Applied AI leaderboard
How well agents built on Tangle do taxes, build apps and handle legal work. Public benchmarks are scored by their owner's evaluator.
As of 2026-11-02: tax, 15 of 46 returns with every line exact. Coding, GPT-5.6 Terra solves 19% of Typesense apps on a real Typesense 30.2. Legal, Claude Sonnet 4.5 scores 98.6 of 100 across 21 personas.
Tax returns
Share of TaxCalcBench returns with every line exact. Higher is better.
Coding
Share of VerticalBench app tasks solved, per board. Higher is better.
Cost per solved task
Spend divided by tasks solved, per board. Lower is better.
Tax: TaxCalcBench
TaxCalcBench (Column Tax), 51 synthetic TY2024 federal returns; our 46 run records at c20ceb4d11.
| # | system | lines exact | returns exact | returns |
|---|---|---|---|---|
| 1 | Gemini 2.5 Prosingle shot, published by Column Tax | 81% | 32% | 51 |
| 2 | Tangle tax agentdeepseek-v4-pro | 78% | 33% | 46 |
| 3 | Claude Opus 4single shot, published by Column Tax | 78% | 27% | 51 |
| 4 | Gemini 2.5 Flashsingle shot, published by Column Tax | 78% | 26% | 51 |
| 5 | Claude Sonnet 4single shot, published by Column Tax | 77% | 23% | 51 |
Returns exact: every evaluated line matches the correct Form 1040. Lines exact: share of lines that match. Published rows are Column Tax's single-shot results, as cited in the write-up; ours ran the production tax agent, median 1.2 min per return.
Every scored return, weakest first
| return | lines exact | return | tool calls |
|---|---|---|---|
| single-senior-blind-over-65 | 0% | no | 8 |
| single-w2-schedule-c-qbi-dependent-estimated-tax | 0% | no | 4 |
| mfj-multiple-w2-schedule-c-qbi-income | 53% | no | 0 |
| single-w2-schedule-c-qbi-loss-carryforward | 53% | no | 0 |
| mfj-spouse-dependent-schedule-c-w2-student-loan-interest | 58% | no | 0 |
| hoh-multiple-w2-box12-codes | 63% | no | 3 |
| hoh-schedule-b-ssa1099-unemployment | 63% | no | 1 |
| mfj-multiple-schedule-c-loss-multi-home-office | 63% | no | 0 |
| mfj-w2-six-dependents-student-over-17 | 63% | no | 0 |
| single-1099k-personal-payments | 63% | no | 1 |
| single-w2-schedulec-1099b-capital-loss-carryover | 63% | no | 3 |
| single-w2-tips-long-employer-name | 63% | no | 0 |
| hoh-w2-dependent-educator-expenses-unemployment | 68% | no | 3 |
| mfj-schedule-2-multiple-w2-excess-social-security-tax | 68% | no | 1 |
| mfj-w2-box12-codes-a-b-1099int-schedulec | 68% | no | 3 |
| mfj-w2-unemployment-1099g-repayments | 68% | no | 1 |
| single-eic-non-dependent-child | 68% | no | 0 |
| single-multiple-w2-schedule-c-qbi-losses | 74% | no | 3 |
| mfj-multiple-1099int-schedule-b-w2 | 79% | within $5 | 1 |
| mfj-multiple-w2s-excess-social-security-tax | 79% | no | 2 |
| single-multiple-w2-state-records | 79% | within $5 | 1 |
| single-retirement-1099r-alaska-dividend | 79% | within $5 | 0 |
| single-w2-box12-code-a-b-alaska | 79% | no | 2 |
| single-w2-multiple-1099int-dividend | 79% | within $5 | 0 |
| single-w2-multiple-1099int-federal-withholding | 79% | within $5 | 1 |
| mfj-multiple-w2-schedule-c-qbi-max-threshold | 84% | no | 3 |
| single-w2-multiple-1099int-withholding-schedule-b | 84% | within $5 | 6 |
| mfj-schedule-c-1099-misc-nec-k-ssa-1099-int-g | 89% | no | 1 |
| single-w2-healthcare-marketplace-1095a | 89% | no | 3 |
| hoh-w2-dependent-care-credit-carryover-no-qualifying-person | 95% | no | 7 |
| single-w2-student-american-opportunity-credit | 95% | within $5 | 1 |
| hoh-w2-1099g-unemployment-schedulec-loss | 100% | exact | 0 |
| mfj-both-blind-nontaxable-social-security | 100% | exact | 0 |
| mfj-dual-w2-over-65 | 100% | exact | 1 |
| mfj-w2-box12-codes | 100% | exact | 2 |
| mfj-w2-multiple-1099g-unemployment-income | 100% | exact | 3 |
| mfj-w2s-dependent-estimated-tax-qbi-loss-carryforward | 100% | exact | 5 |
| single-1099b-long-term-capital-gains-schedule-d | 100% | exact | 0 |
| single-1099int-interest-income-schedule-b | 100% | exact | 0 |
| single-multiple-w2-excess-social-security-tax | 100% | exact | 2 |
| single-multiple-w2-excess-social-security-tax-same-ein | 100% | exact | 1 |
| single-w2-balance-due-no-state-income-tax | 100% | exact | 0 |
| single-w2-direct-debit-payment | 100% | exact | 2 |
| single-w2-minimal-wages-alaska | 100% | exact | 0 |
| single-w2-retirement-sick-pay-social-security-tip | 100% | exact | 0 |
| single-w2-unemployment-1099g | 100% | exact | 4 |
Coding: VerticalBench boards
Real apps built against a pinned backend and graded in a real browser by blueprint-agent; a board is one campaign's record.
Typesense apps on a real Typesense 30.2
2026-09-27, 13 tasks × 2
| # | profile | solved | interval | per solved | attempts |
|---|---|---|---|---|---|
| 1 |  GPT-5.6 TerraOpenCode, no web | 19% | 0% – 38% | $1.09 | 26 |
| 2 | GPT-5.6 LunaOpenCode, no web | 3.8% | 0% – 12% | $0.59 | 26 |
Web apps built on Typesense 30.2, graded in a real browser against the official typesense/typesense:30.2 image. Pass at App Grade B. Record sha256:93f21.
Meilisearch apps on a real Meilisearch 1.54.0
2026-09-27, 13 tasks × 2
| # | profile | solved | interval | per solved | attempts |
|---|---|---|---|---|---|
| 1 | GPT-5.6 LunaOpenCode, no web | 12% | 0% – 31% | $0.21 | 26 |
| 2 | GPT-5.6 TerraOpenCode, no web | 7.7% | 0% – 19% | $3.00 | 26 |
Web apps built on Meilisearch 1.54.0, graded in a real browser against the official getmeili/meilisearch:v1.54.0 image. Pass at App Grade B. Record sha256:c8537.
Legal: persona matrix
Our own eval: 21 synthetic client personas through the production legal agent, scored by a verifier; record at 61f4676566.
| # | model | mean composite | lowest persona | personas |
|---|---|---|---|---|
| 1 | Claude Sonnet 4.5sonnet | 99% | 92 | 21 |
| 2 | Claude Haiku 4.5haiku | 98% | 87 | 21 |
Verifier score out of 100, averaged over 21 synthetic client personas. Lowest persona: the model's weakest single score.
Every persona, lowest score first
| persona | sonnet | haiku |
|---|---|---|
| existing-business-audit | 100 | 87 |
| crypto-exchange-licensing | 96 | 88 |
| saas-company-international | 92 | 94 |
| delaware-llc-multi-state | 96 | 93 |
| nuclear-startup-nrc | 100 | 94 |
| restaurant-formation | 94 | 100 |
| cannabis-dispensary | 96 | 99 |
| construction-contractor | 100 | 99 |
| healthcare-practice | 99 | 99 |
| noncompete-ca-vs-tx | 99 | 100 |
| nonprofit-formation | 99 | 99 |
| source-code-escrow-no-decompile | 99 | 99 |
| cayman-fund-side-letter | 100 | 100 |
| convertible-note-disguised-pref | 100 | 100 |
| earnout-anti-frustration-gap | 100 | 100 |
| founder-83b-missed | 100 | 100 |
| international-distribution-fcpa | 100 | 100 |
| nda-mutual-actually-oneway | 100 | 100 |
| saas-reseller-mfn-trap | 100 | 100 |
| safe-asymmetric-conversion | 100 | 100 |
| working-capital-peg-stale | 100 | 100 |
Health
The health agent's eval runs 12 synthetic patient journeys over 26 sessions through the real product: signup, the audited vault, record import and the chat. It proves the journeys complete; a quality number against a public benchmark is not published yet, so no row appears here.
Research
Discovery scores each research run with the evaluator its play names. Those records are not public yet. When one is, it goes here with its source, the same way.