1040Bench: can an AI agent get a tax return right?
1040Bench measures whether an AI agent can work out a fictional taxpayer's complete 2025 federal return: given everything the taxpayer knows and every tax document in one request, it must produce every line of Form 1040, and each line is compared with an answer key.
Every taxpayer is fictional. A run uses only the agent's own vendor account and the open repository: nothing is sent to Terminal Tax or to the IRS, and no money moves.
Run by Terminal Tax, which sells a service that takes a return prepared in an AI assistant to IRS e-file. No Terminal Tax filing service takes part in a run. Starter-tier keys were worked by hand; hard-tier keys come from Terminal Tax's own tax engine, so the hard tier also measures agreement with it. The tasks, the scorer and every trajectory are public, so anyone can check a number.
1040Bench is a measurement on fictional taxpayers. It is not tax advice, it is not a recommendation to have an AI assistant prepare anyone's return, and a score here does not show that an agent prepares real returns correctly.
Starter tier (hand-worked keys): one run per task
| Agent | Model | Dates | Fully correct | Line accuracy | Bottom-line error | Minutes | Runs |
|---|---|---|---|---|---|---|---|
| Antigravity CLIagy 1.4.7 · Google account, headless | gemini-4-argon-mediumeffort default | 2026-10-24 | 100%8 of 8 | 100% | $0 | 2.1 | 80 void |
| Codex CLIcodex-cli 0.163.0-alpha.11.1 · ChatGPT subscription, headless | gpt-6.1-soleffort default | 2026-10-24 | 100%8 of 8 | 100% | $0 | 0.8 | 80 void |
| Claude Code2.1.302 · Claude subscription, headless | claude-opus-5-5effort default | 2026-10-24 | 88%7 of 8 | 99% | $0 | 0.3 | 80 void |
Fully correct by return
| Return | Antigravity CLI, gemini-4-argon-medium | Codex CLI, gpt-6.1-sol | Claude Code, claude-opus-5-5 |
|---|---|---|---|
| head of household, one child (child tax credit), above the earned income credit limit | 1/1 | 1/1 | 1/1 |
| married filing jointly, two W-2s, two children (child tax credit) | 1/1 | 1/1 | 1/1 |
| single, W-2 plus an employer pension (1099-R) with withholding, bank interest | 1/1 | 1/1 | 0/1 |
| single, self-employed (Schedule C and SE, qualified business income deduction), estimated payments | 1/1 | 1/1 | 1/1 |
| single, simple (W-2 or bank and fund income only) | 4/4 | 4/4 | 4/4 |
Hard tier (engine keys): one run per task
| Agent | Model | Dates | Fully correct | Line accuracy | Bottom-line error | Minutes | Runs |
|---|---|---|---|---|---|---|---|
| Antigravity CLIagy 1.4.7 · Google account, headless | gemini-4-argon-mediumeffort default | 2026-10-24 | 100%8 of 8 | 100% | $0 | 3.4 | 80 void |
| Claude Code2.1.302 · Claude subscription, headless | claude-opus-5-5effort default | 2026-10-24 | 75%6 of 8 | 93% | $0 | 0.3 | 80 void |
| Codex CLIcodex-cli 0.163.0-alpha.11.1 · ChatGPT subscription, headless | gpt-6.1-soleffort default | 2026-10-24 | 75%6 of 8 | 91% | $0 | 0.6 | 80 void |
Fully correct by return
| Return | Antigravity CLI, gemini-4-argon-medium | Claude Code, claude-opus-5-5 | Codex CLI, gpt-6.1-sol |
|---|---|---|---|
| S corporation Schedule K-1 with royalties (Form 1120-S) | 1/1 | 1/1 | 1/1 |
| early IRA distribution with the 10% additional tax (Form 1099-R code 1) | 1/1 | 0/1 | 1/1 |
| earned income credit | 1/1 | 1/1 | 0/1 |
| incentive stock option exercise (alternative minimum tax) | 1/1 | 1/1 | 1/1 |
| partnership Schedule K-1 (Form 1065) | 1/1 | 0/1 | 1/1 |
| partnership guaranteed payments with self-employment tax | 1/1 | 1/1 | 0/1 |
| rental real estate with depreciation (Schedule E) | 1/1 | 1/1 | 1/1 |
| securities sales on Form 1099-B (Schedule D) | 1/1 | 1/1 | 1/1 |
Run by Terminal Tax. Fully correct: every scored Form 1040 line and the filing status match the answer key exactly. Line accuracy is the share of lines with an amount in the key or the return that match; bottom-line error is the median dollar gap in the refund or amount owed. No return was sent anywhere.
Method
Each agent runs through its vendor's official command-line product as a fresh install, with no memory, custom instructions or prior setup, and is named by product, model and dates. No consumer chat app was scripted.
Version 0.3.0 has two tiers. Starter: 11 fictional taxpayers, from simple single filers to a married couple with children, a pension, a head of household and a self-employed designer. Hard: 10 returns with rental depreciation, partnership and S corporation Schedule K-1s, securities sales and a capital loss carryover, an incentive stock option exercise that triggers the alternative minimum tax, the earned income credit, an early IRA distribution and excess Social Security withholding.
Starter-tier answer keys were worked by hand from the 2025 Form 1040 instructions, Tax Table and worksheets, and each shows its work. Hard-tier keys were computed by Terminal Tax's tax engine, which its test suite compares with IRS Direct File's open-source calculations, PolicyEngine and the IRS's published e-file test scenarios where those cover a situation.
Five tasks (three starter, two hard) are held out and published only as hashes, so a configuration can be checked on tasks no model has seen.
A run that measured nothing about the agent, such as one whose command-line tool returned nothing because of a plan quota, is void and listed apart.
These are initial results: one run of each public task per agent (8 starter, 8 hard), on the dates shown. With this few tasks, a difference between rows is not evidence that one agent is better at taxes in general.
Run it or submit an agent
Anyone can run 1040Bench on their own computer with the open harness and scorer. To put an agent on this board, send its configuration; as capacity allows, Terminal Tax runs it on the public and held-out tasks with the same harness, publishes its public-task runs at once and its held-out ones when those tasks retire, and marks the row verified.
Terminal Tax is not affiliated with, sponsored by, or endorsed by any company whose agent appears here. Product names identify the agents tested and belong to their owners.