OpenAI has given away the engine under Codex, and its own numbers say the engine is worth more than the model
On 20 August OpenAI published the harness beneath Codex — the piece that actually runs an agent — under an Apache-2.0 licence. The command-line tool was already open; what is new is the Codex SDK and the app-server, a JSON-RPC engine that holds a session open, and the invitation to point them at things that are not code. The harness handles the loop nobody sees: understanding the task, keeping memory across a long conversation, streaming events, calling tools, letting you interrupt, syncing state, and stopping to ask a human for approval. OpenAI's own measurements are the striking part. Changing only the harness around GPT-5.6 Sol — keeping reasoning between turns and compressing context — moved its ARC-AGI-3 score from 13.3 percent to 38.3 percent and cut token use to a sixth. The same model. One caveat belongs next to that figure: the ARC Prize organisers, running the model their own vendor-neutral way, scored it 7.8 percent, so the 38.3 is OpenAI's harness on OpenAI's terms. Greg Brockman's line for it was that Codex can drive far more than coding tools. Cisco has built an App Builder on the SDK, Thrive Holdings has deployed agents on it, and a tax pilot ran 7,000 returns through it and cut preparation time by roughly a third. The repository is at 112,000 stars and shipping alpha builds daily.
Related stories
A voice model that starts speaking in under 50 milliseconds, and the serving code is open2026-08-22Give a model the reference list of an unpublished paper and ask what the paper says. The best ones get it 15 percent of the time.
2026-08-22Anthropic will let its strongest model hunt flaws in your code — but it will not let you talk to it
2026-08-22The model on its own scored 30 percent. Wrapped in Nvidia's scaffolding, the same model cleared everything.
2026-08-22Apple Music will tell you when a song was made by a machine — but the uploader decides whether to say so
2026-08-21