OpenAI admitted what users suspected: they throttled GPT-5.2’s reasoning capabilities.
The February 4, 2026 release notes quietly stated: “We’re restoring the Extended thinking level for GPT-5.2 Thinking to its prior setting, correcting the inadvertent reduction from January.”
Translation: we slowed it down for cost reasons, you noticed, and we’re rolling it back.
What Happened
In mid-January 2026, users noticed GPT-5.2’s “Extended” thinking mode was producing faster but lower-quality responses. Complex reasoning tasks that previously required 30-60 seconds of processing were completing in 10-15 seconds — with noticeably degraded results.
The change wasn’t announced. Users discovered it through degraded performance on established workflows.
By early February, the backlash was significant enough that OpenAI reversed course. The restoration appeared in the February 4 release notes — an admission that the reduction had been a mistake.
The Cost Pressure
Extended thinking is expensive. Each reasoning token requires multiple forward passes through the model. Longer reasoning chains multiply compute costs.
OpenAI faces a margin squeeze:
- $20/month ChatGPT Plus subscriptions don’t cover heavy reasoning usage
- Enterprise contracts have usage limits that extended thinking rapidly consumes
- API pricing for reasoning models is already premium
The January reduction was a cost-cutting measure disguised as optimization.
User Backlash
The AI community noticed immediately:
- Coding workflows broke — Complex debugging that previously worked started failing
- Research quality declined — Literature reviews missed key connections
- Math accuracy dropped — Competition math scores fell measurably
Power users — the exact demographic OpenAI needs to maintain market position — were loudest in their complaints.
The Reversal
OpenAI’s rollback is unusual. The company typically defends product changes even when unpopular. The speed of this reversal suggests:
- Internal metrics confirmed quality degradation — Not just user complaints, but measurable performance drops
- Enterprise pressure — Paying customers threatened contract cancellations
- Competitive threat — Anthropic’s Claude maintained reasoning quality during the same period
The Deeper Problem
This episode reveals a fundamental tension in AI product management: the gap between demo capability and operational cost.
GPT-5.2 was trained and benchmarked with unlimited reasoning time. The model’s impressive results assume optimal compute allocation. But operational reality requires tradeoffs between quality, speed, and cost.
OpenAI’s challenge: the model they sold isn’t economically viable at current pricing.
What This Means for Users
The restoration is temporary relief, not long-term stability. OpenAI will face the same cost pressures again.
Users should expect:
- Pricing increases for reasoning-heavy features
- Usage limits on extended thinking
- Tier differentiation between Plus, Pro, and Enterprise reasoning access
The free lunch — unlimited frontier AI for $20/month — is ending.
The Competitive Angle
Anthropic’s Claude avoided this specific problem by architecting for efficiency from the start. Claude Opus 4.6 delivers comparable reasoning without the same cost escalation.
Google’s Gemini took a different approach: less capable reasoning, but consistent performance at scale.
OpenAI’s position as “best but expensive” is defensible in the short term. But if quality degradation becomes a pattern, users will migrate to “good enough and reliable” alternatives.
The Lesson
The January-February 2026 episode teaches a lesson about AI infrastructure: capabilities demonstrated in benchmarks aren’t the same as capabilities delivered at scale.
OpenAI can build models that reason brilliantly. The question is whether they can afford to run them.