Client-Ready Decks with GPT-6.1 Terra: Evaluation

7 min read

Accuracy vs time and cost

GPT-6.1 Terra at three reasoning efforts on the 8 new-template requests, with GPT-6.1 Sol medium as the reference point. Colour is the model, shape the reasoning effort. Every setup ran one deck per request, so the points carry no interval.

terra_clean_vs_time.png terra_clean_vs_cost.png

GPT-6.1 Terra is as accurate as GPT-6.1 Sol on templates it has never seen, and cheaper. At medium effort 96% of slides are client-ready against 98% for Sol; at high effort 100%. Terra costs USD 0.44 to 0.47 per deck against USD 0.77 for Sol, and is about a minute faster per deck.

terra_clean_new_vs_tuned.png

On the tuned templates, where the GPT-5.1 optimised setup was developed, that setup stays ahead (97% client-ready slides); Terra and Sol run the standard skill there and reach 78 to 85%.

Key facts

  • What we test: how well Unique AI Conduct builds client-ready PowerPoint decks in a bank's own template with GPT-6.1 Terra, at low, medium and high reasoning effort, with GPT-6.1 Sol medium as the reference. GPT-5.1 high with the optimised setup ran on the tuned templates as the current baseline.
  • Requests: the same 13 requests as the <![CDATA[GPT-5.1 evaluation]]>: 8 on 4 company templates the setup was not developed on (new templates) and 5 on 4 templates it was developed on (tuned templates). 57 decks, graded blind slide by slide (run of 12 November 2026 on the QA environment).
  • How we test: each request runs end to end through Conduct on the Codex harness, as a relationship manager (RM) would send it. A review skill grades every deck blind against a 15-item checklist. The score is client-ready slides: the share of the requested slides that need no fix before the deck goes to the client.
  • Result on new templates: Terra low 96%, medium 96%, high 100% client-ready slides; Sol medium 98%. No setup produced an unusable deck.
  • Time and cost: Terra takes 8 to 9 minutes and USD 0.36 to 0.47 per deck on a new template; Sol takes 10 minutes and USD 0.77. Times were measured on QA with up to eight decks running at once, so compare them only within this page.
  • Recommendation: GPT-6.1 Terra at medium or high for PowerPoint decks. High gave the best result for USD 0.03 more per deck; low works but leaves more fixes.
  • Consistency check: GPT-6.1 Sol medium reaches 98% on the new templates here, the same as in the GPT-5.1 evaluation of 26 October, so the requests and the grading give the same answer across runs.
  • Caveat: one deck per request and setup, so a difference of one deck is within noise and no interval is given.

What to expect

Per deck on a new template, the bank's own template before it has been reviewed with us:

Per deck, new templateGPT-6.1 Terra lowGPT-6.1 Terra mediumGPT-6.1 Terra highGPT-6.1 Sol medium (reference)
Client-ready slides96%96%100%98%
Decks with no fix needed75%88%100%88%
Unusable decks0%0%0%0%
Minutes per deck8.08.89.09.8
Minutes per requested slide1.551.701.771.95
Minutes per client-ready slide1.641.801.772.00
Est. model cost per deckUSD 0.36USD 0.44USD 0.47USD 0.77
Model calls per deck52575950
Input tokens per deck (incl. cached)3.3 M4.0 M4.5 M4.0 M

The fixes that remain on new templates are a design left partly filled (cards with empty bodies, half a slide left empty after a card was removed) and, once, a chart's values repeated as text beside the chart.

Results

Client-ready slides and unusable decks: mean over each request's deck, then over requests, so each request weighs the same. Minutes are wall-clock per deck on QA; cost is the estimated model cost per deck at list prices from the platform's usage ledger. All setups run on the Codex harness. GPT-6.1 models use the standard PowerPoint skill; GPT-5.1 uses the optimised setup it has on master (build tools, working rules and output checks).

New templates

8 requests, 5.4 requested slides on average, one deck per request and setup:

Model and setupReasoning effortClient-ready slidesDecks with no fix neededUnusable decksMinutes per deckEst. cost per deck
GPT-6.1 Terra, standard skilllow96%*75%*0%*8.0USD 0.36
GPT-6.1 Terra, standard skillmedium96%*88%*0%*8.8USD 0.44
GPT-6.1 Terra, standard skillhigh100%*100%*0%*9.0USD 0.47
GPT-6.1 Sol, standard skill (reference)medium98%*88%*0%*9.8USD 0.77

Tuned templates

5 requests, 6.4 requested slides on average, one deck per request and setup:

Model and setupReasoning effortClient-ready slidesDecks with no fix neededUnusable decksMinutes per deckEst. cost per deck
GPT-6.1 Terra, standard skilllow78%*40%*20%*8.3USD 0.39
GPT-6.1 Terra, standard skillmedium81%*40%*20%*9.1USD 0.51
GPT-6.1 Terra, standard skillhigh80%*40%*20%*9.7USD 0.57
GPT-6.1 Sol, standard skill (reference)medium85%*60%*20%*11.6USD 0.96
GPT-5.1, optimised setup (current baseline)high97%*80%*0%*17.1USD 1.02

* One deck per request, so no interval; a difference of one deck is within noise.

On the tuned templates every GPT-6.1 setup leaves the template's own sample icons on one deck, and Terra twice opens a deck without the template's photo cover. The GPT-5.1 optimised setup has checks for both; the GPT-6.1 models run the standard skill without them.

Minutes per slide

Minutes per requested slide: wall-clock minutes of a deck divided by the number of slides its request asked for. Minutes per client-ready slide: minutes per deck divided by client-ready slides per deck.

Model and setupReasoning effortNew templates, per requested slideNew templates, per client-ready slideTuned templates, per requested slideTuned templates, per client-ready slide
GPT-6.1 Terra, standard skilllow1.551.641.321.71
GPT-6.1 Terra, standard skillmedium1.701.801.441.87
GPT-6.1 Terra, standard skillhigh1.771.771.572.01
GPT-6.1 Sol, standard skill (reference)medium1.952.001.802.27
GPT-5.1, optimised setuphighnot runnot run2.742.84

Tokens and cost

Per deck, from the platform's usage ledger. Input tokens include cached input, which makes up most of it.

ModelReasoning effortSetModel callsInput tokensOutput tokensEst. cost
GPT-6.1 TerralowNew templates523.3 M6.4 kUSD 0.36
GPT-6.1 TerramediumNew templates574.0 M7.8 kUSD 0.44
GPT-6.1 TerrahighNew templates594.5 M9.4 kUSD 0.47
GPT-6.1 SolmediumNew templates504.0 M9.6 kUSD 0.77
GPT-6.1 TerralowTuned templates563.7 M6.6 kUSD 0.39
GPT-6.1 TerramediumTuned templates655.0 M9.6 kUSD 0.51
GPT-6.1 TerrahighTuned templates695.8 M11.1 kUSD 0.57
GPT-6.1 SolmediumTuned templates605.5 M12.0 kUSD 0.96
GPT-5.1 (optimised setup)highTuned templates704.5 M30.1 kUSD 1.02

Terra's cost advantage over Sol comes from its price per token: it uses about as many tokens per deck.

Which effort we recommend

Medium or high. On new templates high is the only setup with no fix needed on any deck (100% client-ready slides), for about USD 0.03 and 0.2 minutes more per deck than medium. Medium matches Sol. Low is the cheapest and fastest, with the same client-ready share as medium but more decks needing a fix (75% with no fix needed against 88%). On the tuned templates the three efforts are level.

Limitations

  • One deck per request. Each setup ran every request once, so a single deck moves a rate noticeably and there are no intervals.
  • Times on QA. The decks were built on the QA environment with up to eight running at once. GPT-6.1 Sol took 4.7 minutes per deck in the October run and 9.8 here, so absolute times are not comparable across the two pages; the ranking within this page is.
  • Two tuned templates differ. On QA two of the tuned templates are newer versions of the files used in October (same design, other sample content), so tuned-template results are not directly comparable with the GPT-5.1 evaluation.
  • The new-template requests are known to the team. Earlier GPT-5.1 work looked at some of these requests. Those changes are part of the GPT-5.1 optimised setup and are not active for GPT-6.1 models, which run the standard skill.
  • One grader, written by the team that built the setup. A second blind grader on six borderline tuned-template decks agreed on four verdicts.

Method notes

  • Run: 12 November 2026, QA environment, Conduct on the Codex harness. One space per model and effort, identical apart from the model and the reasoning effort, with workspace persistence off. The model and effort of every turn were checked in its recorded run data before it counted.
  • Grading: the pptx-deck-review skill (rubric as of 23 October 2026, unchanged for this run), blind, one review per deck, with the case's human checks. A deck passes only with no major finding. No verdict was left unsure.
  • A grading calibration for composed components on strict templates and for items laid onto a design with a different number of slots is under review. It would not change any new-template result; on the tuned templates it would raise Terra medium and high to three of five decks with no fix needed.
  • Cost: estimated model cost per deck at list prices from the platform's usage ledger; tokens from the same ledger.
  • All figures in the requests are fictitious; no real CID is involved. Which models may process CID is set per deployment; see the LLM Availability Overview.
Last updated