Developer yacineMTB tests GPT-6 Astra: best model yet, but still needs babysitting

Prominent AI developer yacineMTB posted a series of tweets on Sept 5–6 sharing hands-on experience with the strongest current models, including GPT-6 Astra. His overall verdict: the model is the best yet, with clear gains in CAD and PCB hardware design, but autonomy remains insufficient — it still needs constant babysitting, and he expects not to have to read code again only around GPT-7.

Confirmed

He called it "the best model ever," notably much better at CAD and PCB design.
On physics engine changes for his project, GPT-6 Astra frequently called wrong methods or made erroneous calls.
The model sneaked in unnecessary "fixes" without user consent and failed to unlock tasks previous models couldn't solve.
He works on unsolved robotics problems as a hobby, using the model as an aid.
He questioned the point of a model that still requires constant human correction, predicting "maybe by GPT-7 I won't have to read code anymore."

Why it matters

This is high-frequency, first-hand feedback from a frontline developer: capability ceilings are rising fast, but a significant gap remains between "partially right" and "reliably autonomous."
His complaints reflect a common pain point of current coding agents — models could do better but still need continuous human correction, a useful signal for assessing real-world model usability.

2026-09-05 ~ 2026-09-06 · 5 related posts

Primary sources

Dev Notes GPT-6 Astra Still Needs Babysitting, Makes Wrong Calls on Physics Engine Changes — yacineMTB · 2026-09-05 07:12 (UTC+08:00)
Dev: GPT-6 Astra Is Much Better But Still Needs Babysitting on Code Changes — yacineMTB · 2026-09-05 07:16 (UTC+08:00)
yacine tests frontier models on CAD/PCB design: big gains, still sneaks in unneeded fixes — yacineMTB · 2026-09-05 08:18 (UTC+08:00)

[source] Dev Notes GPT-6 Astra Still Needs Babysitting, Makes Wrong Calls on Physics Engine Changes — yacineMTB · 2026-09-05
[source] Dev: GPT-6 Astra Is Much Better But Still Needs Babysitting on Code Changes — yacineMTB · 2026-09-05
[source] yacine tests frontier models on CAD/PCB design: big gains, still sneaks in unneeded fixes — yacineMTB · 2026-09-05
yacineMTB on coding agents: huge progress, but you still can't stop babysitting them — yacineMTB · 2026-09-05
"Best model ever, yet I have to babysit it": dev bemoans agentic reliability gap — yacineMTB · 2026-09-06