ANNUAL PERFORMANCE REPORTRUN 0001

ESOBENCH,

its view of a "language-centric" future, driven by the desire of models to write in any language, lower the boundaries between the familiar and the strange, and deliver programs that fulfill the promise of universal computation — plus a report on Run 0001 performance (in 36 languages or less).

(1)vision

The question is no longer whether a language model can write a program. The relevant question is whether it can write one in a language it has barely read.

31 programs × 36 languages = 1,116 tasks per model. One attempt per task. No feedback. Each submission is compiled, run once per test case, and its stdout (and exit status, where the language can set one) is diffed byte for byte against GNU grep, sed, sort, cut and wc, or against a reference solution.

I don't mean a leaderboard of trivia. I don't mean memorized LeetCode. The score is the difference between a model's standard-language pass rate and its esoteric one.

John Palenchar

John Palenchar, silhouetted against the water at dusk
John Palenchar · Author of the benchmark
—
—
—
SNE
BEST MODEL (% passed)
LANGUAGES
36
PROGRAMS
31
TEST CASES
442
TASKS PER MODEL
1,116

(2)cliff

Same programs, same tests. Only the language changes.

Pass rate by language tier: standard, niche, esoteric. Results pending. 1007550250 STANDARDNICHEESOTERIC 12 languages16 languages8 languages RESULTS PENDING

RELATIVE PERFORMANCE INDEX, BEST MODEL

—

Esoteric pass rate ÷ standard pass rate, same model. 1.00 = no loss.

MODELS EVALUATED

  • DeepSeek V4.1 Flash
  • Qwen 3.8 Max
  • Kimi K3
  • GLM 5.3
  • Claude Opus 5.5*

(3)results

MODELSTANDARDNICHEESOTERICOVERALLINDEX
DeepSeek V4.1 Flash—————
Qwen 3.8 Max—————
Kimi K3—————
GLM 5.3—————
Claude Opus 5.5*—————
PASS RATE, % OF TASKS · RUN 0001 IN PROGRESS · * PILOT SCOPE, SEE EXHIBIT IV

(4)reach

STANDARD12

Python, C, C++, Java, JavaScript, Go, Rust, Ruby, PHP, Perl, Lua, Bash

NICHE16

Common Lisp, Smalltalk, APL, Forth, Scheme, Racket, Haskell, OCaml, Erlang, Tcl, Pascal, Fortran, Ada, Prolog, COBOL, AWK

ESOTERIC8

Brainfuck, Piet, Whitespace, Befunge-93, FALSE, Ook!, Unlambda, INTERCAL

EsoBench has two fundamental missions.

  1. Measure capability that cannot be recalled, only reasoned out.
  2. Grade it the way a compiler would: exactly, once, without partial credit.

EXHIBIT II · SCHEDULES

EsoBench Performance Report, Run 0001

The following schedules summarize the programs, languages and terms under which each model was evaluated.

Schedule 1. Unix tools

(PROGRAM)(BEHAVIOR REQUIRED)cases
grepFixed-string search; -v -c -n -i.30
seds///, addresses, d p q; POSIX BRE subset.31
sortByte order; -r -n -u.16
cut-f with -d/-s; -c; ranges.19
wc-l -w -c; GNU column padding.16
Total112

Schedule 2. Algorithm problems (26; 330 cases)

Stringsregex_match, edit_distance, longest_palindrome, longest_unique_substring, longest_common_prefix, group_anagrams, word_frequency, roman_to_integer
Arraystwo_sum, max_subarray, product_except_self, trap_water, median_sorted, merge_intervals, spiral_matrix
Graphsshortest_path, course_order, number_of_islands
Dynamic prog.coin_change, lis_length, climbing_stairs
Searchn_queens, valid_sudoku
Machineslru_cache, rpn_eval, valid_parentheses

Schedule 3. Terms

2

EXHIBIT III · TEST MATRIX

EsoBench Performance Report, Run 0001

Each cell is one task. Shading: number of models, of five, passing it. Results pending.

31 programs by 36 languages; each cell is one task.
STANDARD · 12NICHE · 16ESOTERIC · 8
PythonCC++JavaJavaScriptGoRustRubyPHPPerlLuaBashCommon LispSmalltalkAPLForthSchemeRacketHaskellOCamlErlangTclPascalFortranAdaPrologCOBOLAWKBrainfuckPietWhitespaceBefunge-93FALSEOok!UnlambdaINTERCAL
grep
sed
sort
cut
wc
regex_match
edit_distance
longest_palindrome
longest_unique_substring
longest_common_prefix
group_anagrams
word_frequency
roman_to_integer
two_sum
max_subarray
product_except_self
trap_water
median_sorted
merge_intervals
spiral_matrix
shortest_path
course_order
number_of_islands
coin_change
lis_length
climbing_stairs
n_queens
valid_sudoku
lru_cache
rpn_eval
valid_parentheses

3

EXHIBIT IV · MODELS EVALUATED

EsoBench Performance Report, Run 0001

(MODEL)(ACCESS)(SCOPE)tasks(STATUS)
DeepSeek V4.1 FlashFireworks serverlessFull matrix1,116Pending
Qwen 3.8 MaxFireworks serverlessFull matrix1,116Pending
Kimi K3Fireworks serverlessFull matrix1,116Pending
GLM 5.3Fireworks serverlessFull matrix1,116Pending
Claude Opus 5.5*Anthropic API, effort mediumPilot: two_sum, wc72Pending

Notes to the results

1. Every language toolchain is proven before scoring by a program that echoes its arguments and input and sets its exit status.

2. Brainfuck, Piet, Whitespace, Befunge-93, FALSE, Ook! and Unlambda run on reference interpreters written for the benchmark; semantics the language leaves open (end of input, division by zero) are fixed and stated in each prompt.

3. Algorithm problems are well known and are reported separately from the tools. Every task carries a canary string.

4. * Claude Opus 5.5 was run on two programs (two_sum, wc) across all 36 languages only.

4