Basis Completes Tax Workbook 2× Faster With GPT-6 Astra

Basis reports that GPT-6 Astra completed a complicated 50-tab tax workbook in half the time of GPT-5.6 Sol. The accounting-agent company also says Astra raised its internal evaluation score by about 20%, understood user intent better, and used a more direct path through the task.

That is a useful production signal, but it is not proof of a 50% cost saving. OpenAI and Basis disclose no elapsed times, repeated trials, token counts, tool usage, accuracy results, reviewer effort, or dollar bill. The announcement changes no OpenAI API rate.

What Basis tested

Basis compared Astra with GPT-5.6 Sol on one described workload: accurately and reliably completing a workbook containing 50 tabs. Co-founder Mitch Troyanovsky says Astra finished in half the time.

Published detail	What it supports	What remains unknown
Complicated 50-tab tax workbook	A concrete long-horizon accounting task	Workbook contents, formulas, source documents, and expected answers
Half the completion time	Lower wall-clock time in Basis’s reported comparison	Minutes, run count, variance, concurrency, and service tier
About 20% higher internal evaluation score	Better performance under Basis’s own rubric	Baseline score, sample size, weighting, failures, and statistical uncertainty
Adaptive reasoning with cache retained	A plausible route to fewer unnecessary tokens and lower latency	Reasoning settings, cache-hit rate, token mix, and billed usage

Basis says its evaluations inspect both the agent’s work process and final answer, including template compliance, use of primary sources for tax questions, asking questions, flagging assumptions, and self-checking. Those are relevant controls, but the underlying test set, grader outputs, and traces are not public.

Pricing impact: faster does not automatically mean cheaper

The live table, chart, and calculator above use current rates from our canonical pricing dataset. Astra costs $10 input, $1 cached input, $12.50 cache write, and $50 output per 1M short-context tokens. GPT-5.6 Sol costs $4, $0.40, $5, and $20. Astra is therefore 2.5× Sol in every published Standard short- and long-context token category.

With the same token mix, Astra must use 40% or less of Sol’s billed token volume to match Sol’s model-token spend. A faster run can still cost more if it uses expensive reasoning output, large uncached inputs, extra tool calls, or repeated validation passes. Half the wall-clock time is not proof of lower model spend.

The reverse is also possible: better early decisions, fewer corrections, and stronger cache reuse can reduce calls and reviewer time enough to outweigh a higher token rate. Basis says Astra uses tokens more efficiently, but gives no ledger that lets buyers calculate the crossover.

Long workbooks may also carry costs beyond model tokens: file processing, hosted containers, retrieval, storage, orchestration, audit logging, and accountant review. Use the token calculator for the model layer, then add those workflow costs separately. The maintained OpenAI pricing page explains service tiers and long-context rules.

Who benefits—and who may overspend

Tax and accounting teams benefit when a workbook has many linked tabs, ambiguous instructions, varied source documents, and expensive review cycles. Astra can earn its premium if it reduces failed runs, formula repairs, unsupported tax positions, and senior-accountant review time.

Standardized workbooks, fixed extraction, and repeated mappings may not need the premium model. Start those jobs on GPT-6 Sol, GPT-6 Luna, or another lower-cost route and escalate only when the result fails a defined acceptance gate. For an independent comparison, review the Anthropic pricing page and our AI API pricing comparison.

Teams building a custom accounting pipeline can also benchmark a managed open-model control through Novita. Confirm the exact model, spreadsheet tooling, data-retention terms, residency, and live rate before using client tax data.

Affiliate disclosure: AI Pricing Guru may earn a commission from the sponsored Novita link at no extra cost to you. It does not affect this analysis.

What accounting teams should test

1. Freeze representative workbooks, source files, instructions, expected formulas, and approval criteria.
2. Run Astra and a cheaper control with the same tools, context, service tier, retries, and time limits.
3. Grade numerical accuracy, formulas, citations, assumptions, template compliance, and unsupported changes.
4. Record elapsed time, input and output tokens, cache reads and writes, tool charges, failures, corrections, and reviewer minutes.
5. Route to Astra only where total cost per reviewer-approved workbook—or risk reduction—beats the control.

Keep a qualified accountant in the approval path. Faster spreadsheet completion does not establish tax correctness, filing eligibility, or professional sign-off.

Labs coverage decision

This story does not add a score to the AI Pricing Guru Labs leaderboard. Our maintained route is not authorized for Astra, and the fixed text suite cannot reproduce spreadsheet tools or qualified-accountant review.

A fair replay needs authorized Astra access, rights-cleared workbooks and sources, pinned model, reasoning, cache, and tool settings, raw traces, complete token and tool cost records, repeated runs, output diffs, reviewer-time measurement, and blinded qualified-accountant grading.

The comparison also uses a private tax workflow rather than our deterministic text suite. See the related Parallel Astra case study for another example where lower elapsed time did not establish a general API-cost reduction.

Bottom line

Basis gives accounting teams a credible reason to test GPT-6 Astra on difficult, long-running workbooks. Its reported half-time completion and internal evaluation gain do not establish a universal speedup, a lower API bill, or tax accuracy.

Buy on cost per approved workbook. Preserve the cheaper route for routine jobs, escalate hard cases to Astra, and require complete usage and review evidence before expanding deployment.

Sources: OpenAI’s official Basis and GPT-6 Astra customer story, Basis’s accounting-agent overview, OpenAI’s GPT-6 Astra model documentation, and API pricing. Claims, model availability, and pricing rules checked September 28, 2026 at 21:17 UTC.