top result across several tries, not one clean pass. Two of its first try fails ran out of tokens while still reasoning.
• Step 5 Preview: 3 of 7 first try, 7 of 7 with retries.

The bead tracker: a harder test

My bead tracker gauntlet has two parts. Part 1, build a tracking video. Part 2, watch a video and say the exact frame where the bead vanishes.

Grok 4.7 is the only model that passed both parts. Part 2 named the exact frame, 1665. Part 1 passed on its second try through a Cursor cloud agent. One caveat I keep attached. It saw every 15th frame, so its frame numbers were all multiples of 15.

Opus 5.5 passed Part 1 and missed Part 2 by 2 frames. Close, still a fail.

Muse Spark 1.3, Muse Wonder and Muse Spark Open passed Part 1 and failed Part 2. Nemotron 3.5 Super VL failed both.

So on video reasoning, my pick is Grok 4.7. With the caveat.

My favorites, short version

• Cheap clean sweep on the game gauntlet: Muse Spark 1.3 and GPT-6.1 Terra.
• Clean sweep at a higher price: Gemini 4 Argon.
• The bead tracker: Grok 4.7.

These are favorites from my runs, not a universal ranking. Different prompts, different results.

How I keep this honest

My Grok Bot team runs the gauntlets and checks every number against a file. I post the results myself. That flow is in My Grok Bot Team Runs My Content, Site and Video Work.

The hosted runs go through Nous Portal and other routes. The Hermes side is in Hermes Agent on My GX10, Over Telegram. The local side is in The Local Models I Actually Run on My GX10.

Read the rest of the stack

• The AI Hardware I Own: ASUS Ascent GX10, a Mac mini and a MacBook Pro
• Hermes Agent on My GX10, Over Telegram
• My Grok Bot Team Runs My Content, Site and Video Work
• The Local Models I Actually Run on My GX10