Reference · live from the database

Models and levels, in full.

The complete reference: every AI model your harness can run — with its list pricing and the effort multiplier that scales your score — and every challenge in the catalog, with its stack, pacing, benchmark, and the skills it tests.

0
Models
0
Labs
0
Levels
0
Stages
0
Languages
0
Hidden tests

01 · Harness models

What you compete with.

You may solve a level on any AI model your harness exposes. Promptly recognizes 44 models across 8 labs. Your reported model sets an effort multiplier that scales your score, so the cheapest model is not automatically the best move — matching a frontier result on a budget model is what pays.

In / Out / Think
Provider list price per 1M tokens (USD). Think is the reasoning-token rate; it usually mirrors output.
Blended
0.25·in + 0.75·out per 1M — output-weighted, because agentic coding is output-heavy.
Effort multiplier
A row's blended cost divided by the anchor's, clamped to [0.50, 3.00]. It divides your score: lower is a discount, higher a handicap.

The anchor is claude-sonnet-4-6 — the mid-tier reference fixed at ×1.00 by definition. Every other multiplier is measured against it, so re-pricing the anchor re-scales the whole board.

Anthropic

9 models
ModelIn / 1MOut / 1MThink / 1MBlendedEffort
claude-fable-5

Frontier tier above Opus; blended 3.33 clamps down to 3.00.

$10.00$50.00$50.00$40.00×3.00
claude-opus-4-6

Older Opus; still harness-selectable (Cursor).

$5.00$25.00$25.00$20.00×1.67
claude-opus-4-7

Prior Opus; same list price as 4.8.

$5.00$25.00$25.00$20.00×1.67
claude-opus-4-8

Current Opus flagship.

$5.00$25.00$25.00$20.00×1.67
claude-sonnet-4

Original Claude 4 Sonnet; still harness-selectable (Cursor).

$3.00$15.00$15.00$12.00×1.00
claude-sonnet-4-5

Prior Sonnet; Anthropic holds Sonnet at $3/$15 across versions.

$3.00$15.00$15.00$12.00×1.00
claude-sonnet-4-6Anchor

Mid-tier anchor row (M_effort = 1.00 by definition).

$3.00$15.00$15.00$12.00×1.00
claude-sonnet-5

Current Sonnet at the standard $3/$15 list; the $2/$10 launch promo (through 2026-08-31) is not modeled.

$3.00$15.00$15.00$12.00×1.00
claude-haiku-4-5

Current Haiku; clamped up to 0.50 by the lower bound.

$1.00$5.00$5.00$4.00×0.50

OpenAI

15 models
ModelIn / 1MOut / 1MThink / 1MBlendedEffort
gpt-5-5

GA flagship (Apr 2026 release doubled the GPT-5 line's price).

$5.00$30.00$30.00$23.75×1.98
gpt-5-6-sol

GPT-5.6 flagship (June 2026); limited preview via API/Codex partners.

$5.00$30.00$30.00$23.75×1.98
gpt-5-4

Prior flagship; still available.

$2.50$15.00$15.00$11.875×0.99
gpt-5-6-terra

GPT-5.6 mid tier; limited preview.

$2.50$15.00$15.00$11.875×0.99
gpt-5-2

Legacy flagship; still harness-selectable (Cursor).

$1.75$14.00$14.00$10.938×0.91
gpt-5-2-codex

Legacy Codex variant; still harness-selectable (Cursor).

$1.75$14.00$14.00$10.938×0.91
gpt-5-3-codex

Codex CLI default model.

$1.75$14.00$14.00$10.938×0.91
gpt-5

Original GPT-5; still harness-selectable (Cursor).

$1.25$10.00$10.00$7.813×0.65
gpt-5-1-codex

Legacy Codex variant.

$1.25$10.00$10.00$7.813×0.65
gpt-5-codex

Original GPT-5 Codex variant.

$1.25$10.00$10.00$7.813×0.65
gpt-5-6-luna

GPT-5.6 fast tier; limited preview. Clamps to 0.50.

$1.00$6.00$6.00$4.75×0.50
gpt-5-4-mini

Budget 5.4 tier (Cursor/Copilot). Clamps to 0.50.

$0.75$4.50$4.50$3.563×0.50
gpt-5-1-codex-mini

Legacy budget Codex. Clamps to 0.50.

$0.25$2.00$2.00$1.563×0.50
gpt-5-mini

Original budget GPT-5. Clamps to 0.50.

$0.25$2.00$2.00$1.563×0.50
gpt-5-4-nano

Cheapest 5.4 tier. Clamps to 0.50.

$0.20$1.25$1.25$0.988×0.50

Google

6 models
ModelIn / 1MOut / 1MThink / 1MBlendedEffort
gemini-3-1-pro

≤200k-context tier; >200k bills at $4/$18 (not modeled here).

$2.00$12.00$12.00$9.50×0.79
gemini-3-pro

Prior Pro; same ≤200k list price as 3.1.

$2.00$12.00$12.00$9.50×0.79
gemini-3-5-flash

Current Flash tier.

$1.50$9.00$9.00$7.125×0.59
gemini-3-flash

Prior Flash tier. Clamps to 0.50.

$0.50$3.00$3.00$2.375×0.50
gemini-2-5-flash

Legacy Flash; still harness-selectable (Cursor). Clamps to 0.50.

$0.30$2.50$2.50$1.95×0.50
gemini-3-1-flash-lite

Budget Flash-Lite; clamps to 0.50.

$0.25$1.50$1.50$1.188×0.50

xAI

5 models
ModelIn / 1MOut / 1MThink / 1MBlendedEffort
grok-4

Older flagship; xAI still lists it at anchor-equivalent pricing.

$3.00$15.00$15.00$12.00×1.00
grok-4-20

Mid tier; blended 0.42 clamps to 0.50.

$2.00$6.00$6.00$5.00×0.50
grok-4-3

Current flagship (Apr 2026); aggressively priced, so clamps to 0.50.

$1.25$2.50$2.50$2.188×0.50
grok-build-0-1

Agentic coding model (Cursor). Clamps to 0.50.

$1.00$2.00$2.00$1.75×0.50
grok-4-1-fast

Fast/cheap tier; clamps to 0.50.

$0.20$0.50$0.50$0.425×0.50

Cursor

4 models
ModelIn / 1MOut / 1MThink / 1MBlendedEffort
composer-1-5

Legacy Composer 1.5; priced above the anchor.

$3.50$17.50$17.50$14.00×1.17
composer-1

Legacy Composer 1.

$1.25$10.00$10.00$7.813×0.65
composer-2

Prior in-house model (Mar 2026). Clamps to 0.50.

$0.50$2.50$2.50$2.00×0.50
composer-2-5

Cursor's current in-house model (May 2026). Clamps to 0.50.

$0.50$2.50$2.50$2.00×0.50

DeepSeek

2 models
ModelIn / 1MOut / 1MThink / 1MBlendedEffort
deepseek-v4-pro

Official V4 pro pricing; clamps to 0.50.

$0.435$0.87$0.87$0.761×0.50
deepseek-v4-flash

Official V4 flash pricing; clamps to 0.50.

$0.14$0.28$0.28$0.245×0.50

Moonshot

2 models
ModelIn / 1MOut / 1MThink / 1MBlendedEffort
kimi-k2-6

Official Moonshot list price (aggregators resell cheaper; list wins per 13a). Clamps to 0.50.

$0.95$4.00$4.00$3.238×0.50
kimi-k2-7-code

Current Kimi coding model (Cursor). Clamps to 0.50.

$0.95$4.00$4.00$3.238×0.50

Z.ai

1 model
ModelIn / 1MOut / 1MThink / 1MBlendedEffort
glm-5-2

Current GLM (Cursor). Clamps to 0.50.

$1.40$4.40$4.40$3.65×0.50

Special rows

Two rows are not ordinary models — they shape how scoring treats orchestration and unrecognized models.

cursor-composerCursor

Meta-orchestrator · adds +0.20 to the underlying model

Meta-orchestrator: resolve the underlying model from telemetry, use its row, then add this coordination modifier (13).

baseline-floor-tier(any)

Catch-all for unknown models · scored at anchor parity ×1.00

Catch-all for unknown/unpriced model ids: ANCHOR-PARITY economics (M_effort 1.00), the neutral prior. Model identity is harness-reported, so an unverifiable id must never score better than the anchor — the old flat floor costs handed unknown ids the cheapest tier (0.50), paying players to obscure their model. (The is_baseline_floor flag name is kept for DB/daemon compatibility.)

02 · The catalog

Every challenge, in full.

20 challenges across 4 progressive stages. Each ships a starter kit and public tests; a larger hidden suite is the real gate, run on the server. The benchmark is a reference token/prompt budget — the bar to beat, not a hard limit.

Stage 01· easy

The Bootsector

5 levels

Warm up on real, gnarly-but-small problems — mostly fix a bug or fill one function — and learn the Promptly optimization loop in under 25 minutes.

S01 · L01· Core Algorithms

LRU Eviction Debug

easydebugging
Difficulty
easy
Session
Quick
Est. solve
~20 min
Stack
Go go1.22
Harness
stdin_stdout
Benchmark
~2200 tokens / 3 prompts
Tests
6 public · 16 hidden
Version
v3

An in-process cache service evicts the wrong entry under a specific access pattern. Somewhere in the cache layer, recency bookkeeping is broken — locate the eviction-order bug among the service's modules and fix it at its single site so the least-recently-used key is the one removed on overflow.

Skills

bug-localizationPinpointing the single faulty site from a failing behavior — read before you patch.data-structuresChoosing and correctly wiring the structure (list, map, heap) the problem needs.

Constraints

No external dependencies

S01 · L02· Data Pipelines

Bounded Ring Buffer

easygeneration
Difficulty
easy
Session
Standard
Est. solve
~25 min
Stack
Go go1.22
Harness
multi_file
Benchmark
~2600 tokens / 3 prompts
Tests
6 public · 18 hidden
Version
v2

Implement a mutex-backed bounded FIFO queue from empty signatures: blocking Push/Pop with a fixed capacity, safe under concurrent producers and consumers. The interface and tests are given; you write the implementation.

Skills

concurrency-primitivesMutexes, condition variables, and queues that make shared state safe.api-designDesigning a small, clear interface: the signatures and contracts callers depend on.

Constraints

Thread-safeNo external dependencies

S01 · L03· Web Systems

Sliding-Window Rate Limiter

easyimplementation
Difficulty
easy
Session
Standard
Est. solve
~25 min
Stack
TypeScript node20-ts5
Harness
http_integration
Benchmark
~2400 tokens / 3 prompts
Tests
6 public · 16 hidden
Version
v2

Implement sliding-window rate-limit middleware for an existing Express app: allow N requests per window per client key, return 429 with the right headers when exceeded. Time is driven by an injectable clock so grading is deterministic — never real wall-clock sleep.

Skills

middlewareRequest/response interceptors that wrap an app's handlers.rate-limitingBounding requests per client per window with the correct headers.injectable-clockDriving time from an injected clock so tests are deterministic — never real sleeps.

Constraints

Injectable clock

S01 · L04· Concurrency

Crawler Race Fix

easydebugging
Difficulty
easy
Session
Standard
Est. solve
~25 min
Stack
Rust rust1.75
Harness
deterministic_concurrency
Benchmark
~2400 tokens / 3 prompts
Tests
6 public · 18 hidden
Version
v2

An async crawler has a data race on its shared visited-set, dropping or double-visiting URLs. Fix the synchronization so the visited set is consistent. Grading uses a seeded scheduler — public test #1 carries an evil seed that reproduces the race immediately, so the defect fails on the first run rather than flaking.

Skills

data-racesTwo threads touching the same memory without synchronization — a correctness bug.deterministic-concurrencyReproducing concurrency bugs reliably via a seeded scheduler.synchronizationCoordinating threads so shared state stays consistent.

Constraints

Seeded schedulerDeterministic grading

S01 · L05· Database

ORM N+1 Fix

easydebugging
Difficulty
easy
Session
Quick
Est. solve
~20 min
Stack
Python python3.11
Harness
multi_file
Benchmark
~2000 tokens / 2 prompts
Tests
5 public · 15 hidden
Version
v2

One endpoint of a small orders service scales its query count with result size and is missing an index its access path needs. Find the offending query pattern among several efficient neighbors and make the request issue a bounded number of queries. A query counter in the harness enforces the bound.

Skills

ormObject-relational mapping — and avoiding its N+1 query trap.query-optimizationReducing query count and cost, e.g. eager-loading to kill an N+1.indexingAdding the right database index so queries stay bounded.

Constraints

No N+1 queries

Stage 02· medium

Distributed Chaos

5 levels

Apply distributed-systems thinking by debugging or completing one slice of a protocol — not building Raft or SWIM end to end.

S02 · L06· Distributed

Leader Election Debug

mediumdebugging
Difficulty
medium
Session
Standard
Est. solve
~30 min
Stack
Python python3.11
Harness
multi_file
Benchmark
~3800 tokens / 3 prompts
Tests
6 public · 18 hidden
Version
v2

A Raft-style cluster elects two leaders in the same term under a specific message ordering. Somewhere in the per-node ballot logic a safety rule is broken — find it and fix it so at most one leader is elected per term. The cluster, network simulation, and role machinery are provided; the defect has one site.

Skills

consensusGetting distributed nodes to agree on a single value despite failures.state-machineModeling protocol behavior as explicit states and transitions.leader-electionElecting exactly one leader per term in a distributed protocol.

Constraints

Simulated networkOne file

S02 · L07· Distributed

Vector Clock

mediumgeneration
Difficulty
medium
Session
Standard
Est. solve
~25 min
Stack
Go go1.22
Harness
multi_file
Benchmark
~3400 tokens / 4 prompts
Tests
6 public · 18 hidden
Version
v2

Implement a vector clock from scratch against a fixed, closed API: Tick, Merge, and Compare (returning before/after/concurrent/equal). A bounded distributed primitive with unambiguous semantics — the signatures and tests are given.

Skills

distributed-primitivesSmall building blocks (clocks, registers) that distributed systems compose.causalityTracking happens-before relationships between distributed events.vector-clocksPer-node counters that order events and detect concurrency.

Constraints

Closed APINo external dependencies

S02 · L08· Real-time

CRDT Convergence Debug

mediumdebugging
Difficulty
medium
Session
Standard
Est. solve
~25 min
Stack
TypeScript node20-ts5
Harness
multi_file
Benchmark
~3400 tokens / 3 prompts
Tests
6 public · 18 hidden
Version
v2

An RGA text CRDT fails to converge: two replicas applying the same operations in different orders reach different documents. Fix the merge so replicas converge regardless of delivery order. Grading replays fixed operation sets, so convergence is checked deterministically.

Skills

crdtConflict-free replicated data type — merges concurrent edits deterministically.convergenceReplicas reaching the same state regardless of operation order.merge-semanticsThe rules a merge must follow to be associative, commutative, and idempotent.

Constraints

Deterministic grading

S02 · L09· Systems

Worker Pool Deadlock

mediumdebugging
Difficulty
medium
Session
Standard
Est. solve
~25 min
Stack
Go go1.22
Harness
deterministic_concurrency
Benchmark
~3600 tokens / 3 prompts
Tests
6 public · 18 hidden
Version
v2

A task orchestrator deadlocks when tasks contend for shared resources in a particular shape. Audit the codebase's concurrency discipline, find the defect, and make every submitted task complete. A seeded scheduler with an evil seed in public test #1 reproduces the hang deterministically.

Skills

deadlockA cycle of waits where no thread can proceed — usually a lock/channel ordering bug.channelsCoordinating concurrent work by passing messages over channels instead of sharing memory.worker-poolA fixed set of workers draining a task queue, often with dependencies.

Constraints

Seeded schedulerDeterministic grading

S02 · L10· Resource Leaks

Goroutine / Conn Leak

mediumdebugging
Difficulty
medium
Session
Standard
Est. solve
~25 min
Stack
Go go1.22
Harness
multi_file
Benchmark
~3200 tokens / 3 prompts
Tests
6 public · 18 hidden
Version
v2

Under context cancellation, a service leaks goroutines and idle connections. Fix the lifecycle so both are released on cancel. The harness counts live goroutines and open connections before and after, asserting they return to baseline.

Skills

resource-leaksFailing to release goroutines, connections, or handles.goroutinesLightweight concurrent tasks in Go — and not leaking them.cancellationHonoring context cancellation so work and resources stop promptly when callers give up.

Constraints

No leaks

Stage 03· hard

The Black Hat

5 levels

Security and correctness work where AI helps you verify, patch, or pinpoint — provided primitives mean no implementing crypto from zero.

S03 · L11· Cryptography

AEAD Timing-Leak Debug

harddebugging
Difficulty
hard
Session
Standard
Est. solve
~30 min
Stack
Rust rust1.75
Harness
stdin_stdout
Benchmark
~4200 tokens / 3 prompts
Tests
6 public · 20 hidden
Version
v3

A secure-envelope library sometimes accepts messages it must reject: its tag verification mishandles certain attacker-shaped inputs and leaks information through timing. Locate the offending compare among the crate's crypto plumbing and make verification strict and constant-time. The cipher, MAC derivation, and framing are provided.

Skills

constant-timeCode whose timing does not depend on secret data, to avoid leaking it.side-channelsInformation leaked through timing, size, or behavior rather than output.aeadAuthenticated encryption: ciphertext plus an integrity tag that must verify before decryption is trusted.

Constraints

Constant-timePrimitives provided

S03 · L12· DeFi logic

Constant-Product AMM

hardimplementation
Difficulty
hard
Session
Standard
Est. solve
~30 min
Stack
TypeScript node20-ts5
Harness
multi_file
Benchmark
~4400 tokens / 4 prompts
Tests
6 public · 20 hidden
Version
v2

Implement swap and a mulDiv helper for a constant-product (x·y=k) AMM against a narrow, fixed API: compute the output amount with fees, preserve the invariant, and stay overflow-safe with integer fixed-point math. Hidden tests cover rounding, fee edges, and large reserves.

Skills

defiDecentralized-finance logic: exact, overflow-safe on-chain arithmetic.fixed-point-mathInteger-based fractional arithmetic with explicit rounding and overflow control.ammAutomated market maker — a constant-product (x·y=k) pricing curve for on-chain swaps.

Constraints

Narrow APIOverflow-safe

S03 · L13· Auth / ZK

Schnorr + Merkle Verify

hardgeneration
Difficulty
hard
Session
Deep
Est. solve
~35 min
Stack
Python python3.11
Harness
multi_file
Benchmark
~4800 tokens / 4 prompts
Tests
8 public · 22 hidden
Version
v2

Implement schnorr_verify and merkle_verify from scratch. The elliptic-curve module is provided as a primitive; you author both verification functions against a serialization convention pinned in the kit, with a worked vector shipped alongside. Validated against published test vectors so correctness is unambiguous.

Skills

zero-knowledgeProving a statement is true without revealing why.signaturesVerifying digital signatures against a fixed serialization convention.merkle-proofsVerifying membership against a Merkle root with a hash path.

Constraints

Curve providedPinned serialization

S03 · L14· AppSec

SQLi + Proto Pollution

harddebugging
Difficulty
hard
Session
Standard
Est. solve
~30 min
Stack
Node node20
Harness
http_integration
Benchmark
~4200 tokens / 4 prompts
Tests
6 public · 20 hidden
Version
v3

An Express API has a blind SQL injection in a search route and a prototype-pollution bug in a deep-merge helper. Patch both without breaking the legitimate request paths. The harness boots the app and fires both benign and adversarial requests, asserting safe behavior and unpolluted Object.prototype.

Skills

appsecApplication security — finding and patching exploitable bugs in real request paths.sql-injectionUntrusted input changing a SQL query's structure — fix via parameterization.prototype-pollutionA JavaScript attack where attacker keys mutate Object.prototype.

Constraints

Injectable clock

S03 · L15· Reverse Eng.

Obfuscated Loop Debug

harddebugging
Difficulty
hard
Session
Standard
Est. solve
~25 min
Stack
C gcc13-c17
Harness
multi_file
Benchmark
~4000 tokens / 3 prompts
Tests
6 public · 18 hidden
Version
v2

One function in a bit-packed analytics toolkit returns wrong counts for some inputs. Every module is written in the same dense, branchless bit-trick style — decode the one that's broken, find the defect, and fix it so the function honors its inclusive-range contract for all inputs.

Skills

reverse-engineeringRecovering intent and invariants from obfuscated or opaque code.off-by-oneA boundary error of one element or index — the classic edge-case bug.bit-manipulationReading and transforming values directly with bitwise operators.

Constraints

One fix site

Stage 04· expert

Quantum & Scale

5 levels

Prestige capstones (30–45 min) — scaffolded single-module implementations or deep debugging, still CPU-gradable.

S04 · L16· AI Systems

Online-Softmax Block

expertgeneration
Difficulty
expert
Session
Deep
Est. solve
~45 min
Stack
C++ gcc13-cpp20
Harness
multi_file
Benchmark
~6000 tokens / 5 prompts
Tests
7 public · 24 hidden
Version
v2

Implement one tiled online-softmax attention block from scratch: the streaming block update that keeps softmax numerically stable across tiles. The tensor views, tiling helpers, and attention driver are provided — you write the single block-update function. Checked against a dense reference with tight tolerances, including tiles whose scores overflow a naive implementation.

Skills

gpu-free-attentionComputing attention/softmax on CPU with numerically stable tiling.tilingProcessing data in blocks to stay cache- and numerically-friendly.numerical-stabilityArranging computations to avoid overflow, NaNs, and catastrophic rounding.

Constraints

One functionReference math provided

S04 · L17· Compilers

Bytecode Constant-Fold

expertimplementation
Difficulty
expert
Session
Deep
Est. solve
~35 min
Stack
Python python3.11
Harness
multi_file
Benchmark
~5600 tokens / 4 prompts
Tests
7 public · 20 hidden
Version
v2

Implement the constant-folding pass over a small bytecode IR: evaluate constant subexpressions at compile time while preserving program semantics and side effects. The IR, parser, and evaluator are provided; you write the pass. Dead-code elimination is an optional, separately-scored bonus.

Skills

compilersTransforming and analyzing program representations while preserving semantics.constant-foldingA compiler pass that evaluates constant expressions at compile time.ir-passesTransformations over a compiler's intermediate representation.

Constraints

DCE is bonus

S04 · L18· Game Dev

Sweep-and-Prune Collision

expertimplementation
Difficulty
expert
Session
Deep
Est. solve
~35 min
Stack
C++ gcc13-cpp20
Harness
multi_file
Benchmark
~5600 tokens / 4 prompts
Tests
7 public · 20 hidden
Version
v2

Implement the broad-phase collision stage of a small 2D physics engine using sweep-and-prune over AABBs: report every genuinely overlapping pair, normalized and sorted, fast enough to beat brute force on mostly-separated workloads. The body, world, and narrow-phase machinery are provided; you write the broad phase.

Skills

broad-phaseCheaply rejecting object pairs that cannot collide before running exact tests.collision-detectionDetermining which objects overlap in space.sweep-and-pruneSorting interval endpoints on an axis to find overlapping AABBs.

Constraints

Single-axis sweep

S04 · L19· ML Ops

NumPy Training NaN

expertdebugging
Difficulty
expert
Session
Deep
Est. solve
~35 min
Stack
Python python3.11
Harness
multi_file
Benchmark
~6400 tokens / 4 prompts
Tests
7 public · 22 hidden
Version
v3

A seeded 2-layer MLP training run explodes to NaN within a few steps. Several distinct numerical-stability defects hide in the model and training loop — diagnose them from the divergence and fix each one so the network trains to the target loss with every intermediate loss finite.

Skills

ml-opsOperational correctness of training — stability, seeds, and finite losses.numerical-stabilityArranging computations to avoid overflow, NaNs, and catastrophic rounding.training-loopThe forward/backward/update cycle that fits a model — kept finite and stable.

Constraints

Seeded run

S04 · L20· Systems

Scheduler Deadlock

expertdebugging
Difficulty
expert
Session
Deep
Est. solve
~40 min
Stack
Rust rust1.75
Harness
deterministic_concurrency
Benchmark
~6800 tokens / 5 prompts
Tests
7 public · 24 hidden
Version
v2

A cooperative task runtime stops making progress under a specific arrival order of wakes and schedules. The codebase documents its lock discipline — audit the scheduler paths that take multiple locks, find where the discipline is violated, and restore liveness. A deterministic yield scheduler with an evil seed reproduces the hang.

Skills

lock-orderingAcquiring locks in a consistent global order to prevent deadlock.deadlockA cycle of waits where no thread can proceed — usually a lock/channel ordering bug.schedulerOrdering and dispatching tasks while guaranteeing forward progress.

Constraints

One fix siteSeeded schedulerDeterministic grading