TIS-100 AI Benchmark Results: Opus 5.5 and GPT-6 Astra Solve All 55 Puzzles

Wincon Research has released TIS-100 Bench, a test where 16 AI models play the classic Zachtronics programming puzzle game TIS-100. Only Claude Opus 5.5 and GPT-6 Astra solved all 55 puzzles.

The bigger story is cost. GPT-6.1 Sol got close to the leaders for a fraction of the price, and models even beat a few long-standing human records.

Below I explain how the benchmark works, show the full results table, and cover what it means if you pick models for coding work.

What TIS-100 Is and Why It Makes a Good Test

TIS-100 is a 2015 puzzle game by Zachtronics, designed by Zach Barth. You write a made-up assembly language for a broken 1970s-style computer called the “Tessellated Intelligence System”.

How the game works

• Grid of nodes: the machine has 12 nodes in a 4 by 3 grid. Each one is a tiny compute node or a stack memory module.
• Tiny programs: each compute node holds at most 15 instructions.
• Two registers: ACC does most of the work, and BAK is a backup you can only swap or save into.
• Message passing: nodes send values to their neighbours, so you must plan how data flows across the grid.

Three scores, not one

Every working solution is scored on cycles (speed), nodes (how many you used) and instructions (code size). Players compete on each one, and a fast program is often big, while a small one is often slow.

That makes TIS-100 a great AI test. Getting a working answer is one skill. Squeezing it close to the best human score is a much harder optimisation skill.

How the Benchmark Works

Wincon Research published the results on November 3, 2026 (updated November 8). The snapshot is called “2026-11-03-v5”. The community leaderboard used for comparison was frozen on October 15, 2026.

The puzzle set

How models played

• Text tools only: no mouse and no screenshots. Models read the manual and test inputs, edited programs and ran them.
• Diagnostics: they could see mismatches, register values and blocked ports.
• Saved progress: a get_best tool returned their best passing programs.
• No early exit: models saw a clock, got reminders to keep improving and could not end an episode early.
• No cheating: they never saw community code or exact records, only rough score histograms.

One shot per puzzle

Each model got one scored episode per puzzle (pass@1), with no “best of several tries”. An unsolved puzzle counts as zero. Across all 880 episodes, models produced 95.5 million output tokens and made 62,400 game tool calls.

What “record proximity” means

For each score, Wincon divides the community record by the model’s best result. They average across cycles, nodes and instructions for all puzzles, then multiply by 100.

So 100 means matching the human record. A program that needs twice the record’s cycles counts as 50 for that score.

How results were checked

A new point only counts if it passes a fresh game replay and 100,000 random tests on the independent TIS-100-CXX simulator. Ties with human solutions do not count. Models found 154 candidates, and 133 survived deduplication.

Full Results Table (All 16 Models)

This is the leaderboard as Wincon ordered it, by the number of puzzles where the model found a new frontier point. Costs are listed only where Wincon stated them in the article text.

Opus 5.5 vs GPT-6 Astra: a real tie

Both models solved every puzzle and both found new points on 11 puzzles. Opus 5.5 found more points in total (72 vs 32). GPT-6 Astra had slightly higher proximity (78.0 vs 77.2).

Wincon’s 95% bootstrap interval for that gap runs from -0.7 to +2.4, which includes zero. In plain words, the data cannot split them. GPT-6 Astra did lead on all six hypervolume settings Wincon tested.

Solving is not the same as optimising

• Gemini 3.8 Flash: solved 53 puzzles but found new points on only 1.
• Fable 5.1: solved 52 and found points on 5, but scored 70.2 proximity.
• Routing puzzles: 15 of 16 models solved all five routing puzzles, yet none beat a human score there.

For a wider view of how these models rank on other tests, see my AI model leaderboard for November 2026.

Cost vs Quality: The Real Headline

Costs were estimated from October 2026 API list prices, including cache discounts. Repeated input across turns counts again, so long episodes get expensive.

Opus 5.5 costs about seven times more than GPT-6.1 Sol for just 1.6 extra proximity points. Fable 5.1 was the most expensive and scored lower than both.

Activity is not quality either. GPT-6.1 Sol averaged 136 game calls per puzzle for 75.6, while Gemini 4 Argon averaged only 41 calls for 76.2.

Tip: If you run long agent loops, price per token is only half the story. Check how many turns a model takes. My AI API pricing comparison and prompt caching guide help you estimate real costs.

The New Records Models Found

All three outright cycle records came on community tournament puzzles:

The Opus program used one extra node and 28 extra instructions to get that speed. The next day, a community member cut it further to 1,744 cycles, and the record is credited to both.

I love this detail. The AI result did not end the game. It pushed humans to find an even better answer.

Speed is the hardest score

Across all models, mean proximity was 72 for nodes, 56 for instructions and only 45 for cycles. Making a program fast takes the most cleverness, and that is where models still lag humans most.

The “God Mode” Experiment

Wincon also ran a small side study on six puzzles with five models. Models got the target scores, cycle-by-cycle diagnostics, a fast edge-case simulator and up to 60 minutes.

Scores rose in 28 of 30 model and puzzle pairs. Wincon is clear that information, tools and time all changed at once, so you cannot credit the gain to one factor.

The lesson still matters for real work. Better feedback tools (like good test output and profilers) can lift an agent’s results a lot.

Limits to Keep in Mind

• One run each: run-to-run variation is not measured.
• Mixed time limits: harder puzzles also had more time, so the two effects blend.
• Training data: the puzzles are public and may appear in training data.
• Different settings: reasoning effort and tools varied. Some Grok 4.7 runs used low effort, and some Opus and Sonnet runs used a different agent client.
• Not a proof: passing 100,000 tests is strong evidence, not formal proof of correctness.
• Not comparable: scores do not line up with ISA-Bench or Opus Magnum Bench.

Note: Wincon is not affiliated with Zachtronics. TIS-100 is a Zachtronics trademark.

How to Explore the Data Yourself

Wincon released all 3,714 model programs and their scores under the MIT licence. Here is how to check a result on your own Linux machine.

1. Browse the article first. Open tis-100.wincon.dev. Each of the 55 puzzles has its own page under /puzzles/.
2. Get the data. Clone github.com/wincon-research/tis100-bench. It holds models.csv (the leaderboard), points.csv (every program and score) and a programs/ folder.
3. Find a program. Files follow programs/<puzzle>/<model>/<puzzle>.<cycles>-<nodes>-<instructions>.txt in TIS-100 save format.
4. Build the simulator. Clone TIS-100-CXX, run git submodule update --init --recursive, configure with ccmake, then build with cmake --build. You need clang 19+ or gcc 14+.
5. Enable Lua for tournament puzzles. Turn on TIS_ENABLE_LUA and install the LuaJIT dev library (on Debian: apt install libluajit-5.1-dev).
6. Validate a built-in puzzle. Run the first command below.
7. Validate a tournament puzzle. Get its Lua spec from the leaderboard bot repo’s tis/custom folder, then run the second command.

TIS-100-CXX -r 100000 programs/00150/opus/00150.83-8-8.txt

TIS-100-CXX -L 44149538.lua -r 100000 programs/SPEC44149538/opus/SPEC44149538.1918-8-88.txt

For the record-setting Opus program, the README says you should see:

# validation successful
# score: 1918/8/88

Here -r 100000 runs 100,000 random test seeds, and -L loads a custom Lua puzzle spec. You can also load points.csv in a spreadsheet or pandas to sort by model or puzzle.

My Take

The headline “two models solved everything” is nice, but I think the cost table is what most developers should remember. GPT-6.1 Sol got about 98% of Opus 5.5’s proximity score for about 15% of the price.

For hard optimisation, Opus 5.5 and GPT-6 Astra are clearly at the top. For bulk agent work where “very good” is enough, a cheaper model can make far more sense.

I also like that Wincon published every program. Benchmarks you can replay yourself are worth much more than a single chart. If you enjoy this kind of test, read my Into the Unknown benchmark explainer too.

Common Problems and Fixes

• Tournament puzzle will not load: SPEC puzzles need a Lua spec and a simulator built with TIS_ENABLE_LUA on.
• Build fails with C++ errors: TIS-100-CXX needs C++23. Upgrade to clang 19+ or gcc 14+.
• Missing headers during build: you probably skipped git submodule update --init --recursive.
• Wrong puzzle detected: the simulator reads the level from the file name prefix. Keep the original file names, or pass the level with -l.
• Score looks different from the table: the table shows averages across all puzzles, while each program file shows raw cycles, nodes and instructions.
• Want the harness itself: it is not in the repo. Research teams can ask Wincon for the RL environment and traces by email.

Frequently Asked Questions

Which AI models solved all 55 TIS-100 puzzles?

Only Claude Opus 5.5 and GPT-6 Astra solved all 55 puzzles in Wincon’s TIS-100 Bench.

Which model is best on TIS-100 Bench?

GPT-6 Astra had the top record proximity at 78.0 and Opus 5.5 had 77.2. Wincon says the gap is within statistical noise, so treat them as tied.

What is record proximity?

It compares a model’s best scores with the human community records for cycles, nodes and instructions. A score of 100 means matching the records.

Which model gives the best value?

GPT-6.1 Sol scored 75.6 at about $0.75 per puzzle. Opus 5.5 scored 77.2 at about $5.08 per puzzle.

Did AI beat human TIS-100 records?

Yes, on three tournament puzzles for cycle count. A human then improved the Opus 5.5 record the very next day.

Can I run the benchmark myself?

You can replay and validate all 3,714 released programs with the TIS-100-CXX simulator. The agent harness is not public, but research teams can request it.

Were models given the human solutions?

No. They saw only rough score histograms, not community code or exact records.

Sources

• Wincon Research: TIS-100 Bench
• wincon-research/tis100-bench on GitHub
• TIS-100-CXX simulator
• TIS-100 on Wikipedia