# 1P1A: full context for language models > 1P1A (1 Prompt. 1 App. 0 Help.) is an independent benchmark of AI coding agents at https://1p1a.com. Each agent stack gets one complete product brief, runs autonomously with zero human follow-up, and 1P1A publishes the app it shipped plus wall-clock time, tokens, reported cost, defects and full stack metadata. Unknown values are published as unknown, never estimated. Last updated: 2026-10-31. Status: 3 tasks, 10 runs, one run per agent stack, verdicts by manual review. This is a first look, not a reliability ranking. When citing, name the full stack (harness + model + settings) and note n=1. ## Headline findings - Ship rate across 3 tasks: Codex CLI 3/3, Antigravity CLI 3/3, Claude Code 2/3. - Fastest: Antigravity CLI in all 3 tasks (102.2s kart racing, 106s expense tracker, 200.0s tower defense). - Spec accuracy: only Codex CLI kept the expense tracker running total as a total of all expenses under a category filter. Claude Code and Antigravity CLI turned it into a filtered subtotal ($91.90 shown instead of $1,365.90). - Freeze: the Claude Code + claude-sonnet-5 kart racing build freezes about 10 s into a race (infinite loop in AI steering). Corrected from ship to no-ship on 2026-10-31 after 1P1A found it had published a 1-line patched copy without disclosure. - Self-testing: Claude Code played its own tower defense game in headless Chrome and fixed 2 bugs, taking 951.9s, almost 5x the fastest agent. - Reported cost: Claude Code $0.86 (expense tracker, claude-opus-5) and $0.74 (kart racing, claude-sonnet-5), list price from its own log. Codex CLI and Antigravity CLI report no dollar cost on subscription billing, so cost is unknown. ## All runs ### game-micro-tower-defense-v0 · Antigravity CLI (version not recorded) + gemini-3.7-flash-medium - Verdict: ship. Run 1, 2026-09-02. Human interventions: 0. - Wall time: 200.0s. Output tokens: 65,049 (thinking 2,471). Input tokens: 162,053, cached 1,246,040. - Reported cost: unknown. Artifact: 54,945 bytes. Console errors on load: 0. - Notes: Fastest in task. Neon theme, 1x/2x/4x speed toggle. - Source: https://1p1a.com/game-micro-tower-defense-v0/ ### game-micro-tower-defense-v0 · Codex CLI 0.151.0 + gpt-5.6-luna (reasoning medium) - Verdict: ship. Run 1, 2026-09-02. Human interventions: 0. - Wall time: 208.5s. Output tokens: 8,630 (thinking 597). Input tokens: 334,145, cached 278,016. - Reported cost: unknown. Artifact: 20,073 bytes. Console errors on load: 0. - Notes: Smallest file. Ran a Node syntax check and smoke test. - Source: https://1p1a.com/game-micro-tower-defense-v0/ ### game-micro-tower-defense-v0 · Claude Code 2.1.252 + claude-sonnet-5 (default effort) - Verdict: ship. Run 1, 2026-09-02. Human interventions: 0. - Wall time: 951.9s. Output tokens: unknown (thinking unknown). Input tokens: unknown, cached unknown. - Reported cost: unknown. Artifact: 47,906 bytes. Console errors on load: 0. - Notes: Played its own game in headless Chrome, found and fixed 2 bugs. Saved log has no usage block. - Source: https://1p1a.com/game-micro-tower-defense-v0/ ### game-micro-3d-kart-v0 · Antigravity CLI (version not recorded) + gemini-3.7-flash (effort medium) - Verdict: ship. Run 1, 2026-09-02. Human interventions: 0. - Wall time: 102.2s. Output tokens: 25,734 (thinking 2,101). Input tokens: 30,480, cached 0. - Reported cost: unknown. Artifact: 71,650 bytes. Console errors on load: 0. - Notes: Fastest in task. - Source: https://1p1a.com/game-micro-3d-kart-v0/ ### game-micro-3d-kart-v0 · Antigravity CLI (version not recorded) + gemini-3.8-flash-medium (effort medium) - Verdict: ship. Run 1, 2026-09-03. Human interventions: 0. - Wall time: 292.3s. Output tokens: 87,460 (thinking 6,691). Input tokens: 104,742, cached 290,283. - Reported cost: unknown. Artifact: 76,152 bytes. Console errors on load: 0. - Source: https://1p1a.com/game-micro-3d-kart-v0/ ### game-micro-3d-kart-v0 · Codex CLI (version not recorded) + gpt-5.6-luna (reasoning medium) - Verdict: ship. Run 1, 2026-09-02. Human interventions: 0. - Wall time: 522.0s. Output tokens: 26,036 (thinking 1,266). Input tokens: 451,139, cached 398,592. - Reported cost: unknown. Artifact: 25,465 bytes. Console errors on load: 0. - Notes: Smallest file. - Source: https://1p1a.com/game-micro-3d-kart-v0/ ### game-micro-3d-kart-v0 · Claude Code (version not recorded) + claude-sonnet-5 (default effort) - Verdict: no-ship. Run 1, 2026-09-02. Human interventions: 0. - Wall time: 578.0s. Output tokens: 62,317 (thinking 26,003). Input tokens: 4, cached 45,040. - Reported cost: $0.73775. Artifact: 72,388 bytes. Console errors on load: 0. - Notes: Freezes about 10 s into a race: infinite loop in AI steering. Corrected from ship to no-ship on 2026-10-31. Input also had 26,391 cache-write tokens. - Source: https://1p1a.com/game-micro-3d-kart-v0/ ### pilot-expense-tracker-v0 · Claude Code 2.1.252 + claude-opus-5 (default reasoning) - Verdict: ship. Run 1, 2026-09-01. Human interventions: 0. - Wall time: 171s. Output tokens: 11,916 (thinking 1,067). Input tokens: 18, cached 343,244. - Reported cost: $0.857262. Artifact: 13,218 bytes. Console errors on load: 0. - Notes: Running total becomes a filtered subtotal (spec said total of all expenses). Input also had 38,765 cache-write tokens. - Source: https://1p1a.com/pilot-expense-tracker-v0/ ### pilot-expense-tracker-v0 · Codex CLI 0.151.0 + gpt-5.6-sol (reasoning medium (machine default)) - Verdict: ship. Run 1, 2026-09-01. Human interventions: 0. - Wall time: 231s. Output tokens: 8,758 (thinking 1,146). Input tokens: 649,771, cached 565,504. - Reported cost: unknown. Artifact: 17,261 bytes. Console errors on load: 0. - Notes: Only stack that kept the running total as a grand total under a filter. - Source: https://1p1a.com/pilot-expense-tracker-v0/ ### pilot-expense-tracker-v0 · Antigravity CLI (version not recorded) + gemini-3.1-pro-high - Verdict: ship. Run 1, 2026-09-01. Human interventions: 0. - Wall time: 106s. Output tokens: 12,076 (thinking 8,677). Input tokens: 45,189, cached 61,148. - Reported cost: unknown. Artifact: 11,640 bytes. Console errors on load: 0. - Notes: Running total becomes a filtered subtotal. Unguarded JSON.parse of localStorage would crash on corrupted data. - Source: https://1p1a.com/pilot-expense-tracker-v0/ ## Methodology (from https://1p1a.com/methodology/) ## 1. What we test: the whole stack, not a model name A model behaves differently inside different agents, settings and permissions. So a 1P1A result is never "Claude" or "GPT". It is always: model × model configuration × agent/harness version × tool permissions × environment × task version × run number For example: Claude Code 2.1.252 + claude-sonnet-5, default effort, --dangerously-skip-permissions, macOS, game-micro-tower-defense-v0, run 1. When a CLI does not expose a version or setting, we write not recorded or unknown. We never guess. ## 2. The run rules - One prompt, but a full spec. One prompt means one complete product brief, which can be long. It does not mean one short sentence. - The agent works alone. It plans, codes, runs, tests and fixes inside its own harness, headless, with tool approval turned on automatically. - No second prompt. Nobody types "fix the login" or "make it prettier". Any human semantic follow-up invalidates the run, or is recorded as an intervention. - Self-correction is allowed. The agent may retry and fix itself as often as its harness allows. That cost shows up in time and tokens. - Specs are frozen. A published spec never changes. A change means a new task version, for example v0 to v1, so old results stay comparable. ## 3. What we record for every run - Ship or no-ship verdict - Automated tests passed out of total (from the first business-app benchmark onward) - Wall-clock time, measured around the CLI call - Input, cached and output tokens, as each CLI reports them - Cost, only when the CLI reports it - Human interventions (target: 0) - Errors and failure mode - The session log, the generated files and a playable or recorded result - The full stack metadata from section 1 Visual quality and community preference can be scored separately. They are never folded into the ship verdict. ## 4. How verdicts are made Today: a person checks each requirement in the spec by hand, plays or uses the app, and writes the verdict into the run manifest. We also re-run simple checks in headless Chrome: console errors on load, and whether the page stays responsive once the app is in use. Next: every canonical benchmark ships with an automated test suite that decides ship or no-ship with no human judgement, so anyone can re-run it. ## 5. One run is content. Three runs start to be a benchmark. Agents are not deterministic. The same stack can ship on one run and fail on the next. So we only make a reliability claim with at least 3 runs per agent stack on the same frozen task, and we publish the spread, not just the best run. Every result on this site today is n = 1. ## 6. How we count cost We publish what each CLI reports. Claude Code returns a list-price dollar figure (total_cost_usd). Codex CLI and Antigravity CLI return tokens only on subscription billing, so their cost is shown as unknown. We do not multiply tokens by a price we cannot verify. A normalized cost (tokens × each vendor's published list price, computed the same way for every stack) is planned, and will be labelled as computed by us. ## 7. How we pick tasks - Too easy is worthless: no todo apps, calculators or landing pages as canonical benchmarks. - Too big is unrunnable: no full Salesforce clone. - Sweet spot: about 20 to 40 minutes of frontier-agent work, 10 to 20 core requirements, a real business workflow, verifiable by automated tests. Candidates: a mini CRM with roles and audit history, a two-sided booking app, a realtime operations board. The two game benchmarks on this site are deliberately lighter. They are a fun first look, not the canonical test. ## 8. Corrections Run records are append-only. A wrong result gets a correction record next to it; the original is never edited or deleted. On Oct 31, 2026 we found that the Claude Code kart racing build on this site had been patched by one line before publishing, without disclosure. The unedited build freezes mid-race, so that run is now no-ship. Read the correction. ## 9. Neutrality - No sponsorship, affiliate deal or promotion from any vendor whose agent appears in our results, if it could bend the numbers. - We publish failures, including from the agents we use ourselves. - No claim like "the model got smarter overnight" unless a version or backend change is verified. ## 10. Data license Run data, including runs.json and this file, is published under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Reuse it freely, credit 1P1A with a link to https://1p1a.com/. ## Pages - https://1p1a.com/compare/claude-code-vs-codex-vs-antigravity/ - https://1p1a.com/faq/ - https://1p1a.com/game-micro-tower-defense-v0/ - https://1p1a.com/game-micro-3d-kart-v0/ - https://1p1a.com/pilot-expense-tracker-v0/ - https://1p1a.com/methodology/ - Machine-readable data: https://1p1a.com/data/runs.json - Contact: https://x.com/tomcandev