GPT-6 Astra Outpaces Opus, Argon Fails CSV Test

Three generalist models were tested on a wireframe, a project report, a data analysis, and a go-to-market strategy. Two completed all four exercises, while one stumbled twice on a CSV file. The discrepancies are also evident in the GTM plans and API list prices as of October 22.

Verdict, scoring, and API list prices as of October 22

The scoring system distinguished three outcomes: 10/10, 9/10, and NC for not completed or not comparable. GPT-6 Astra was narrowly selected, thanks to a slight advantage in the GTM contingency plans. Claude Opus completed all four tasks and matched three, but its GTM fallback plan needed adjustment. Argon's work was useful where it was complete, but the failure to handle the CSV prevented a fair comparison; since NC is not zero, no overall total was assigned to Argon. Regarding verified API prices as of October 22, Astra listed $50 per million tokens for input and output, compared to $10 for Opus. Until November, Argon listed a price of $10, then increased to $20 starting in December. Tasks performed in the browser were not subject to cost evaluation, and the prices indicated do not correspond to the actual execution costs. Astra and Opus used the Max setting, while Argon was on High, without standardizing the calculation between models.

BrightPath Data Analysis: two successes and one CSV failure

The data analysis exercise required allocating the next $1,000 after comparing the last four weeks with the first four, including a before-and-after table, three insights, two actions, and one limit, in under 350 words. Astra and Opus identified that the search conversion increased from 10% to 11%, and social media conversion rose from 4% to 5%, but the overall rate declined as the share of sessions from social media increased from 16.7% to 75%. Both prioritized a monitored search test, without guaranteeing marginal returns, relying on recent activation costs of $19.70 for search versus $33.33 for social media. Argon, on the other hand, submitted an empty input CSV and returned an unpopulated framework, despite a second attempt with the same files; both responses were from the High setting. A partial output mentioned a 6.5% conversion for sign-ups and $25.44 cost per activation, with a cost increase of 23.3%.

GTM ShelfCheck: quantitative objectives met, divergences on fallback

The strategy aimed for 12 paying stores by day 42, with 14-day trials and a capacity for six weekly integrations. All models targeted independent stores, excluded unsupported integrations, and used warm contacts. The plans adhered to an effective acquisition cap of $3,000, equating to twelve customers at $250 CAC. At best, 24 trials could start by day 28; reaching twelve payers assumed a conversion rate of at least 50%. Astra's plan, budgeted at $1,200, explicitly raised the required conversion forecast after starting lower than planned. Opus's plan, at $1,250, remained solid, but its fallback to "at least 55%" was too broad in a scenario with nine early starts, where 21 remaining trials imply 57.1%. Argon's plan, at $1,200, did not compensate for lost capacity by merely loading the following weeks and exceeded the 450-word constraint. Web research activity was observed on Argon's side despite the instruction to use only the brief, but the final material facts remained within its scope.

Wireframing ShiftBridge: three standalone HTMLs, execution detail

The expected dashboard was to be delivered in standalone HTML, focused on the coordinator's work, with included interactions and an explanation of fewer than 60 words. The requested desktop display was 1440 × 900, with operational mobile use and control of uncovered shift filters, the attribution panel, closing, and the Escape key with focus return. All three models produced functional HTML, with compliant filters and panels, and successful mobile validation at 390 × 844. Astra and Opus provided interfaces adhering to the 1440 × 900 resolution. Argon generated a layout reaching 987 pixels in width and requested a manual recording of its chat code without modification; its interface was the most rudimentary, and some buttons were non-functional.

The synthesis reconciles the divergent data from the PDF file

The synthesis was to reconcile a file including a two-page PDF with discrepancies, for a report dated October 22, 2026, and limited to 350 words, including a risk-action table, progress status, recommendation on the pilot for November 2, budget position, and decisions to be made before the checkpoint on October 30. All three models relied on an acceptance result of 18/24 and two open priority defects, and converged on a forecast of $53,600 highlighting a gap of $5,600. None considered scheduled reviews as approvals nor validated the pilot in advance; information on owners, actions, and dates was usable. The formats varied: Astra provided a concise document, Argon submitted a report organized according to usage, and Opus mentioned an unverified privacy review falling on the decision day. All deliverables were rated as satisfactory.

Testing framework: attempts, settings, and non-standardization

The tests were conducted with one attempt per task, accompanied by a specific retry for Argon's CSV task. Argon operated in High mode, its highest thinking level, while Astra and Opus used Max. The selected parameters did not standardize either computing power or tools, with the exercise being a practical comparison rather than a benchmark protocol with repeated trials.