Summary in three lines
- When talk spread that "Claude Opus 5.5 has already been nerfed," BridgeMind released NerfBench, which re-measures performance now against launch day.
- In the first retest on September 27, Opus 5.5 scored 99.2% of its launch level and GPT-6 Astra scored 102.8%. Both were within the normal range (±10%), so there was no nerf.
- It measures performance together with tokens and cost, and anyone can see the results at bridgebench.ai/nerf-bench.
A few days after a new AI model ships, the claim that "it has already been nerfed" tends to follow. To look at the feel-based debate in numbers, BridgeMind built NerfBench, which compares launch day with the present.
On September 26 at 9:34 PM (Korea time), BridgeMind said there was a lot of nerf talk and announced it would release NerfBench the next day.
Source: X @bridgemindai
1. First result: no nerf
The results came out on September 28 at 1:10 AM (Korea time).
Source: X @bridgemindai
| Model | Launch baseline date | September 27 retest | vs. launch | Verdict |
|---|---|---|---|---|
| Claude Opus 5.5 | September 22 | 99.2% | −0.8% | Normal variation |
| GPT-6 Astra | September 3 | 102.8% | +2.8% | Normal variation |
Both models landed within the normal range of 90 to 110%. The board's verdict was "no nerf detected," and the number of models outside the normal variation range was shown as 0.
2. What NerfBench measures
NerfBench takes the value measured on launch day as that model's 100% baseline, retests the same model, and compares its current "power." Power combines three things.
- Task quality: How well it does the task
- Tokens: How many tokens it uses to do the same job
- Cost: How much one run costs
Even if answer quality stays the same, power can drop if tokens or cost go up. BridgeBench treats 90 to 110% as the normal variation range, and views anything under 90% as less power than at launch and anything over 110% as stronger. It does not rule on a drop in a single task, but looks at the change across tasks spanning many areas.
The official post explains that it does not publish the tasks and weights, to prevent "benchmaxxing," which means pushing up only benchmark scores.
| Item | Launch day | Retest |
|---|---|---|
| Task score | 0.9 | 0.72 |
| Tokens | 10,000 | 14,000 |
| Cost | $0.10 | $0.13 |
| power | 100% | About 77.7% (lost power) |
The numbers are an example the official post gave for explanation.
3. Where the "it got nerfed" talk came from
On September 27 at 12:18 AM (Korea time), one developer posted that Opus 5.5 did not feel like it did in the first two days.
Source: X @DevDacian
There was also a post (September 25, 8:04 PM, @batuhan4) saying the author's own SWE Pro benchmark score had fallen from 83% to 71%. NerfBench is an attempt to re-measure these gut-feel differences with the same tasks.
4. BridgeBench and BridgeMind
BridgeBench is an AI coding model benchmark site built by BridgeMind, and it runs a model leaderboard, Design Bench, and Nerf Bench. BridgeMind describes itself as "the home of the vibe coding movement," and its founder is @matthewmillerai. It also offers an agent app for Mac, Windows, and Linux.
On its separate coding leaderboard, as of September 22, Claude Opus 5.5 was first with 760 points and GPT-6 Astra second with 734 points (25 models). Those are coding leaderboard scores, a different metric from NerfBench power, which shows a ratio against launch.
5. More models to be retested
BridgeMind said it will spend $10,000 over the next few months to keep retesting models, not just on launch day but afterward. In the results post, it hinted at more models and more frequent retests and asked, "Which model should we retest next?"
6. How to see the results board yourself
- Open bridgebench.ai/nerf-bench.
- Click a model to see its full history by date.
- You can download a result as a 10-second 1080p MP4 or post it straight to X.
- You can subscribe to get new results by email.
If Opus 5.5 feels off lately, do not rely on feel alone. Check the history by date on the NerfBench board first.