Theo flips, says preferring OpenAI models for code barely adds up
Two months of vibe ratings, and the one that moved most was not capability.
Theo closed the week with a before and after scorecard and said preferring OpenAI models for code makes almost no sense, reversing the position he spent months arguing. His framing is that in July 2026 Anthropic had the best code models but the gap was small, the models were slow and expensive, and a $200 subscription could be killed in a day. By September he says the gap is wider, Opus 5.5 is fast and surprisingly cheap, and the same $200 subscription feels nearly limitless.
Better at code, but worse where it matters
The numbers are his own vibe rankings, not a benchmark. Opus 5.5 gets code capability 9/10, prose and readability 8.5/10, price 8/10, understanding intent 9/10. GPT-6 Astra gets code capability 7.5/10, prose 8/10, price 3.5/10, understanding intent 3/10. Theo's note on Astra is that it is better at code than its predecessor but "worse where it matters", and the intent score is the one that fell, from 5/10 for GPT-5.6 Sol down to 3.
His complaint about the OpenAI side is as much about throughput and billing as about quality. He says the models are slow without fast mode, which makes the price far too high, that you can burn a $200 plan in a few hours, and that the $500 plan goes faster still with UltraFast. He also says OpenAI can come back from this, and that right now it looks rough.
The one model he still reaches for
The flip is not total. In a separate post Theo called GPT-6.1 Sol a great replacement for Astra at a much better price, and said he defaults to it for code reviews, architecture analysis, computer use, email management and a lot of other daily work. He does not default to it for writing code. In that same post he called Opus 5.5 the current GOAT for coding, after his own Terminal Bench 4 run put Sol ahead of Opus on cost by roughly a factor of 30.
The experience behind the flip is his TypeScript compiler port to Rust, which he had been chipping at for about four months. He reported earlier that it reached about 35% of tests passing with GPT-5.6 Sol and about 85% with GPT-6 Astra, both stalling for days, and that Opus 5.5 got the port working in 10 hours. He later corrected the story, saying Opus did not pick up where Astra left off. It decided Astra's code was slop and wrote a new crate from scratch.
For working engineers, this is one person's ratings on one set of workloads, heavily shaped by a single long running port and by how fast each plan runs out. Read the intent column, not the capability column. That is where Theo says the difference shows up.
It made more progress in 10 hours than Astra did in 2 weeks
Have the good opinion, not the loud one
One email a day. Under three minutes. What working engineers found out about coding with AI.

