TL;DR
Luna recorded higher judged scores and lower historical cost than GLM 5.2 on the August 2026 planning and next-action replays. Execution scores were 89.2 versus 82.4, at $1.67 versus $25.37 per 1,000 decisions. On one sandbox research task, Luna reached B+ three times while GLM reached it once and finished at D. Judge dependence, cached billing and uneven CPU throttling limit broader quality or speed conclusions.
Planning models on 29 frozen conversations: published results
Tick: recorded mean score. Whisker: mean ± one recorded standard error (SE), not a confidence interval. Overlap alone does not establish a ranking.
16-model campaign, 29 frozen conversations; 14 model rows shown in the public table. Published 2026-08-08.
Show the table · 2 rows, 5 columns
Scroll sideways to compare all columns. Select a heading to sort.
| openai/gpt-5.6-luna | 0.731 | 0.068 | $0.0012 | 8.6s |
| z-ai/glm-5.2 | 0.515 | 0.079 | $0.0246 | 44.8s |
No matching rows. Clear the filter to see all records.
16-model campaign, 29 frozen conversations; 14 model rows shown in the public table. Published 2026-08-08.
Which fits the tested job?
Luna had the higher observed decision scores and lower historical cost across the two replay tasks. It also reached B+ more often on the one sandbox task. GLM's single favorable sandbox grade is useful evidence, but its later regressions prevent presenting that run as a stable successful artifact.
Planning models on 29 frozen conversations
The planning means were 0.731 for Luna and 0.515 for GLM, with costs $0.0012 and $0.0246 per decision and p50 8.6s and 44.8s. The table supports the observed operating-cost difference without supplying a pair-specific statistical quality verdict.
Next-action execution model bakeoff
Luna scored 89.2 versus GLM's 82.4, with schema validity 99.0% versus 98.1%. Recorded cost was $1.67 versus $25.37 per 1,000 decisions and p50 5.6s versus 15.7s. Keep the judge and prompt-cache billing limitations alongside that comparison.
Five models completing one sandbox research task
Luna's trajectory returned to B+ three times; GLM reached B+ once and finished at D. Recorded wall/cost totals were 59m/$2.63 versus 91m/$2.72. This favors investigating Luna for that particular remediation workflow, while the one-task sample and CPU-throttling differences limit any broader speed claim.
Route a request and plan the work: follow the published task
A useful planner chooses the right kind of work before doing it. The complete customer conversations and system prompts remain private.
Published task summary
An incoming request and its existing conversation.
Choose whether to write a plan, call a catalog tool, ask for clarification or continue an existing agent.
Read what came back
A routing decision and plan. Every model receives the same saved conversation; the test compares that decision before fresh research begins.
Replay 29 frozen conversations across the model arms, then compare graded plan scores, billing and response time.
Results from tasks both models attempted
The tested model IDs are openai/gpt-5.6-luna and z-ai/glm-5.2. Every result table below comes from a study in which both participated.
Compare the measured quality, historical billed cost and elapsed time within the same task and rubric.
Scores from different tasks are not combined into one winner score.
Planning models on 29 frozen conversations: read these measurements
Read the complete rubric, sample and billing notes
16-model campaign, 29 frozen conversations; 14 model rows shown in the public table
Route an incoming request: write a plan, call a catalog tool, ask for clarification or continue an existing agent.
Scoring
Top planning means were close at this sample size. The source's paired comparison with Opus was inconclusive.
Cost units
Cost is provider-billed dollars per decision. Prefix reuse can make it differ from list price multiplied by every input token.
Time units
p50 is median latency in seconds, not the duration of complete research.
Frozen conversation
A saved input kept identical across model arms so their decisions can be compared.
Standard error (SE)
An estimate of uncertainty around the sample mean. A small mean difference needs more evidence than the ranking alone.
Next-action execution model bakeoff: published results
23-model campaign; 69 frozen decisions × 3 samples; 12 published model rows. Published 2026-08-08. Each bar uses the original task scale. The original table retains all other dimensions and qualifications.
Show the table · 2 rows, 5 columns
Scroll sideways to compare all columns. Select a heading to sort.
| openai/gpt-5.6-luna | 89.2 | 99.0% | $1.67 | 5.6s |
| z-ai/glm-5.2 | 82.4 | 98.1% | $25.37 | 15.7s |
No matching rows. Clear the filter to see all records.
23-model campaign; 69 frozen decisions × 3 samples; 12 published model rows. Published 2026-08-08.
Choose the next allowed agent action: follow the published task
A well-formed action can still be the wrong decision. Keep output validity and judged decision quality separate.
Published task summary
The existing plan, conversation and results returned by tools.
Choose one next allowed action, or finish with an answer when the evidence is sufficient.
Read what came back
A structured action. The replay records the proposed action without executing its tools.
Replay 69 saved decisions three times each. Check the required output structure separately from the judge score.
Next-action execution model bakeoff: read these measurements
Read the complete rubric, sample and billing notes
23-model campaign; 69 frozen decisions × 3 samples; 12 published model rows
Choose one next action in an unfinished trading-agent workflow. This replay measures individual decisions, not a fresh complete trading run.
Scoring
Mean action score comes from the judge. Schema validity is the share of actions matching the required output structure; a valid structure can still contain the wrong action. Switching judges reduced the Luna-versus-DeepSeek score gap from 13.9 to 1.5 points.
Cost units
Cost uses historical provider billing per 1,000 decisions. Reused prefixes earned cache discounts; the cached-token split was not retained.
Time units
Prod p50 is median production response latency, separate from replay quality and complete agent wall time.
Five models completing one sandbox research task: published results
5 models on the same single research task; real Modal virtual machines. Published 2026-08-08.
Show the table · 2 rows, 4 columns
Scroll sideways to compare all columns. Select a heading to sort.
| openai/gpt-5.6-luna | D → B+ → D → B+ → D → B+ | 59m | $2.63 |
| z-ai/glm-5.2 | D → D → B+ → D → D → D | 91m | $2.72 |
No matching rows. Clear the filter to see all records.
5 models on the same single research task; real Modal virtual machines. Published 2026-08-08.
Write research code and repair its artifacts: follow the published task
The grade sequence shows whether a repair held up at the next review. One B+ followed by D is a different outcome from repeated favorable reviews.
Run one concrete research task
Correlate Strait of Hormuz ship traffic with the next USO trading session.
Write and run research code on real Modal machines, return the artifacts for review and repair defects raised by the grader.
Inspect the recorded repairs and regressions
Luna returned D → B+ → D → B+ → D → B+; GLM returned D → D → B+ → D → D → D. DeepSeek returned D on all six reviews.
The source found self-check code that assigned the same variable to delivered_stats and recomputed, then hard-coded a match. That explains why a passing self-check could survive while the delivered analysis still failed review. The publication does not assign this particular code defect to one model in the table.
Five models completing one sandbox research task: read these measurements
Read the complete rubric, sample and billing notes
5 models on the same single research task; real Modal virtual machines
Correlate Strait of Hormuz ship traffic with the next USO trading session.
Scoring
Grades describe the delivered artifact after successive reviews. Oscillation exposed holes in agent-written self-checks; a successful process exit was insufficient.
Cost units
Cost is historical billed run cost. Flash Lite has no reported total after failing out; a dash is missing data, not zero.
Time units
Wall time covers the complete workflow. Uneven web-tier CPU throttling affected timing, so the source treats it as a soft comparison.
Sandbox
An isolated machine where the agent runs research code.
No comments yet.