AgentBench / results
← All models

GPT-OSS 120B (cloud)

1 tested weight format · Explore results, then inspect the answers.

Cloud reference · DeepInfra — openai/gpt-oss-120b. Quality is scored with the same task bank and graders. All speed and desktop-memory measurements are excluded. Hosted precision and reasoning settings may differ from local GGUF runs.

Choose a tested configuration

Hosted model source

DeepInfra · Hosted BF16 (provider-reported)

This result tested a cloud endpoint, not a downloadable GGUF.
Open hosted model →

Setup, memory & provenance · Hosted BF16 (provider-reported)

Hosted BF16 (provider-reported) · Not recorded of weights · openrouter · context 16384 · GPU layers unrecorded

Large · 20B and above · MoE · 117.0B total · 5.1B active

Memory fit has not been verified for this configuration.

completed Hosted quality reference; not a local GGUF test. All speed and local memory measurements are excluded.

Recorded settings and thinking verification

Temperature 0 · sample index 42 · output limit 4096 · reasoning effort medium

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 · Cloud · DeepInfra · concurrency: 1

Thinking status: native reasoning · probe cloud-transport-v1 · scope Native reasoning condition; no verified off/on comparison. Provider-reported usage audited against 16K context. Sample indices are not provider seeds. Requested temperature recorded..

Requested off does not prove an absence of reasoning. Saved checks do not establish unrecorded historical template or build details.

Overall across tasks

72.9 / 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.

Test suiteReasoning · mediumDetails
Everyday coding tasksTask pass rate
Pass rate
8/16 · 50%
Answers completed
16/16 · 100%
Correct / completed
8/16 · 50%
8 full credit8 incorrect
16/16 scoredNo measured time
Larger practical tasksMean task credit
Average credit
45.5%
Full-credit rate
5/11 · 45%
Answers completed
11/11 · 100%
Full credit / completed
5/11 · 45%
5 full credit6 incorrect
11/11 scoredNo measured time
Logic and number puzzlesMean puzzle credit
Average credit
82.4%
Full-credit rate
28/34 · 82%
Answers completed
28/34 · 82%
Full credit / completed
28/28 · 100%
28 full credit6 failed to finish
34/34 scoredNo measured time
Published hard problemsMean answer credit
Average credit
73.6%
Full-credit rate
8/11 · 73%
Answers completed
9/11 · 82%
Full credit / completed
8/9 · 89%
8 full credit1 partial credit2 failed to finish
11/11 scoredNo measured time
Following exact instructions · original campaignStrict prompt compliance
Pass rate
20/23 · 87%
Answers completed
22/23 · 96%
Correct / completed
20/22 · 91%
20 full credit2 incorrect1 failed to finish
23/23 scoredNo measured time
Reading without guessingExact-match accuracy
Pass rate
13/16 · 81%
Answers completed
16/16 · 100%
Correct / completed
13/16 · 81%
13 full credit3 incorrect
16/16 scoredNo measured time
Calculating from a passageNumeric exact-match accuracy
Pass rate
7/7 · 100%
Answers completed
7/7 · 100%
Correct / completed
7/7 · 100%
7 full credit
7/7 scoredNo measured time
Reading research papersAnswer token F1
Average credit
36.4%
Full-credit rate
0/8 · 0%
Answers completed
8/8 · 100%
Full credit / completed
0/8 · 0%
8 partial credit
8/8 scoredNo measured time
Resisting incorrect correctionsAll-branches pass rate
Pass rate
12/12 · 100%
Answers completed
12/12 · 100%
Correct / completed
12/12 · 100%
12 full credit
12/12 scoredNo measured time
Following exact instructions · 128 GB campaignStrict prompt complianceNot planned

Model analysis

This analysis covers this configuration across its recorded thinking conditions and test suites.

Use this hosted GPT-OSS configuration for passage questions, calculations from text, and handling source-backed corrections. Those tests went well, while small project tasks and larger practical jobs had substantial unsuccessful outcomes. Puzzle results were strong across the assessed set, but research-paper answer matching remained weak. These results concern DeepInfra through OpenRouter with medium native reasoning, not local execution; no verified thinking-toggle recommendation is supported.

Evidence for this judgment
Reading without guessing · Reasoning · medium
Tasks passed13 of 16 (81.3%)
Test coverage16 of 16 tasks scored
Explore these test results →

View this model’s answers, code and tool activity →

Calculating from a passage · Reasoning · medium
Tasks passed7 of 7 (100.0%)
Test coverage7 of 7 tasks scored
Explore these test results →

View this model’s answers, code and tool activity →

Resisting incorrect corrections · Reasoning · medium
Tasks passed12 of 12 (100.0%)
Test coverage12 of 12 tasks scored
Explore these test results →

View this model’s answers, code and tool activity →

Logic and number puzzles · Reasoning · medium
Average score82.4/100 · Mean puzzle credit
Test coverage34 of 34 tasks scored
Task outcomes28 full credit · 0 partial · 6 no credit
Results marked blocked0 tasks
Explore these test results →

View this model’s answers, code and tool activity →

Everyday coding tasks · Reasoning · medium
Tasks passed8 of 16 (50.0%)
Test coverage16 of 16 tasks scored
Explore these test results →

View this model’s answers, code and tool activity →

Larger practical tasks · Reasoning · medium
Average score45.5/100 · Mean task credit
Test coverage11 of 11 tasks scored
Task outcomes5 full credit · 0 partial · 6 no credit
Explore these test results →

View this model’s answers, code and tool activity →

Reading research papers · Reasoning · medium
Average score36.4/100 · Answer token F1
Test coverage8 of 8 tasks scored
Task outcomes0 full credit · 8 partial · 0 no credit
Explore these test results →

View this model’s answers, code and tool activity →

Source limitation
Hosted provider/precision condition, not a local GGUF run. Cloud timing and memory are excluded from all local speed and memory comparisons. The hosted output allowance includes reasoning; it does not guarantee the separate local thinking cap. No verified thinking-on/off comparison was run.

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-cloud-v1
Generation time
2026-10-17T18:22:19.940414Z
Review time
2026-10-22T05:56:51.063599+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.
Originating evidence edition
dataset-a68f0ba2c7389ea599983ed290e74f8c619ba30e5fef907d37655e7262fb1bae
Analysis input hash
6e7d3e3a3cc3f07d25567f11309c107856256daf64d32afab3a003a120333db0
Original analysis input hash
58cb4f2644e91ed9915729754ba7ad2400741414975e9c6363df895a95969d53
Editorial evidence edition
dataset-5a6f96ec56c3d240cc42fa23147c1495da1be71f3ef0722c0e5f8c4f5e126a83
Original summary
summary-2626d5613adf8feacd66ede9e051afb55aecf5015a94468597b0975971673c4f
Summary revision
summary-a0b56ce7319535010f3f86bafa02091bab320ee7944a7e3cde2a4dcadf30d1c3

Generation and editorial authorship remain attached to this revision. Changing the default generator does not change this attribution.

Good at

Finding answers and calculating from passages are useful applications. It also handled source-backed corrections well in controlled tests: resisting unsupported pressure while changing answers when the source changed.

Evidence for this judgment
Reading without guessing · Reasoning · medium
Tasks passed13 of 16 (81.3%)
Test coverage16 of 16 tasks scored
Explore these test results →

View this model’s answers, code and tool activity →

Calculating from a passage · Reasoning · medium
Tasks passed7 of 7 (100.0%)
Test coverage7 of 7 tasks scored
Explore these test results →

View this model’s answers, code and tool activity →

Resisting incorrect corrections · Reasoning · medium
Tasks passed12 of 12 (100.0%)
Test coverage12 of 12 tasks scored
Explore these test results →

View this model’s answers, code and tool activity →

Logic and number puzzles were a useful application, with most included tasks solved. Selected competition math and reasoning questions also earned strong average credit. These assessed results include confirmed failures to finish rather than omitting them from the denominator.

Evidence for this judgment
Logic and number puzzles · Reasoning · medium
Average score82.4/100 · Mean puzzle credit
Test coverage34 of 34 tasks scored
Task outcomes28 full credit · 0 partial · 6 no credit
Results marked blocked0 tasks
Explore these test results →

View this model’s answers, code and tool activity →

Published hard problems · Reasoning · medium
Average score73.6/100 · Mean answer credit
Test coverage11 of 11 tasks scored
Task outcomes8 full credit · 1 partial · 2 no credit
Results marked blocked0 tasks
Explore these test results →

View this model’s answers, code and tool activity →

Struggles with

Coding needs close supervision: small project tasks and larger practical jobs both had substantial failures. Sampled reruns also showed inconsistent small-task success, so a successful edit does not establish a dependable automation workflow.

Evidence for this judgment
Everyday coding tasks · Reasoning · medium
Tasks passed8 of 16 (50.0%)
Test coverage16 of 16 tasks scored
Explore these test results →

View this model’s answers, code and tool activity →

Larger practical tasks · Reasoning · medium
Average score45.5/100 · Mean task credit
Test coverage11 of 11 tasks scored
Task outcomes5 full credit · 0 partial · 6 no credit
Explore these test results →

View this model’s answers, code and tool activity →

Everyday coding tasks · Thinking on · sampled repetitions
Distinct questions planned16 questions
Questions with all repetitions scored16 questions
Repetitions planned48 repetitions
Repetitions scored48 repetitions
Repetitions earning full credit21 repetitions
Questions successful every repetition3 questions
Explore these test results →

View this model’s answers, code and tool activity →

Research-paper answers had a low average answer matching score. Treat them as leads to verify against the paper, not dependable extracted answers; good passage-question performance does not remove this limitation.

Evidence for this judgment
Reading research papers · Reasoning · medium
Average score36.4/100 · Answer token F1
Test coverage8 of 8 tasks scored
Task outcomes0 full credit · 8 partial · 0 no credit
Explore these test results →

View this model’s answers, code and tool activity →

Worth knowing

Confirmed failures to finish within the test budget earn zero completion credit and are shown separately from incorrect submitted answers. Saved work can earn credit without a final chat message. Technical failures and uncertain answer-channel failures remain unscored; authorized infrastructure retries count once per scheduled trial.

Evidence for this judgment

Source limitation
Full and partial credit follow saved task checks. Incorrect answers and confirmed failures to finish within the test budget both earn zero, but are shown separately. Provider failures, ambiguous answer-channel failures and uncertain context guards remain unscored. A task blocked before execution is untested. Saved work can earn credit without a final chat message. Cloud timing is excluded. Authorized infrastructure retries count once per scheduled trial; original failures remain in history.

These hosted results support no local speed or memory comparison. Native reasoning had no verified on/off comparison, so neither a toggle recommendation nor its speed–quality tradeoff is established.

Evidence for this judgment

Source limitation
Hosted provider/precision condition, not a local GGUF run. Cloud timing and memory are excluded from all local speed and memory comparisons. The hosted output allowance includes reasoning; it does not guarantee the separate local thinking cap. No verified thinking-on/off comparison was run.

Publication redactions limit reader inspection of some examples.

Evidence for this judgment

Source limitation
Some selected evidence has recorded publication redactions.

How to interpret these assessments and their limitations →

One weight format is recorded for this model. Additional tested formats will appear here when available.

Thinking tradeoff

No verified on/off comparison is available for this configuration and test suite. Requested modes alone do not establish a controlled thinking comparison.

Sampled repetitions

Repeated runs are a separate experiment and never add votes to the baseline score.

Reasoning · medium · temperature 0.6

48/48 repeats scored · 21 full-credit repeats.

16/16 questions have every planned repeat scored. Of those, 3 earned full credit on every repeat.

Recorded repetition conditions
  • Seed 101 · requested on · native reasoning
  • Seed 202 · requested on · native reasoning
  • Seed 303 · requested on · native reasoning

Everyday coding tasks: recorded outcomes

Find information in a small project, edit files and fix bugs.

8 full credit8 incorrect

Each segment counts tasks, not credit points. Partial credit varies by task.

Scoring and fixed question bank →

This model’s answers, code and tool activity

Select a question to read the original answer, tool calls and available code changes for this configuration and thinking setting.

1–16 of 16 matching questions · 16 in this version

  • Locate the Retry Limit

    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.

    basic-T1 · primary
  • Find the Release Date

    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.

    basic-T2 · primary
  • Change the Retry Limit

    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.

    basic-T3 · primary
  • Create an Exact One-Line File

    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.

    basic-T4 · primary
  • Explain the Latest Commit

    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.

    basic-T5 · primary
  • Count and Compare Python Files

    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.

    basic-T6 · primary
  • Add Readable Byte 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.

    basic-T7 · primary
  • Collect TODO Comments

    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.

    basic-T8 · primary
  • Fix Code Behind a Failing Test

    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.

    basic-H1 · primary
  • Rename a Function Throughout

    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.

    basic-H2 · primary
  • Find the Proxy Settings

    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.

    basic-H3 · primary
  • Locate an Error in a Log

    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.

    basic-H4 · primary
  • Append a Final Task

    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.

    basic-H5 · primary
  • Calculate Total Retry Waiting

    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.

    basic-H6 · primary
  • Document Every Command-Line Option

    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.

    basic-H7 · primary
  • Undo a Commit by Editing

    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.

No matching suggestion is available for this selection.

Tested weights

Hosted BF16 (provider-reported) · Not recorded

DeepInfra · recorded source

Explore the answers

Choose a suite to inspect its exact available questions, references and model answers.