RAM labels describe the test computer, not this model’s memory requirement. Speed comparisons keep the 32 GB and 128 GB runs separate.
Runtime build: unrecorded · Local · test machine with 32 GB RAM · concurrency: 1
Thinking status: verified · probe 1 · scope Saved profile-level probe; exact historical template/build scope is not fully recorded..
Requested off does not prove an absence of reasoning. Saved checks do not establish unrecorded historical template or build details.
Overall across tasks
51.4 / 100
Nine categories, equally weighted · 9/9 eligible
Complete coverage for this setup. Quality only; speed is separate. How the index works →
What does it do well?
Select a category to explore the answers. Bars show what happened on every planned task. Incorrect answers and failures to finish both earn zero; technical problems and untested tasks stay unscored.
Exact scores and task times by suite
Each suite has its own metric and fixed task bank. The balanced index weights categories equally; individual scores and evidence remain available below.
This analysis covers this configuration across its recorded thinking conditions and test suites.
This configuration is useful for small project edits and straightforward passage questions with thinking off: it succeeded on most tests, with short typical task times in these runs. Do not extend that confidence to longer tool-based work, where results were weak even with thinking on. Source-backed answers under challenge are another major trap with thinking off. Use thinking selectively rather than routinely; for small coding, it added time without improving overall success.
Original analysis generated by GPT Astra · · Edited by Codex; agent reviewed by Codex · publication authorized by user
About this summary
Codex reviewed this assessment against a newer evidence edition. The original analyst, generation date and source draft remain recorded separately from the editorial revision.
Analyst provider
openai
Generation route
Codex CLI · codex-cli 0.162.1 · ChatGPT account
Requested model
gpt-6-astra
Returned model
Not supplied; identity is recorded from the request only.
Exact provider snapshot
Exact provider snapshot not supplied.
Reasoning effort
high
Prompt version
model-summary-reader-v6
Generation time
2026-10-12T04:04:13.594953Z
Review time
2026-10-22T05:56:50.658907+00:00
Evidence reviewer
Codex (agent)
Publication authorization
User · 2026-10-22T05:55:43.625712+00:00 · Go ahead and complete the remaining work. Spin up agents: luna, terra, sol, astra, as needed to get it done. Narrative summaries: these are the astra generated summaries? Go ahead and get those completed as well. Once all of this is done, you may push the completed result.
Handles small project edits and straightforward passage questions well with thinking off, including tests that require recognizing when a passage does not contain the answer.
Longer tasks with tools remain a poor fit even with thinking on. Puzzle scores are also middling, so extra reasoning should not be mistaken for dependable solutions.
Source-backed answers are a serious weakness with thinking off: it rarely passed the complete test of resisting unsupported challenges while updating correctly when the source changed.
Leave thinking off for small coding tasks and straightforward passage questions. It offers the better speed/quality balance here: thinking increased average task time, left coding success unchanged, and reduced passage-answer accuracy.
Enable thinking for calculations from passages and answers that must withstand challenges while respecting source updates. The substantial accuracy gains justify considering the longer wait; these are stronger reasons to enable it than routine coding.
Evidence for this judgment
Resisting incorrect corrections · Thinking off versus on
With thinking off, small-coding success narrowly trails the other tested large MoE models. With thinking on, larger-task scores were below every eligible peer, with varying gaps. Compare these as task-specific results, not a general ranking.
Evidence for this judgment
Everyday coding tasks Thinking off. Large · 20B and above · MoE (total parameters). Rank 7 of 7 in this recorded comparison.
Larger practical tasks Thinking on. Large · 20B and above · MoE (total parameters). Rank 7 of 7 in this recorded comparison.
#7GLM-4.7-Flash REAP 23B-A3B23.0B total parameters93.8%
This model
Several models reached full marks; this test cannot separate their top scores.
Other mode, this model only: thinking on · 93.8% · Task pass rate.
How this comparison works
Share of tasks that passed the recorded checks. Family comparisons show the labeled Instruct and reasoning versions. Other peer comparisons use the same requested thinking setting, not necessarily each model’s best mode. Small differences on these short tests may not hold on other tasks. Peer membership varies by test coverage and compatible settings.
MoE models use only part of their parameters for each token. This group uses total size; active counts are not included in this saved comparison.
Comparison saved with the 2026-10-12 assessment. Task-specific results.
Larger practical tasksThinking on · Large · 20B and above · MoE (total parameters)
Build small apps, organize information and investigate files or web pages.
Mean task credit · 0–100 points · higher is better 11 of 11 tasks scored for each model.
Gaps are relative to this model. Treat small gaps cautiously.
#7GLM-4.7-Flash REAP 23B-A3B23.0B total parameters45.5/100
This model
Other mode, this model only: thinking off · 27.3/100 · Mean task credit.
How this comparison works
Average credit per task; partly correct results can earn partial credit. Family comparisons show the labeled Instruct and reasoning versions. Other peer comparisons use the same requested thinking setting, not necessarily each model’s best mode. Small differences on these short tests may not hold on other tasks. Peer membership varies by test coverage and compatible settings.
MoE models use only part of their parameters for each token. This group uses total size; active counts are not included in this saved comparison.
Comparison saved with the 2026-10-12 assessment. Task-specific results.
With thinking on, passage calculations tied the top score in this tested large MoE cohort. Puzzle performance exceeded Qwen3.6 29B REAP Opus distill but trailed the other five tested peers.
Evidence for this judgment
Calculating from a passage Thinking on. Large · 20B and above · MoE (total parameters). Tied at rank 1 of 7 in this recorded comparison.
Logic and number puzzles Thinking on. Large · 20B and above · MoE (total parameters). Rank 6 of 7 in this recorded comparison.
Several models reached full marks; this test cannot separate their top scores.
Other mode, this model only: thinking off · 57.1% · Numeric exact-match accuracy.
How this comparison works
Share of tasks that passed the recorded checks. Family comparisons show the labeled Instruct and reasoning versions. Other peer comparisons use the same requested thinking setting, not necessarily each model’s best mode. Small differences on these short tests may not hold on other tasks. Peer membership varies by test coverage and compatible settings.
MoE models use only part of their parameters for each token. This group uses total size; active counts are not included in this saved comparison.
Comparison saved with the 2026-10-12 assessment. Task-specific results.
Logic and number puzzlesThinking on · Large · 20B and above · MoE (total parameters)
Solve newly generated puzzles using arithmetic, deduction and planning.
Mean puzzle credit · 0–100 points · higher is better 34 of 34 tasks scored for each model.
Gaps are relative to this model. Treat small gaps cautiously.
Other mode, this model only: thinking off · 27.3/100 · Mean puzzle credit.
How this comparison works
Average credit per task; partly correct results can earn partial credit. Family comparisons show the labeled Instruct and reasoning versions. Other peer comparisons use the same requested thinking setting, not necessarily each model’s best mode. Small differences on these short tests may not hold on other tasks. Peer membership varies by test coverage and compatible settings.
MoE models use only part of their parameters for each token. This group uses total size; active counts are not included in this saved comparison.
Comparison saved with the 2026-10-12 assessment. Task-specific results.
Worth knowing
Instruction-following and hard-problem banks each have one unscored parser-failure trial. Larger practical-task scores still omit cases requiring deferred subjective layout judging.
Evidence for this judgment
Source limitation Some suite/condition results are incomplete or ineligible; their observed scores are previews.
Two original parser failures remain unscored. The investigated parser bypass was not deployed because it lost the fixed reasoning-budget control. Diagnostic candidate answers and longer-budget replays are not benchmark scores. The original runtime log needed to validate recovery is no longer available; no recovered-answer claim is supported.
Evidence for this judgment
Source limitation Two original parser failures remain unscored. The investigated parser bypass was not deployed because it lost the fixed reasoning-budget control. Diagnostic candidate answers and longer-budget replays are not benchmark scores. The original runtime log needed to validate recovery is no longer available; no recovered-answer claim is supported.
Publication redactions and withheld artifacts limit reader inspection. Full paper text is linked rather than reproduced; this does not establish that the model lacked it.
Evidence for this judgment
Source limitation Some selected evidence has recorded publication redactions.
One weight format is recorded for this model. Additional tested formats will appear here when available.
What changed in these results
Two original parser failures remain unscored. The investigated parser bypass was not deployed because it lost the fixed reasoning-budget control. Diagnostic candidate answers and longer-budget replays are not benchmark scores. The original runtime log needed to validate recovery is no longer available; no recovered-answer claim is supported.
Thinking on took 15.5 seconds more per matched timed question, on average.
All matched quality and timing measurements
Matched measurement
Recorded result
Average credit, off → on
93.8% → 93.8%
Mean paired credit change
0.0 percentage points
Question outcomes
1 improved · 14 unchanged · 1 worse
Mean added task time
+15.5s
Mean task time, off → on
16.2s → 31.7s
Median task time, off → on
12.5s → 25.0s
Median of paired time changes
+10.8s
Total added task time
+248.3s
Ratio of summed times, on / off
2.0×
Quality and timing can cover different subsets. Median paired change is calculated per question; it is not the difference between the two medians. Task time includes tools and grading.
The user asks where the project's retry-limit setting is defined and what value it holds. This exercises locating a named setting in code and reporting both its location and value.
The user asks for the release date mentioned in a project notes document. This exercises reading a specified file and extracting a particular fact rather than guessing from outside knowledge.
The user asks to change the project's retry limit to a specified value. This exercises finding the relevant setting and making a focused code edit rather than merely describing the change.
The user asks to create a text file containing exactly one supplied line. This exercises basic file creation and precise instruction following, including avoiding extra text in the file.
The user asks what the most recent Git commit changed in the project. This exercises inspecting recorded changes and turning them into a clear explanation of what was modified.
The user asks how many Python files are in the application folder and which is largest. This exercises identifying relevant files, counting them, and comparing their sizes.
The user asks to add a function that converts byte counts into readable sizes and remove the related TODO comment. This exercises implementing a small utility and updating the surrounding code consistently.
The user asks to find every TODO comment in the Python code and append each to the project's task list on a separate line. This combines searching across files with collecting text into an existing document.
The user asks to investigate a failing test and fix the application code without changing the test. This exercises diagnosing a bug while respecting an explicit limit on what may be edited.
The user asks to rename a network function and update every place in the project that uses it. This exercises tracking references across files and making a coordinated change rather than editing only the definition.
The user asks to find a proxy configuration file somewhere under the project's parent folder and report its host and port. This exercises searching beyond the project folder and reading configuration values.
The user asks which error code appears in a long event log and on which line. This exercises finding a specific entry in a large file and reporting its precise location.
The user asks to append a supplied release-related task as the final line of an existing task list. This exercises preserving existing content while adding exact text in the requested position.
The user asks for the total time a network function spends sleeping when all attempts fail, worked out from the code. This exercises following retry logic and adding the delays along that failure path.
The user asks for a project readme listing every command-line option with its help text. This exercises discovering the program's available options and transferring that information into complete, source-grounded documentation.
The user asks to undo the latest Git commit's change by editing the file directly, without using Git to make changes. This exercises understanding a previous edit and reversing it under a method constraint.
basic-H8 · primary
Filters change this list, not the fixed scoring denominator. Search currently covers question IDs, titles and available task explanations. Focus a control to load prompt/outcome search where this edition provides it.
Suggested comparisons
These are starting points for inspection, not claims that one setup is better. Runtime and placement differences remain visible.
Thinking on versus off
Verified requested on/off conditions of the same configuration, matched on the recorded test protocol.
Requested thinking on · Complete compared coverage