How We Test Prompts: Methodology

In short: we ran every prompt on this site 3 times on each of five named models, using the same fictional input each time. The fourth model, Claude Fable 5.1, and the fifth, Gemini 4 Argon, ran later than the other three on the same locked prompts, inputs and rubrics. Each answer was graded against a rubric we wrote and locked before the first run. The grader was an LLM from the other model family, and a separate AI editor agent re-checked a sample of its verdicts. These checks covered all prompt test rounds, including a new round of Excel and Microsoft 365 tasks. The bench now also includes a second Spanish round and French and Portuguese versions of our Claude Skills. We also tested our Claude Skills separately, without a spot-check or judge-swap calibration. Every transcript is now public. That is all “tested” means. It is not a professional sign-off, and it is not a test of the ChatGPT, Claude or Gemini apps, or of Copilot.

What did we test, and when?

The first test round ran on October 19, 2026. It covered 40 prompts (30 in English, 10 in Spanish) across 10 professions, with 5 rubric criteria per prompt.

For the three original models, that makes 360 runs and 1,800 criterion verdicts. Every run completed, and we didn’t have to drop any. The claude-fable-5-1 row was run later, on October 20, 2026. It added 120 runs and 600 verdicts in this round. The gemini-4-argon-medium row also ran on October 20, 2026. It added another 120 runs and 600 verdicts in this round.

The second round ran the same day, October 19, 2026, on the same bench with the same judges. It covered 40 prompts adapted to local law and practice: 20 for France and 20 for Brazil, written in French and Portuguese. Each prompt again had 5 rubric criteria.

Again, the three original models made 360 runs and 1,800 criterion verdicts. Here too, the claude-fable-5-1 row was run on October 20, 2026, and added 120 runs and 600 verdicts. The gemini-4-argon-medium row was run on the same day and also added 120 runs and 600 verdicts.

Claude Fable 5.1 ran on the same locked prompts, inputs and rubrics as the other three models, with the same harness. Gemini 4 Argon (gemini-4-argon-medium, Gemini 4 Argon at its medium reasoning setting) also ran on the same locked prompts, inputs and rubrics, but through Google’s own command-line client, described below. Across both rounds, the three original models make 80 prompts, 720 runs and 3,600 verdicts. We dropped 0 runs from those original tests. With Claude Fable 5.1 added, the totals are 960 runs and 4,800 verdicts. With both Claude Fable 5.1 and Gemini 4 Argon added, rounds 1 and 2 total 1,200 runs and 6,000 verdicts.

On October 20, 2026, we also ran a new round of 10 Excel and Microsoft 365 prompts on all five models, 3 runs each. That round made 150 runs and 750 verdicts, and it has its own section below. Counting all prompt rounds before the second Spanish round, including the Excel and Microsoft 365 round but excluding the separately reported skills round, the bench covers 90 prompts, 1,350 runs and 6,750 verdicts.

On October 21, 2026, we ran a second Spanish round of 10 new prompts, 2 each for engineers, travel agents, construction managers, event planners and interior designers. Four are set in Mexico, three in Spain and three in Colombia. We wrote and locked their rubrics as a separate set, with its own lock, before the first run. Each prompt has 5 criteria, and all 10 include a SAFETY criterion. All five models ran every prompt 3 times, with the same judge per family: gpt-6-astra for Opus, Sonnet, Fable and Gemini, and claude-opus-5-5 for gpt-6-astra. That round made 150 runs and 750 verdicts, and every run completed. With it, all prompt rounds together, still excluding the skills rounds, cover 100 prompts, 1,500 runs and 7,500 verdicts.

How is a prompt run?

• Every model gets the same input. Each prompt comes with a fictional, de-identified sample input, such as a client record, a ledger or a transcript. We save the exact query text to a file and send it unchanged to every model.
• Each Claude and GPT run goes through an agent client, not a bare API call. For these models, we send the query with an agent command-line client in its safe mode: tools, web access, memory, skills and custom instructions are off. The client still adds its own generic assistant system instruction. We did not write it, and it is the same for every Claude and GPT model and prompt. Claude models were reached through a routing service on a subscription, and gpt-6-astra through an OpenAI subscription sign-in, not through each vendor’s plain API. Nothing is edited by hand between the prompt and the output.
• Gemini runs went through Google’s own client. We reached gemini-4-argon-medium through Antigravity CLI, Google’s command-line agent. That is not a direct API request, and it is not the Gemini app. For each run we used a custom agent with no tools, skills or rules loaded, and a one-line system prompt (“You are a helpful assistant.”). Each run started in a fresh, empty folder in sandbox mode, and permission prompts were not auto-approved. The client still wraps the query with the local time and a model-selection note. It also still exposes one built-in tool, for listing or stopping background tasks. After each run we read the client’s own log. A run with any tool call would not count as completed. In the 240 Gemini runs of rounds 1 and 2 there were 0 tool calls, and each run took 1 turn.
• Each model gets 3 runs. Output changes from run to run, so one good answer proves little. The site always says how many runs were graded.
• We keep the transcripts, and they are public. Each run is saved word for word, along with its model id, start time, duration, exit code and the rubric’s hash. You can download the full archive: test transcripts, October 2026 (ZIP, 9.2 MB). It holds 3,015 files, including 2,052 transcripts: the 1,200 prompt runs from rounds 1 and 2 on all five models, the 150 runs of the Excel and Microsoft 365 round, the 150 runs of the second Spanish round, 327 skill runs, including runs on earlier skill versions, and 225 runs from the French and Portuguese skills round. Of those 225, 180 are on the current texts and 45 are on the earlier version 1.0 and 1.1 texts of the revised adaptations, kept as labeled evidence. It also holds each criterion verdict with the judge’s one-line reason, the locked rubrics with their hashes, the calibration and spot-check files, and earlier skill files. A README in English explains the structure. We removed command lines, session and conversation ids and local file paths, and reduced provider names to the vendor (anthropic, openai or google). All inputs are fictional. With the archive, anyone can re-check any verdict on this site.

How is an answer graded?

1. We write the rubric first. Before any model saw a prompt, we wrote 4–6 pass/fail criteria for it (5 in every prompt round). Some check correctness: does the reconciliation balance, and is the deflection recomputed with consistent units? In round 1, 32 of the 40 rubrics have one criterion labeled SAFETY: no invented figures, citations, dates or legal conclusions. The other 8 check the same risk inside a correctness criterion, for example, “adds no exam findings that are not in the transcript.” In round 2, all 40 rubrics have a SAFETY criterion. The rubric is hashed and locked, and the harness refuses any run whose rubric hash doesn’t match.
2. The primary grader is always from the other model family. gpt-6-astra grades Claude outputs and claude-opus-5-5 grades GPT outputs, so no model grades its own family in the primary grading. gpt-6-astra also grades Claude Fable 5.1, like the other Claude models. It grades Gemini 4 Argon too, so Gemini has the same judge as Opus, Sonnet and Fable. For each criterion the grader returns strict JSON: pass or fail, plus a one-line reason. If it is unsure, it has to mark fail.
3. An AI editor agent spot-checks the grader. The editor is an AI agent separate from the grader (the agent that operated the test bench), not a human. In round 1 it read 20 verdicts with the full transcripts. Fourteen were picked at random, half passes and half fails. The other six were GPT passes on safety or length criteria, picked to catch a lenient grader. The editor wrote down its own verdict before reading the grader’s. The two agreed on 18 of 20 (90%), and in both disagreements the grader was stricter than the editor. In round 2 the editor checked 10 verdicts against the full transcripts, five passes and five fails, on French and Portuguese outputs. It agreed with the grader on 9 of 10. In the one disagreement, the grader was again the stricter one. For Claude Fable 5.1 the editor checked 12 verdicts, six passes and six fails, on English, Spanish, French and Portuguese outputs. It agreed with the grader on 8 of 12 (67%). In all 4 disagreements, the grader was stricter than the editor. For Gemini 4 Argon the editor checked 10 verdicts, five passes and five fails, on English, French and Portuguese outputs. It agreed with the grader on 7 of 10. In all 3 disagreements, the grader was stricter than the editor. In the Excel and Microsoft 365 round, it checked 10 verdicts, five passes and five fails, across all four languages, and agreed with the grader on 10 of 10.
4. We swapped the judges to calibrate them. In round 1 we re-graded 60 sampled runs with the other judge. On the same Claude runs, gpt-6-astra passed 93% of criteria and claude-opus-5-5 passed 98%. On the same GPT runs, the two judges came out the same (97% each). In round 2 we re-graded 10 Claude runs and 10 GPT runs. On the Claude runs, gpt-6-astra passed 74% of criteria and claude-opus-5-5 passed 86%. On the GPT runs, both judges gave 98%. The two judges are not equally strict, and the gap grew. On the same Claude runs it was 12 points in round 2, compared with 5 points in round 1. For Claude Fable 5.1 we re-graded 10 runs per round with claude-opus-5-5. In round 1, gpt-6-astra passed 88% of criteria and claude-opus-5-5 passed 100%, a 12-point gap. In round 2 it was 82% against 90%, a gap of 8 points. The judge asymmetry shows again on Fable outputs. For Gemini 4 Argon we also re-graded 10 runs per round with claude-opus-5-5. In round 1, gpt-6-astra passed 92% of criteria and claude-opus-5-5 passed 98%, a 6-point gap. In round 2 it was 86% against 88%, a gap of 2 points. On Gemini outputs too, gpt-6-astra was the stricter judge. In the Excel and Microsoft 365 round we re-graded 6 runs per model family with the other judge. On the Claude runs, gpt-6-astra passed 93% of criteria and claude-opus-5-5 passed 97%. On the Gemini runs it was also 93% against 97%. On the GPT runs, claude-opus-5-5, which grades GPT on the site, passed 87%, and gpt-6-astra passed 90%. In that round the gaps were small, 3 points on each family, but the samples were small too. So part of every Claude-versus-GPT gap on this site comes from the judge, not the model. That includes Claude Fable 5.1, and the same holds for any Gemini-versus-GPT gap. We cannot say exactly how much. Treat any Claude-versus-GPT comparison here as rough. Opus and Sonnet share a judge, so that comparison is like-for-like. Fable and Gemini have that same judge, so comparing them with Opus and Sonnet is like-for-like too.
5. The second Spanish round had its own checks. We re-graded 6 runs per model family with the other judge. On the Claude runs, gpt-6-astra passed 77% of criteria and claude-opus-5-5 passed 93%. On the Gemini runs it was 73% against 90%. On the GPT runs, both judges gave 93%. The judge asymmetry was larger in this round: 17 points on the Claude runs and 17 points on the Gemini runs. So the GPT figure in this round is not like-for-like with the other four models. The editor agent also checked 12 verdicts on Spanish outputs, six passes and six fails. It agreed with the grader on 8 of 12 (67%). In all 4 disagreements, the grader was stricter than the editor.

What do the labels mean?

• “Tested · date” on a prompt means at least 2 of its 3 runs on that model completed and were graded. It does not mean every run passed every criterion. Among the original three models in round 1, 48 of the 120 prompt-and-model results had at least one graded run that missed a criterion. In round 2, it was 79 of 120. If a prompt does not meet that minimum on a model, we publish it without a tested label for that model. A run that did not complete never counts toward a “tested” mark.
• “4.3 of 5 criteria on average” is the average number of rubric criteria passed across the graded runs.
• “3 runs graded” is how many runs were completed and graded.
• “all criteria met in 1 of 3” (in the test details) is how many of those runs passed every criterion.
• “judge: …” names the model that graded those runs.
• For the earlier prompts in the prompt library, the model shown on a prompt card has the highest average score among the original three models. If two models tie, the card shows the one with more runs that passed every criterion. On the Excel and Microsoft 365 prompts, which are new, the card can name any of the five models. On the earlier prompts it still names the best of the original three, and Claude Fable 5.1 and Gemini 4 Argon appear as extra lines in the test details. Earlier “tested · date” records were not changed when we added Fable and Gemini. Because of the judge gap above, a card showing GPT may reflect a gentler judge rather than a better answer.

What “tested” does NOT mean

• It isn’t the consumer app, and it isn’t the plain API. We tested named model versions through agent clients with a generic system instruction and no tools in use. Gemini 4 Argon ran through Google’s command-line client, not the Gemini app. ChatGPT, Claude.ai, the Gemini app and Copilot add their own instructions, tools and memory, so your results in an app may differ. The Excel and Microsoft 365 round is not a test of Copilot in Excel or Microsoft 365.
• It isn’t a professional sign-off. No accountant, clinician, engineer or lawyer reviewed every output. Passing the rubric means an output met our criteria on our fictional input. Check every figure, citation and deadline before you rely on it.
• It isn’t a guarantee for your data. Real inputs are messier than our samples.
• It isn’t an overall ranking of models. Ten professions with a few tasks each is a narrow slice, and the grader-strictness gap above applies.

What did the first round show?

These findings cover the original three models; Fable’s and Gemini’s results follow below.

• All three models handled most tasks. No model averaged below 60% of the criteria (3.0 of 5) on any prompt, and no run was unusable.
• Most lost points came from SAFETY criteria: the model added a warning, date, benefit or legal assumption that wasn’t in the input. In one example, a patient leaflet added an emergency warning the doctor never gave. In another, a sales email hinted at a result nobody had approved.
• Word limits were the other common miss, mostly for the two Claude models. Between them they had 8 length-failure verdicts. gpt-6-astra had none.
• The raw averages were 99% of criteria for gpt-6-astra, 92% for claude-opus-5-5 and 91% for claude-sonnet-5. Read them alongside the calibration above: on 30-run samples graded by the same judge, Claude scored 98% and GPT 97%.

For prompt-by-prompt numbers, see the round 1 results in detail.

What did round 2 show?

These findings also cover the original three models.

• The inputs were harder on purpose. They contained traps: a quote with a wrong total, a wrong unit conversion, sensitive data sent over a messaging app, or a discriminatory request from a client. A good answer had to catch the trap, not just format the output.
• The raw averages were 98% of criteria for gpt-6-astra, 79% for claude-opus-5-5 and 70% for claude-sonnet-5. The lowest average on any prompt for any model was 40% of criteria (2.0 of 5).
• Claude lost points mostly on word limits and added facts. The two Claude models had 107 length-failure verdicts between them. They also added details the input did not contain, such as a quality of a property that was not in the agent’s notes.
• One error repeated. In the calculation checks, claude-sonnet-5 repeatedly copied the wrong unit conversion from the input instead of correcting it.
• Read these numbers with the calibration. In this round the judge gap on Claude runs was 12 points. Part of the gap between GPT and the Claude models comes from the judge. The 9-point gap between Opus and Sonnet is like-for-like, because the same judge graded both.

How did Claude Fable 5.1 do?

We ran claude-fable-5-1 on October 20, 2026, on the same 80 locked prompts, inputs and rubrics, with the same harness. All 240 runs completed and were graded.

• Against the other Claude models, the comparison is like-for-like. gpt-6-astra graded Fable, Opus and Sonnet alike. In round 1, Fable passed 90% of criteria, against 92% for claude-opus-5-5 and 91% for claude-sonnet-5. That is 2 points below Opus and 1 point below Sonnet. Prompt by prompt, Fable scored higher than Opus on 4 prompts, lower on 11 and the same on 25. It met all criteria in 67 of 120 runs (56%), against 79 of 120 (66%) for Opus. Its lowest average on any prompt was 3.0 of 5.
• Round 2 was close. Fable passed 78% of criteria, against 79% for Opus and 70% for Sonnet. That is 1 point below Opus and 8 points above Sonnet. It scored higher than Opus on 17 prompts, lower on 14 and the same on 9. It met all criteria in 28 of 120 runs (23%), against 35 of 120 (29%) for Opus. Its lowest average on any prompt was 2.33 of 5.
• The honest reading. On these locked prompts, Fable did not beat Opus 5.5 in round 1. In round 2 it was level with Opus within a point. A flagship label is not a result.
• Fable versus GPT is not like-for-like. gpt-6-astra is graded by claude-opus-5-5, a different judge. On the Fable calibration samples, gpt-6-astra as a judge was 12 points stricter than claude-opus-5-5 in round 1 and 8 points stricter in round 2. Read any Fable-versus-GPT gap with that in mind.
• Where Fable lost points. Like the other Claude models, it missed word limits: 5 length-failure verdicts in round 1 and 38 in round 2. It also added facts the input did not contain, or extra regulatory checks nobody asked for.
• Claude Skills. Fable ran the same 20 skill-language pairs 3 times each, on the same skill versions as the other Claude models. That makes 60 graded runs. Averages ranged from 4.0 to 5.0 of 5, and 15 of the 20 pairs were at 5 of 5. The same caveat as the skills round applies: same inputs, no judge-swap calibration and no spot-check on skills.

How did Gemini 4 Argon do?

We ran gemini-4-argon-medium on October 20, 2026, on the same 80 locked prompts, inputs and rubrics, through Google’s command-line client described above. All 240 runs completed and were graded.

• Against the Claude models, the comparison is like-for-like. gpt-6-astra graded Gemini, Opus, Sonnet and Fable alike. In round 1, Gemini passed 91% of criteria, against 92% for claude-opus-5-5, 91% for claude-sonnet-5 and 90% for claude-fable-5-1. Prompt by prompt, Gemini scored higher than Opus on 9 prompts, lower on 13 and the same on 18. It met all criteria in 75 of 120 runs (63%), against 79 of 120 (66%) for Opus. Its lowest average on any prompt was 3.0 of 5.
• Round 2 was close too. Gemini passed 80% of criteria, against 79% for Opus, 78% for Fable and 70% for Sonnet. It scored higher than Opus on 16 prompts, lower on 15 and the same on 9. It met all criteria in 36 of 120 runs (30%), against 35 of 120 (29%) for Opus. Its lowest average on any prompt was 2.67 of 5.
• The honest reading. Gemini was level with Opus within about a point in both rounds: 1 point below in round 1 and 1 point above in round 2. Differences that small on 40 prompts per round are not a ranking.
• Gemini versus GPT is not like-for-like. gpt-6-astra is graded by claude-opus-5-5, a different judge. On the Gemini calibration samples, gpt-6-astra as a judge passed 92% of criteria against 98% for claude-opus-5-5 in round 1, and 86% against 88% in round 2. Part of every Gemini-versus-GPT gap on this site is the judge.
• Where Gemini lost points. Like the Claude models, it missed word limits: 5 length-failure verdicts in round 1 and 45 in round 2. It also added facts the input did not contain.
• No Claude Skills. We did not run Gemini on our Claude Skills. Skills are a Claude format.

How did we test Excel and Microsoft 365 tasks?

On October 20, 2026, we added a round of 10 prompts for everyday office work: 4 in English, 2 in Spanish, 2 in French and 2 in Brazilian Portuguese. Six are Excel tasks and four are Microsoft 365 tasks. They are not tied to one profession. Each prompt has 5 criteria, and all 10 include a SAFETY criterion. As before, we wrote and locked each rubric before the first run. All five models ran every prompt 3 times, with the same judge per family as in the other rounds.

• Results. claude-opus-5-5 passed 93% of criteria, gemini-4-argon-medium 91%, claude-fable-5-1 88% and claude-sonnet-5 83%. gpt-6-astra passed 88% under its own judge, claude-opus-5-5. All 150 runs completed, and all 50 prompt-and-model results are tested. 26 of the 50 had at least one run that missed a criterion.
• By tool. On the Excel prompts, Gemini passed 97% of criteria, Opus 96%, Fable 92%, GPT 87% and Sonnet 83%. On the Microsoft 365 prompts, GPT passed 90%, Opus 88%, and Fable, Gemini and Sonnet 82% each.
• Older lookup functions. In the three non-English lookup prompts, the rubric asked for XLOOKUP in its localized name (BUSCARX, RECHERCHEX or PROCX) or for INDEX/MATCH. 20 failed verdicts came from runs that used the older VLOOKUP family instead (BUSCARV, RECHERCHEV or PROCV). Most of them were from Sonnet and GPT, 9 each. That formula can work on the data, but it did not meet the locked rubric.
• Open items listed as decisions. In the two meeting-minutes prompts, 22 failed verdicts were items listed as decisions that were still open in the input.

Read these results with care. The answers are text only. We judged formulas by their logic and by the result they stated on the given data. We did not run them in Excel. These are chat models answering a pasted task. This is not Copilot inside Excel or Microsoft 365, and Copilot is not in our bench. GPT is graded by a different judge, so read its figure with the calibration above.

What did the second Spanish round show?

We ran the 10 new Spanish prompts on October 21, 2026, on all five models, 3 runs each. All 150 runs completed and were graded.

• Opus was clearly ahead of the other same-judge models. gpt-6-astra graded Opus, Sonnet, Fable and Gemini, so those four compare like-for-like. Opus passed 83% of criteria. Gemini, Fable and Sonnet were close to each other, between 72% and 74%.
• GPT is not like-for-like. gpt-6-astra passed 98% under its own judge, claude-opus-5-5. In this round’s calibration, gpt-6-astra as a judge was 17 points stricter than claude-opus-5-5 on Claude runs and on Gemini runs. Part of every gap between GPT and the other four models here is the judge.
• Missed criteria were common. 38 of the 50 prompt-and-model results had at least one run that missed a criterion. No prompt had all 15 runs meet every criterion.
• Word limits cost points again, for every model except GPT. The last column shows the length-failure verdicts per model.
• The honest reading. This is 10 prompts. It shows how the models did on these tasks, not a ranking.

The prompts from this round are in our prompt library.

How did we test the Claude Skills?

Our Claude Skills are instruction packages (a SKILL.md file) for Claude. We tested them in a separate round. The Skills are published under CC BY 4.0 (Creative Commons Attribution 4.0 International), so anyone may use, adapt and share them, including commercially, with attribution, and each skill download includes a LICENSE.txt file.

• What was tested. 10 skills, one per profession, each in English and Spanish. That makes 20 skill-language pairs.
• Models and judge. Initially, only claude-opus-5-5 and claude-sonnet-5, because skills are a Claude format. Each model ran each pair 3 times on the same bench. gpt-6-astra graded every run, so the judge is again from the other family. Claude Fable 5.1 ran the same pairs later. Its results are in the section above.
• Inputs and rubrics came first. For each pair we wrote one fictional input and a 5-criterion rubric, and hash-locked both before we wrote the skills.
• Version 1.0 had real defects. The model stopped to ask clarifying questions instead of delivering the output in one reply. It cited official sources that the skill lists for the professional, rather than for inclusion in the output. It also added things the input did not contain. 12 of the 40 model-and-pair results for version 1.0 were below 4 of 5. The lowest was 0.33 of 5.
• What we changed. We revised 8 skills with general behavior rules, not fixes aimed at one input: deliver in one reply, with assumptions labeled, and do not cite the skill’s reference sources in the output. 3 skill-language pairs needed a second revision.
• Where they stand now. For Opus and Sonnet on the current versions, averages range from 4.0 to 5.0 of 5, and 27 of the 40 model-and-pair results are at 5 of 5. These results cover 120 graded runs.

The re-runs used the same locked inputs and rubrics as the first run. The improvement is shown on these inputs only. It is not proven for every input. We ran no judge-swap calibration and no spot-check on the skills round. We therefore have not measured judge asymmetry on these outputs or independently checked a sample of their verdicts. A skill is listed as tested on a model only if at least 2 of 3 runs on its current text completed and were graded.

How did we test the French and Portuguese Skills?

On October 21, 2026, we tested French (France) and Brazilian Portuguese adaptations of the same 10 skills. That makes 20 skill-language pairs, all in the skills library.

• Models and judge. claude-opus-5-5, claude-sonnet-5 and claude-fable-5-1 each ran every pair 3 times. That made 180 runs on the first texts. gpt-6-astra graded every run.
• Inputs and rubrics came first, with their own lock. For each pair we wrote a fictional input and a 5-criterion rubric, and hash-locked them separately before we wrote the adaptations.
• First texts. Averages ranged from 3.67 to 5.0 of 5. 1 of the 60 model-and-pair results was below 4 of 5.
• What we revised. We revised 4 adaptations with general behavior rules. In the event-planner vendor RFP, in French and in Portuguese, evaluation criteria were left blank when the brief gave none. In the French construction daily report and delay notice, the notification deadline was hedged. In the French sales pipeline review, a deal with a single contact was not flagged. The French construction adaptation needed a second revision. We re-tested the revised texts in 45 runs on the same locked inputs.

Average criteria passed out of 5, before → after revision, with models listed in the order Opus / Sonnet / Fable:

• Event-planner vendor RFP, French: 4.0 / 4.0 / 4.0 → 4.33 / 5.0 / 4.67
• Event-planner vendor RFP, Portuguese: 4.0 / 4.0 / 4.0 → 5.0 / 5.0 / 4.33
• Construction daily report and delay notice, French (version 1.2): 4.0 / 4.0 / 4.0 → 5.0 / 5.0 / 5.0
• Sales pipeline review, French: 3.67 / 4.67 / 4.33 → 5.0 / 5.0 / 4.67

Where they stand now. On the current texts, averages range from 4.33 to 5.0 of 5, and 46 of the 60 model-and-pair results are at 5 of 5. These results cover 180 graded runs.

The same caveat applies here. The re-tests used the same locked inputs and rubrics as the first runs, so the improvement is shown on these inputs only. We ran no judge-swap calibration and no spot-check on this round. The earlier texts and their runs stay in the archive.

How do we handle re-tests and corrections?

A test record is a dated historical fact. When a new model version comes out, we add it and mark the affected prompts “needs re-test.” We never relabel an old result as a newer model. Adding Claude Fable 5.1 and Gemini 4 Argon on October 20, 2026, followed this rule: we added their runs as new models and kept the earlier records as they were; on October 21, 2026, we added a separately locked Spanish round and French and Portuguese Skills, and revised 4 adaptations while keeping their earlier versions and runs in the archive. If you spot an error, including a verdict you disagree with after checking the archive, tell us through the contact page and we’ll correct it with a dated note.