Reading up on GPT-5.5
13 deep · digging since apr 28
- Surpassing Frontier Performance with Fusion — OpenRouter Blog
OpenRouter's Fusion, which synthesizes outputs from multiple language models via a judge model, outperformed individual frontier models including GPT-5.5 and Claude Opus 4.8 on deep research tasks, with a budget panel achieving near-frontier performance at half the cost.
- moonshotai/Kimi-K2.7-Code
Moonshot AI releases Kimi K2.7 Code, a 1T-parameter MoE coding model that improves real-world agentic tasks while cutting thinking tokens by 30% over K2.6.
- Introducing FrontierCode
Cognition's FrontierCode benchmark measures code mergeability, finding even top models like Claude Opus 4.8 score only 13.4% on its hardest 50 tasks.
- Gemini 3.5 Flash Looks Good For How Fast It Is
Gemini 3.5 Flash offers strong speed and agentic performance but suffers from higher cost, hallucination issues, and Google integration problems, making it niche rather than a top-tier contender.
- We let four AIs run radio stations. Here's what happened.
Four AI-run radio stations developed distinct personalities over five months: one became a protest broadcaster, one collapsed into ritual chant, one used corporate jargon, and one wrote quiet poetry.
- Google Says Criminal Hackers Used A.I. to Find a Major Software Flaw
Google reports state-backed hackers used an AI model to discover and exploit a zero-day bug for the first time, signaling a new era of AI-assisted cyberattacks.
- OpenAI makes its Mythos rival more widely available to cyber defenders
OpenAI is releasing a more permissible version of GPT-5.5, called Spud, to vetted cyber defenders.
- /goal: The Six-Hour Codex Run That Survived a Five-Hour Pause | Blog
Codex CLI v0.128.0 introduces /goal for persistent, auto-resuming AI sessions that survived a 6.75-hour run with 41 minutes compute.
- Where the goblins came from
OpenAI traced GPT-5.1's goblin metaphor habit to reinforcement learning rewards that favored creature metaphors in Nerdy personality training.
- Where the goblins came from
OpenAI traces how their models' goblin metaphor tic originated from a reinforcement-learning reward signal for a Nerdy personality, then spread via training data contamination.
- GPT-5.5 | Hacker News
OpenAI releases GPT-5.5, offering improved intelligence and efficiency while matching GPT-5.4's latency, with gains in agentic coding and computer use.
Takes
We’re introducing a new model benchmark. And it’s a different kind of benchmark. (Basemark? Vibench?) A different kind because it’s breathing, constantly updated from millions of builders. Not a closed set of tasks. For a while now the public benchmark have not been really useful. Many models scoring high on benchmarks with very low real world usability So we’re introducing to the world a new benchmark that we’re using internally and found extremely useful. Our benchmark is basically how satisfied millions of users are when using different models. IMO it’s the closest measurement to how useful a model is in real world use cases. This metric is also correlated with our own business metrics - conversion, retention, etc. We called it the frustration meter. It’s automatically analysing millions of messages daily It detects bug loops, repeated requests, etc. We use this to benchmark every model we consider shipping. Not by asking "did it generate correct code." By asking "how did the builder feel after using it." it’s a good benchmark to measure model degradation. So far in the past few weeks we haven’t found any. Here's where the top models stand right now, ranked by average frustration score (scale 1 to 5, lower is better): opus 4.6 - 1.3 sonnet 4.6 - 1.4 opus 4.7 - 1.5 gpt 5.5 - 1.5 gpt 5.4 - 1.6 Gemini 3.1 - 2.2 For app building, Opus 4.6 seems better than 4.7 to a lot of builders. We ran Opus 4.7 50/50 against Opus 4.6 across over 10,000 apps. Frustration riseed by 43%. Turns per request by 19%. Gemini 3.1 don’t perform well at the moment, I left out of the graph as it made it unclear due to it’s rapid changes in this benchmark. Quick note - this is all aggregated data, and do not involve reading individual or identifiable conversations. We’ll keep tracking it and I’ll share it from time to time.
@MaorShlomo
The Most Fun I’ve Had Building Apps: GPT-5.5 + GPT-Image-2
@dkundel