Open Source Tax Engine outperforming GPT sol and Fable 5

I'll pull the article and HN thread so the paragraphs stick to verifiable facts only.Invaro’s **OpenTax** engine posted the highest score yet on **TaxCalcBench TY25**: **Claude Sonnet** plus the **OpenTax MCP** produced exact returns on **48 of 50** full U.S. filings (**96%**), versus **6%** for the same Sonnet with no engine. On the same strict leaderboard—every line must match the reference return—**GPT-5.6 Sol with web search** scored **58%**, **GPT-5.5 with web search** **54%**, and **Claude Fable 5 with web search** **34%**. Across **820** scored lines, the OpenTax run hit **98.2%** exact; the two full-return misses were traced by Invaro to reference worksheets that disagree with their own inputs, with a bug filed upstream on the **column-tax** benchmark.

OpenTax is not another model. It is a **deterministic computation layer**: fixed facts and fixed rules yield the same return every time, using exact arithmetic rather than sampling. Rules carry **26 U.S.C.** citations, effective windows, and verbatim statute excerpts; each computation can emit a **Merkle-rooted proof** of facts, rules, and arithmetic for offline re-verification. A compiler turns **W-2** and **1099** boxes plus birth dates into engine facts without judgment calls, and inputs outside the corpus are refused with a named reason. Coverage spans individual, corporate, and fiduciary forms (**1040**, **1120**, **1041**) and **29** states under the same citation discipline. Agents connect via a hosted MCP endpoint (`https://opentax.invaro.ai/mcp`) or a local package (`@invaro/opentax`); the stack is **AGPL-3.0** (plus commercial) at `github.com/Invaro/opentax-engine`.

For builders, the result reframes tax work as **tool-amplified reasoning**, not raw LLM accuracy. A model that alone clears **6%** of full returns jumps **16×** once line math and statute lookup leave the probabilistic path. That split matters for agent design: the model can still parse docs and draft client language, while refunds, safe-harbor estimates, **§ 6654** annualized installments, **QBI** wage limits, and cliff sweeps stay in a testable engine. Review workflows get a second, cited recomputation against draft software; audit trails get re-derivable proofs instead of chat logs.

Competitive context is the gap between **closed model+search** setups and a **public rule engine**. Solo Sonnet and top GPT/Claude runs with web search all sit well below the OpenTax-assisted score on TaxCalcBench’s pass@1, cold-agent protocol as published for July 2026. Invaro also claims it is the only open-source tax engine on the market; whether that exclusivity holds, the open corpus and the benchmark harness that produced the **96%** figure are inspectable—unlike proprietary prep stacks whose line logic stays opaque. Language models remain probabilistic; tax law is not, and the leaderboard numbers show how hard full-return exactness is without a dedicated calculator.

Practical next steps: wire the MCP into Claude, ChatGPT, Cursor, or a custom agent and re-run a known client file line-for-line against your current software, keeping the proof artifacts with the workpapers. Watch TaxCalcBench for fixes to the two contested reference cases and whether other engines publish comparable strict scores. Track state-line growth past **29** jurisdictions and whether AGPL versus commercial licensing fits your distribution model. For CI, the CLI’s expect-style checks (`opentax check --expect`) are the natural place to lock regression coverage as rules and forms change.