Max Blade says Gemini 4 Neon is broken in his agent CLI
Two builders posted opposite readings of the same model on the same day, and neither ran the other's test.
Max Blade says Gemini 4 Neon is completely broken in agy cli and acp. He posted what the model produced for a 3d printable fidget spinner and a lawn mow simulator, and says that output came after sending multiple follow up prompts, not on the first try. He also says xhigh is not supported.
it is completely broken in agy cli and acp??
The post was a reply to Eco, who had put up an Opus 5.5 versus Neon comparison under the line "Gemini 4 Neon is the real deal" and asked which one won. Blade's follow up to Eco was short and to the point, asking where he was running it. That question is the actual content of the disagreement. Blade is reporting a specific harness, agy cli and acp, on two generative tasks. Eco is reporting something else, and neither posted the other's setup.
The other reading
Greg Isenberg got early access and came away with the opposite impression. He calls it really good, really fast and about half the price of what he is used to, and says the benchmarks matched his experience once he saw them. His number is that Neon scored about the same as Opus 5.5 on coding for around half the cost per task. He does not name the benchmark or the dollar figure.
Isenberg also carries his own caveats, which is more than most launch day impressions do. He says Claude is still often what he reaches for on the hardest problems, and that Opus still edges Neon on a couple of coding benchmarks. His argument is a throughput one. When you run agents all day, fast and half the cost is hard to ignore, and the competition pushes prices down for everyone.
What you actually have here
Nothing in this is a head to head. Isenberg is reporting benchmark scores and a price ratio from early access. Blade is reporting one agent harness on two tasks, with screenshots of the result. Those are different kinds of evidence and they can both be true, because a model that scores well on a coding benchmark can still fall over inside a specific CLI integration that was written against something else.
For people building on it, the failure Blade is describing sits at the harness level, not obviously at the model level. The missing xhigh setting points the same way. If you are planning to swap Neon in behind an existing agent runner on the strength of a price per task figure, the thing worth testing first is whether your runner speaks to it correctly at all, before you measure anything about quality.
It scored about the same as Opus 5.5 on coding for around half the cost per task!
Know what Jev is good at before you bet on it
One email a day. Under three minutes. What builders found out about Jev yesterday.

