Glowing Spoon

Industry / Models / OpenAI

Codex

An engineering agent backed by OpenAI models for repository work, tool execution, and reviewable software changes.

Model snapshot
GPT-6 Astra benchmark snapshot; GPT-6.1 Sol also available in Codex
Deployment
Codex CLI, desktop and IDE surfaces; model access also through the API
Context & outputs
Model-dependent: GPT-6 Astra API 1,050,000 tokens; GPT-6.1 Sol in Codex defaults to 500,000
Access & license
Proprietary models and service

The agent and the model are different layers

Codex provides the engineering workflow; the selected model supplies reasoning and generation. OpenAI documents GPT-6 Astra for demanding work and GPT-6.1 Sol for repeated, long-running work. A benchmark for one underlying model is not a score for every Codex configuration. Official OpenAI model guidance.

Technical capabilities

GPT-6 Astra supports text input/output and image input, with a 1,050,000-token API context and a maximum output of 128,000 tokens. Its API supports reasoning tokens. GPT-6.1 Sol’s default Codex context is a separate, surface-specific limit. OpenAI API specifications.

Our engineering approach

We give the agent a bounded specification, repository instructions, and explicit verification commands. It produces changes that can be inspected as a diff, exercises the affected behavior, and records validation evidence. We separate permission to inspect, permission to modify, and permission to deploy. Acceptance is based on the resulting software and tests, with a human release decision.

Published evidence

Benchmark snapshot.

Codex is an agent, not one model. These scores describe GPT-6 Astra in Google’s October 2026 table, not Codex as a whole or GPT-6.1 Sol. Terminal-Bench is the Artificial Analysis result; other benchmark conditions differ.

Published results for GPT-6 Astra benchmark snapshot; GPT-6.1 Sol also available in Codex; higher is better within each metric. Results have different scopes and are not a universal ranking.
Parameter / benchmarkResultEvaluation settingSource
Software engineeringFrontierSWE 265.5%GPT-6 Astra; public leaderboard resultPublished result
Terminal agentsTerminal-Bench 4.0 (AA)59.1%GPT-6 Astra; Artificial Analysis resultPublished result
Science and mathRiemannBench72.0%GPT-6 Astra; published comparisonPublished result

These are published results, not Glowing Spoon measurements. Benchmark percentages, F1 scores, and Elo ratings use different scales. Production evaluation also includes latency, cost, reliability, and human review effort on your workload.

Sources & methodology