AI Tax Agent Test: Best Model Passes Just 10.8%

Key takeaways

Surge AI's new DAYJOB: Tax benchmark, published October 27, 2026, gives AI agents 50 professional tax assignments built from 1,789 files.
The best AI tax agent, Claude Sonnet 5.5, passed only 10.8% of attempts at a 95% rubric threshold; three models scored 0%.
The failures were often about judgment, such as missing a home-sale deadline the client never asked about.
The tasks include small-business tax, partnerships and property and investments, so the results are directly relevant to owners, landlords and home sellers.
Use AI to organize records and draft questions, and keep a qualified tax professional in charge of the final return and advice.

Can an AI tax agent do a tax professional's job from start to finish? Not yet, according to a new benchmark. On October 27, 2026, Surge AI published DAYJOB: Tax, which hands AI agents 50 realistic tax assignments with the messy records real clients send, and the best model passed only 10.8% of its attempts (Surge AI). The write-up reached 231 points and 150 comments on Hacker News.

If you have been wondering whether ChatGPT, Claude or another agent can file your business return, this is the most detailed public answer so far, and it includes a home-sale example every property owner should read.

What the DAYJOB: Tax benchmark tests

Fifty assignments, 1,789 files

The benchmark covers eight areas of tax practice: individual and family tax, small-business tax, partnerships, property and investments, estates, gifts and trusts, international tax, corporate transactions, and tax review and disputes. Agents get a client request plus the supporting records and must produce a return, workpapers, a technical review or client advice (Surge AI).

The files are what a real practice sees: 1,487 PDFs, 117 Excel workbooks and 38 Word documents, plus emails, images and CSVs, with a median of 30 files per assignment and as many as 123. They include draft returns, K-1s, W-2s, IRS notices, general ledgers and payroll records (Surge AI).

Built and graded by tax experts

Tax professionals turned real workflows into the tasks, and each one went through three layers of expert review. Contributors include a tax attorney with more than 10 years at PwC, Deloitte and Exxon, and a specialist in real-estate partnerships and individual tax with Deloitte and EY experience (Surge AI). The Hugging Face dataset card adds that each task has a median of 30.5 grading criteria and that agents had no web browsing during the test (Hugging Face).

A tough pass mark

An attempt only passes if it meets at least 95% of the task's criteria, which cover calculations, supporting evidence, consequential issues and the finished deliverable (Surge AI). That is a high bar on purpose: a return that is 90% right can still be wrong.

The results: no AI tax agent came close

Model	Pass rate	Cost per task
Claude Sonnet 5.5	10.8%	$10.71
Claude Fable 5.1	9.6%	$15.37
Claude Opus 5.5	9.2%	$12.96
Muse Wonder 1.0	6.0%	$5.60
Grok 4.7	5.0%	$4.57
GPT-6 Astra	4.8%	$13.88
Gemini 4 Argon	3.2%	$2.96

Pass rates come from Surge AI's leaderboard; the cost column comes from the Fireworks Specialized Intelligence Index, which lists 16 models. Kimi K3, Mistral Large 4 and Qwen 3.8 Max scored 0%. Fireworks notes that the best open model, DeepSeek V4.1 Flash, reached 1.2% at $0.49 per task.

Surge sums it up in one line: "Agents haven't yet passed the audit" (Surge AI).

Where AI tax agents go wrong

The condo sale deadline

The example that matters most for home sellers involves a client named Anna. She moved out of her condo in August 2024 and asked whether closing in October 2026 or January 2027 would be better for taxes. The agent correctly said both dates met the residence test for a home-sale exclusion worth up to $250,000 of gains. What it missed: under the two-out-of-five-years rule, Anna would stop qualifying around the end of August 2027, so a delayed sale could cost her the break. The model Surge calls Astra never flagged that deadline in five attempts, while Opus caught it all five times (Surge AI).

An assumption nobody checked

In another task, a separating couple asked how to divide a Roth IRA and a 401(k). Opus 5.5 built its advice on the assumption that they were unmarried in all five attempts, even though their files showed they had lived in Texas and held a ceremony in Colorado, which could matter for an informal marriage. Astra said it needed to confirm their marital status first in four of five attempts (Surge AI). No single model got every kind of judgment right.

Small-business records that tell a different story

Two tasks look a lot like small-business life. One includes a photo of a handwritten receipt for goods bought for resale that must be matched to spending records and inventory. Another involves a silo-services company that thinks it only operates in Alabama, while its dispatch log shows crews in five states, which affects its multistate tax obligations (Surge AI).

What we did not verify or test

We did not run the benchmark ourselves; only two of the 50 tasks are public (Hugging Face). The scores and cost figures are as published by Surge AI and Fireworks. The tasks use US forms and IRS rules, so they say nothing direct about Canadian or other tax systems. Agents were also tested without web access, so tools that look up current rules may perform differently.

What this means for small businesses

The takeaway is not "never use AI for taxes." It is "use AI for the grunt work and keep a professional on the judgment calls." Here is a practical split:

Let AI organize the records. Turn receipt photos into entries with receipt photos to bookkeeping entries and clean up your ledger with categorize transactions.
Let AI chase what's missing. Use the missing-documents chaser so your accountant gets a complete file the first time.
Ask for a draft, not a filing. If you use an agent on a return, have it produce a review-ready draft with every exception flagged, as in let AI agents prepare individual tax returns while you review.
Make it list the questions you didn't ask. Anna's case shows the costly mistakes are often omissions. Ask any agent for deadlines, assumptions and facts it couldn't confirm, and tighten that instruction with the prompt grader or the AI prompt generator.
Selling property? Get a human check. Home-sale exclusions, rentals and partnerships were exactly where judgment mattered in the test.
Count the real saving. At $10 or more per attempt for the top models, and a pass rate near 1 in 10, work out the value of your time with the AI time savings calculator before replacing a preparer.

Frequently asked questions

Can AI do my taxes?

It can help, but on DAYJOB: Tax the best AI agent passed only 10.8% of professional tax assignments at a 95% threshold, and three models scored 0% (Surge AI). Use it to prepare and organize, and have a professional review the result.

What is the best AI for taxes?

On this benchmark, Claude Sonnet 5.5 led at 10.8%, followed by Claude Fable 5.1 at 9.6% and Claude Opus 5.5 at 9.2% (Surge AI). Sonnet 5.5 cost about $10.71 per task in Fireworks' run (Fireworks).

Can an AI agent file a tax return on its own?

The benchmark suggests it should not. Agents had to deliver returns, workpapers and advice from messy files, and Surge concludes that reliably carrying out real tax workflows "remain[s] out of reach" for frontier models (Surge AI).

Can AI do my taxes in Canada?

DAYJOB: Tax uses US documents such as W-2s, K-1s and IRS notices (Surge AI), so it does not measure Canadian returns. Treat its findings on judgment and missed deadlines as a general warning rather than a Canadian test.

Sources

How we report: every fact, number and quote above comes from the linked sources, which we read before publishing. Product details can change quickly during previews, so check the vendor page before you buy. Spotted an error? Tell us via the link on our home page.

AI tax agent benchmark: the best model passed just 10.8% of real tax assignments