November 5, 2025
Team Speed vs Team Think
Developers are choosing older AI models, and the data explains why
Dev drama: “Old” models win on speed, GPT‑5 roasted for lag, locals feel snubbed
TLDR: Developers are picking models by task, not version—older Sonnet 4.0 rose while 4.5 fell, and GPT‑5 held steady. The comments erupted: speed lovers dunked on GPT‑5’s lag, pragmatists said “stick with what works,” and local‑model fans called the data biased—proof that AI upgrades are now sideways, not up.
The plot twist no one saw coming: developers aren’t chasing the newest AI toy. According to Augment Code’s data, Sonnet 4.5 usage dipped while older Sonnet 4.0 climbed, and GPT‑5 just sat there, steady. The internet promptly split into camps. One side is Team What Works—users like rcarmo who say constant model hopping is “cognitive overload.” Translation: don’t make me relearn my job every other week.
Then there’s Team Big Brain, led by fans like KronisLV, swearing Sonnet 4.5 thinks deeper and handles complex stuff. But the loudest chorus? The roast of GPT‑5’s speed. gptfiveslow brought the house down: “HELLISHLY slow.” Memes followed: “GPT‑5 = Go Prepare Tea,” and “I asked for one nginx line and got a TED Talk.”
The data says 4.5 writes more and calls tools less—more thinking, fewer actions—while 4.0 is snappier and predictable. Devs translated that into real life: Team Speed vs Team Think. Meanwhile, skeptics like Manfred question the whole sample—if we don’t know the customer mix, how universal is this? And local‑model die‑hards felt erased, calling out the post for ignoring free/self‑hosted tools. Bottom line: upgrades aren’t “better,” they’re different, and the community is loudly choosing by job, not hype.
Key Points
- •Model adoption is fragmenting: Sonnet 4.5’s share fell while Sonnet 4.0’s rose during the first week of Oct 2025; GPT-5 remained ~10–12%.
- •Sonnet 4.5 produces larger outputs with fewer tool calls, implying deeper internal reasoning; Sonnet 4.0 makes more frequent tool calls for faster execution.
- •Average output per message: Sonnet 4.5 ~7.5k tokens vs. Sonnet 4.0 ~5.5k tokens, a ~37% increase; Sonnet 4.5 shows slightly lower throughput (qualitative).
- •Compute/cache data (several billion tokens sample): Sonnet 4.5 has roughly one-third more cache reads than Sonnet 4.0; GPT-5 has a lighter footprint.
- •Early specialization: Sonnet 4.5 for long-context and complex refactoring; Sonnet 4.0 for deterministic, structured tasks; GPT-5 for explanatory and hybrid code-doc tasks.