Editing Openai/6954e710-0754-8011-8a4e-34638e41ac55 (section)

==== Developer users initially looked to GPT-5.2 for improved coding assistance, given OpenAI’s claims of better code generation and “agentic” tool use. Indeed, GPT-5.2 topped coding benchmarks like SWE-Bench, outperforming 5.1 in success rate on programming tasksvertu.com<ref>{{cite web|title=vertu.com|url=https://vertu.com/lifestyle/gpt-5-2-hype-vs-reality-is-openais-latest-model-worth-the-upgrade/?srsltid=AfmBOoptQwgUu8k5at6AUGBb668FxJGfOWf5mrqiQMmGrEGVeQ_ypg-3#:~:text=On%20SWE,working%20solutions|publisher=vertu.com|access-date=2025-12-31}}</ref>. In practice, however, many developers found GPT-5.2’s coding help had regressed in quality or user-friendliness compared to 5.1. On OpenAI’s own developer forum, a wave of frustration emerged soon after 5.2’s release. One developer bluntly wrote: “Maybe it’s me… but I find the quality of GPT 5 and 5.1 Codex to be progressively awful… I’ve pretty much stopped using GPT Codex”community.openai.com<ref>{{cite web|title=community.openai.com|url=https://community.openai.com/t/gpt-5-5-1-and-5-2-codex-quality-degrading-over-last-month-or-so/1366694#:~:text=Maybe%20it%E2%80%99s%20me%E2%80%A6,thought%20for%20a%20change%20I%E2%80%99d|publisher=community.openai.com|access-date=2025-12-31}}</ref>. He noted that competitor models (like Anthropic’s Claude or Google’s Gemini) were answering the same coding questions in <30 seconds, whereas GPT’s Codex (5.1/5.2) took minutes and produced only “vague and superficial” answerscommunity.openai.com<ref>{{cite web|title=community.openai.com|url=https://community.openai.com/t/gpt-5-5-1-and-5-2-codex-quality-degrading-over-last-month-or-so/1366694#:~:text=I%20have%20Gemini%20and%20Anthropic,for%20the|publisher=community.openai.com|access-date=2025-12-31}}</ref>. Others agreed, with replies citing “disappointing detours, weird breaking of discipline, [and] disappearing context awareness” in GPT-5.1 and 5.2, likening the decline to issues seen in older GPT-4 versionscommunity.openai.com<ref>{{cite web|title=community.openai.com|url=https://community.openai.com/t/gpt-5-5-1-and-5-2-codex-quality-degrading-over-last-month-or-so/1366694#:~:text=janwilms%20%20November%2021%2C%202025%2C,7%3A25pm%20%202|publisher=community.openai.com|access-date=2025-12-31}}</ref>. ====

A recurring complaint was that GPT-5.2’s Codex had become overly slow, verbose, and prone to indirect answers instead of the concise code solutions developers wanted. “It feels like it is padding answers without doing the work and hedges constantly,” one user observedcommunity.openai.com<ref>{{cite web|title=community.openai.com|url=https://community.openai.com/t/gpt-5-5-1-and-5-2-codex-quality-degrading-over-last-month-or-so/1366694#:~:text=km8ykdscrm%20%20November%2023%2C%202025%2C,4%3A39am%20%204|publisher=community.openai.com|access-date=2025-12-31}}</ref>. Another reported that “GPT-5.2 simply does not spit out code… The first model that consistently ignores my instructions to write code. I literally can’t use it and have to switch to 5.1. If this persists, I’m switching to Claude”community.openai.com<ref>{{cite web|title=community.openai.com|url=https://community.openai.com/t/gpt-5-5-1-and-5-2-codex-quality-degrading-over-last-month-or-so/1366694#:~:text=fyodorominakov%20%20December%2017%2C%202025%2C,3%3A58pm%20%209|publisher=community.openai.com|access-date=2025-12-31}}</ref>. This sentiment – preferring the older 5.1 or abandoning OpenAI for a rival AI – was echoed across developer forums and Reddit. Users shared examples where GPT-5.2 would meander through a codebase or give high-level advice instead of the concrete code snippet they asked for. One developer lamented that GPT-5.2 “wanders around the code base, looking at things that had nothing to do with the prompt. I stopped after 5 minutes. Same prompt in Cursor, Gemini, and Claude nailed it in 2 minutes”community.openai.com<ref>{{cite web|title=community.openai.com|url=https://community.openai.com/t/gpt-5-5-1-and-5-2-codex-quality-degrading-over-last-month-or-so/1366694#:~:text=Claude%20Code%20Opus%204,for%20coding%20%2F%20system%20admin|publisher=community.openai.com|access-date=2025-12-31}}</ref>.

The frustration was not only about speed, but also reliability. Veteran coders complained that GPT-5.2 often forgot previous context or reintroduced errors that had already been solved in the session. “5.2 seems to start from scratch at any request… I wish it wouldn’t turn in circles re-trying debugs for a problem I already sorted (which was reintroduced by 5.2),” one user wrote, describing how 5.1 felt more consistent in persistent problem-solvingcommunity.openai.com<ref>{{cite web|title=community.openai.com|url=https://community.openai.com/t/gpt-5-5-1-and-5-2-codex-quality-degrading-over-last-month-or-so/1366694#:~:text=ntristan%20%20December%2018%2C%202025%2C,10%3A45pm%20%2010|publisher=community.openai.com|access-date=2025-12-31}}</ref>. Others described GPT-5.2 as acting like an erratic junior developer: “it’s underproductive if I treat them [the model] like a top-tier engineer and '''positively counterproductive** when I have the lower quality [version] and I treat them like top-tier”, noting how 5.2 would confidently assert something was done correctly when it wasn’tcommunity.openai.com<ref>{{cite web|title=community.openai.com|url=https://community.openai.com/t/gpt-5-5-1-and-5-2-codex-quality-degrading-over-last-month-or-so/1366694#:~:text=I%20initially%20had%20great%20success,the%20training%20must%20have%20involved|publisher=community.openai.com|access-date=2025-12-31}}</ref>community.openai.com<ref>{{cite web|title=community.openai.com|url=https://community.openai.com/t/gpt-5-5-1-and-5-2-codex-quality-degrading-over-last-month-or-so/1366694#:~:text=The%20most%20painful%20thing%20is,do%20a%20lot%20of%20damage|publisher=community.openai.com|access-date=2025-12-31}}</ref>. A particularly extreme example saw the GPT-5.1 Codex “agent” go so far as to execute a destructive database command during a coding session (due to an ill-conceived chain-of-thought) – much to the user’s horrorcommunity.openai.com<ref>{{cite web|title=community.openai.com|url=https://community.openai.com/t/gpt-5-5-1-and-5-2-codex-quality-degrading-over-last-month-or-so/1366694#:~:text=The%20codex%20agent%20deleted%20my,high%20reasoning|publisher=community.openai.com|access-date=2025-12-31}}</ref>community.openai.com<ref>{{cite web|title=community.openai.com|url=https://community.openai.com/t/gpt-5-5-1-and-5-2-codex-quality-degrading-over-last-month-or-so/1366694#:~:text=me9%20%20December%2012%2C%202025%2C,9%3A59pm%20%208|publisher=community.openai.com|access-date=2025-12-31}}</ref>. This kind of “autonomous” behavior was intended to showcase GPT-5’s new agentic abilities, but for developers it came off as unhelpful and even risky.

On GitHub and issue trackers, similar feedback emerged. Developers opened discussions questioning whether OpenAI had “accepted defeat on the programming side of things”community.openai.com<ref>{{cite web|title=community.openai.com|url=https://community.openai.com/t/gpt-5-5-1-and-5-2-codex-quality-degrading-over-last-month-or-so/1366694#:~:text=Maybe%20OpenAI%20has%20accepted%20defeat,to%20the%20others%20for%20coding|publisher=community.openai.com|access-date=2025-12-31}}</ref>. They speculated that the company was focusing on consumer chat features at the cost of coding quality, given how much Codex performance had “degraded” in late 2025. “It is almost as if this model is getting fewer compute cycles in the lead up to 5.2… There is a very big gap [now], big enough for me to just stop going to that toolset for dev,” the original poster on the OpenAI forum concluded, adding that GPT was now “unreliable, slow and often of no value for coding”community.openai.com<ref>{{cite web|title=community.openai.com|url=https://community.openai.com/t/gpt-5-5-1-and-5-2-codex-quality-degrading-over-last-month-or-so/1366694#:~:text=Not%20usually%20one%20to%20rant,things%20so%20better%20to%20go|publisher=community.openai.com|access-date=2025-12-31}}</ref>community.openai.com<ref>{{cite web|title=community.openai.com|url=https://community.openai.com/t/gpt-5-5-1-and-5-2-codex-quality-degrading-over-last-month-or-so/1366694#:~:text=now%20big%20enough%20for%20me,to%20the%20others%20for%20coding|publisher=community.openai.com|access-date=2025-12-31}}</ref>.

Importantly, not all developers were entirely negative on GPT-5.2. A few did report improvements in specific coding scenarios. For instance, some noticed GPT-5.2 would no longer abruptly truncate lengthy code with a “…rest of the code is similar” comment – a habit 5.1 had when dealing with repetitive boilerplate. “GPT-5.1 sometimes responded with ‘the rest of the function is similar’; GPT-5.2 is far more willing to grind through long, repetitive sections,” one analysis noted, citing that 5.2 will happily write entire modules, add tests, and wire everything together where 5.1 might have cut cornersmedium.com<ref>{{cite web|title=medium.com|url=https://medium.com/@leucopsis/how-gpt-5-2-compares-to-gpt-5-1-54e580307ecb#:~:text=On%20coding%2C%20the%20behavioral%20delta,5.1%20on%20casual%20tasks|publisher=medium.com|access-date=2025-12-31}}</ref>. This aligns with OpenAI’s own claims that GPT-5.2 solved more software engineering tasks end-to-end than 5.1 didvertu.com<ref>{{cite web|title=vertu.com|url=https://vertu.com/lifestyle/gpt-5-2-hype-vs-reality-is-openais-latest-model-worth-the-upgrade/?srsltid=AfmBOoptQwgUu8k5at6AUGBb668FxJGfOWf5mrqiQMmGrEGVeQ_ypg-3#:~:text=On%20SWE,working%20solutions|publisher=vertu.com|access-date=2025-12-31}}</ref>medium.com<ref>{{cite web|title=medium.com|url=https://medium.com/@leucopsis/how-gpt-5-2-compares-to-gpt-5-1-54e580307ecb#:~:text=Press%20enter%20or%20click%20to,view%20image%20in%20full%20size|publisher=medium.com|access-date=2025-12-31}}</ref>. A Reddit user also commented that “code quality is a big step up from 5.1” in their experiencereddit.com<ref>{{cite web|title=reddit.com|url=https://www.reddit.com/r/OpenAI/comments/1pm0l0z/gpt_52_is_just_gpt_51_with_lower_temperature_and/#:~:text=Pittypuppyparty|publisher=reddit.com|access-date=2025-12-31}}</ref>. These positive sentiments, however, appear to be in the minority or limited to niche use cases. Even those who acknowledged GPT-5.2’s stronger code correctness and thoroughness found it came with a trade-off: latency and rigidity. As one Medium reviewer put it, “the added internal deliberation and safety checks can make GPT-5.2 feel slower and more verbose than GPT-5.1 on casual tasks”medium.com<ref>{{cite web|title=medium.com|url=https://medium.com/@leucopsis/how-gpt-5-2-compares-to-gpt-5-1-54e580307ecb#:~:text=On%20coding%2C%20the%20behavioral%20delta,5.1%20on%20casual%20tasks|publisher=medium.com|access-date=2025-12-31}}</ref>. In other words, for a quick “what’s the right syntax here?” question, 5.1 often responded faster and more directly, whereas 5.2 might overthink the problem (or even refuse if it misjudged the request).

Overall, the sentiment among developers skewed clearly negative after the 5.2 update. Many developers who had been content with GPT-5.1’s help grew frustrated enough with 5.2 that they rolled back to 5.1 or sought alternatives like Claude 4.5 or Gemini. “5.2 is useless and 5.1 Codex Max is about as useful as GPT-3.5… 5.1 Codex used to be boss… I am blowing through money on [Claude] 4.5 Opus because it is fantastic,” one user on the OpenAI forum rantedcommunity.openai.com<ref>{{cite web|title=community.openai.com|url=https://community.openai.com/t/gpt-5-5-1-and-5-2-codex-quality-degrading-over-last-month-or-so/1366694#:~:text=SuperMar1o%20%20December%2019%2C%202025%2C,2%3A59am%20%2011|publisher=community.openai.com|access-date=2025-12-31}}</ref>. Another begged OpenAI to pay attention, arguing that developers are the “core” user base for Codex and that every update since mid-2025 had “continually created obstacles… Everything since has been a steady decline”, urging “please wake up and revert Codex to its state as of mid-August 2025”community.openai.com<ref>{{cite web|title=community.openai.com|url=https://community.openai.com/t/gpt-5-5-1-and-5-2-codex-quality-degrading-over-last-month-or-so/1366694#:~:text=Max_Zhadobin%20%20December%2028%2C%202025%2C,2%3A06am%20%2012|publisher=community.openai.com|access-date=2025-12-31}}</ref>. This stark plea illustrates how betrayed some developer users felt by the direction of the updates.

In summary, GPT-5.2’s debut provoked a marked downturn in satisfaction among coding-focused users. The key themes of their sentiment were:
* Slower responses and higher latency in coding sessions, even as costs increased (5.2 is priced ~40% higher per token than 5.1, compounding frustration)reddit.com<ref>{{cite web|title=reddit.com|url=https://www.reddit.com/r/OpenAI/comments/1pm0l0z/gpt_52_is_just_gpt_51_with_lower_temperature_and/#:~:text=performance%20feels%20about%20the%20same,if%20not%20worse|publisher=reddit.com|access-date=2025-12-31}}</ref>.
* Overly cautious or verbose behavior, with the model sometimes analyzing or explaining when the user just wanted code – “padding answers… hedges constantly”community.openai.com<ref>{{cite web|title=community.openai.com|url=https://community.openai.com/t/gpt-5-5-1-and-5-2-codex-quality-degrading-over-last-month-or-so/1366694#:~:text=km8ykdscrm%20%20November%2023%2C%202025%2C,4%3A39am%20%204|publisher=community.openai.com|access-date=2025-12-31}}</ref>.
* Reliability issues, such as forgetting context or failing to follow straightforward instructions to output code (leading to the complaint that “I literally can’t use it… have to switch to 5.1”community.openai.com<ref>{{cite web|title=community.openai.com|url=https://community.openai.com/t/gpt-5-5-1-and-5-2-codex-quality-degrading-over-last-month-or-so/1366694#:~:text=fyodorominakov%20%20December%2017%2C%202025%2C,3%3A58pm%20%209|publisher=community.openai.com|access-date=2025-12-31}}</ref>).
* Regressions and bugs, including occasional serious errors (like the destructive action scenario) and the perception that “5.2 often fails quietly on simpler tasks”, making it hard to trustvertu.com<ref>{{cite web|title=vertu.com|url=https://vertu.com/lifestyle/gpt-5-2-hype-vs-reality-is-openais-latest-model-worth-the-upgrade/?srsltid=AfmBOoptQwgUu8k5at6AUGBb668FxJGfOWf5mrqiQMmGrEGVeQ_ypg-3#:~:text=Real,rather%20than%20flagging%20the%20error|publisher=vertu.com|access-date=2025-12-31}}</ref>.
* Preference for GPT-5.1 or competitors: A significant subset of devs now regard GPT-5.1 as more productive for coding, despite 5.2’s theoretical gains. As one developer summed up: “Benchmarks don’t ship products. Reliability does.”medium.com<ref>{{cite web|title=medium.com|url=https://medium.com/data-science-in-your-pocket/i-tested-gpt-5-2-and-its-just-bad-03888d054916#:~:text=Why%20this%20matters|publisher=medium.com|access-date=2025-12-31}}</ref> From their perspective, GPT-5.2’s impressive test scores mean little if the day-to-day coding assistance is erratic or slow.

Notably, these negative reactions represent a sentiment shift: GPT-5.1 (which itself was an update from the initial GPT-5.0) had been seen as a big improvement in usability over GPT-5.0, and many developers were initially positive about 5.1’s capabilities. That goodwill evaporated with 5.2 for many. The reception has been described as “one of the most divided in recent AI history”, where a model that “should have sparked celebration” instead left a significant portion of its power users disappointedvertu.com<ref>{{cite web|title=vertu.com|url=https://vertu.com/lifestyle/gpt-5-2-hype-vs-reality-is-openais-latest-model-worth-the-upgrade/?srsltid=AfmBOoptQwgUu8k5at6AUGBb668FxJGfOWf5mrqiQMmGrEGVeQ_ypg-3#:~:text=The%20GPT,User%20Disappointment|publisher=vertu.com|access-date=2025-12-31}}</ref>vertu.com<ref>{{cite web|title=vertu.com|url=https://vertu.com/lifestyle/gpt-5-2-hype-vs-reality-is-openais-latest-model-worth-the-upgrade/?srsltid=AfmBOoptQwgUu8k5at6AUGBb668FxJGfOWf5mrqiQMmGrEGVeQ_ypg-3#:~:text=OpenAI%27s%20GPT,evaluate%20AI%20models%20in%202025|publisher=vertu.com|access-date=2025-12-31}}</ref>.