Quick Answer

In our May 2026 testing, Claude Sonnet 4.6 made fewer errors than Gemini 2.0 Pro on multi-step tax logic, especially deduction phase-outs and cross-document reconciliation. The same scratchpad fix reduces errors on both. Neither replaces a licensed tax professional.

Last Updated: August 2026 · Cole Bridges Research Lab

Verified Snapshot: Tested on Claude Sonnet 4.6 and Gemini 2.0 Pro, May 2026, Reviewed August 2026 (current default: Sonnet 5)

All test results in this article used identical prompts on both models via API. Tests were conducted by the Cole Bridges Research Lab. Neither AI is a substitute for a licensed tax or legal professional.

Quick Answer

In our May 2026 test run, Claude Sonnet 4.6 produced fewer errors on multi-step deduction logic and cross-document reconciliation. Gemini 2.0 Pro is competitive on single-document summarization. For OBBBA-related changes, both models struggle with phase-out threshold math, use the scratchpad prompt fix for any numerical analysis.

Model status: checked August 6, 2026

Anthropic's current lineup is Claude Haiku 4.5, Claude Sonnet 5 (released June 30, 2026), Claude Opus 5 (released July 24, 2026), and Claude Fable 5, with Claude Mythos 5 restricted to approved partners. Sonnet 5 is the default model on the Free and Pro plans, Opus 5 is the default on Max and the strongest model available on Pro. Source: Anthropic.

The test runs on this page were performed on Claude Sonnet 4.6, which Sonnet 5 replaced as the default model on June 30, 2026. We publish these results as a dated snapshot rather than restating them as current-model behavior. Where a finding has not been re-tested on Sonnet 5 or Opus 5, treat it as history, not as a live bug report.

What We Tested

We gave both Claude Sonnet 4.6 and Gemini 2.0 Pro the same five tax document analysis tasks using identical system prompts and user prompts. All tests were conducted in May 2026 via direct API access.

  • Single W-2 summarization with bracket identification
  • Schedule C self-employment deduction extraction
  • Cross-document reconciliation (W-2 + 1099 + Schedule C)
  • Deduction phase-out threshold calculation
  • Multi-year comparison summary (2024 vs 2025 returns)

Head-to-Head Results

TaskClaude Sonnet 4.6Gemini 2.0 ProWinner
Single W-2 summarization0 errors0 errorsTie
Schedule C deduction extraction1 minor error3 errorsClaude
Cross-document reconciliation2 errors5 errorsClaude
Deduction phase-out calculation3 errors4 errorsClaude (marginal)
Multi-year comparison1 error2 errorsClaude

Where Both Models Struggle: Phase-Out Math

Both Claude Sonnet 4.6 and Gemini 2.0 Pro produced their highest error rates on deduction phase-out threshold calculations, the type of multi-step math required for OBBBA-impacted deduction analysis. This is consistent with the pre-calculation failure pattern documented in our Pre-Calculation Fix guide.

The fix is the same for both models: use a <scratchpad> tag to force step-by-step calculation before the model writes any summary. In our testing, this sharply reduced phase-out errors on both Claude and Gemini.

Important Disclaimer

Neither Claude Sonnet 4.6 nor Gemini 2.0 Pro should be used for actual tax filing decisions without verification by a licensed tax professional. These results reflect document analysis capability only, not legal or financial advice accuracy. Always verify AI-generated tax analysis with a CPA.

Best Prompt Structure for Tax Document Analysis

<task>Analyze the attached tax document and extract the following.</task>

<extract>
1. Gross income
2. All deductions listed
3. Taxable income after deductions
4. Any phase-out thresholds that apply
</extract>

<scratchpad>
Show all calculations step by step before writing the summary.
Do not skip arithmetic. Flag any number you are uncertain about.
</scratchpad>

<output>
Structured summary only after scratchpad is complete.
Flag any item that requires CPA verification.
</output>
Anthropic shipped Claude Sonnet 5 on June 30, 2026 and made it the default model for Free and Pro accounts, replacing Sonnet 4.6 across its consumer and developer products the same day.
Anthropic — Introducing Claude Sonnet 5

Model defaults move every few weeks now. Any technique you rely on needs a tested-on date attached, which is why every page here carries one.

Stanford HAI's 2026 AI Index puts organizational AI adoption at 88 percent, and finds generative AI reached 53 percent population adoption within three years — faster uptake than either the personal computer or the internet.
Stanford HAI — 2026 AI Index Report

Adoption is no longer the hard part. Getting reliable output is. That gap is what the testing on this site is built to close.