Impact of Kimi K2 and Qwen 3 Coder on the LLM market, based on the @OpenRouterAI data in the 'programming' category. What we see is quite interesting:. - Sonnet 4 models keep growing as if nothing happened. - Gemini 2.5 Pro is losing share very quickly, from 15% to 9% in a

This is interesting, in my tests, Summit and Lobster (never got Zenith) were way better than Qwen3-Coder every single time. Expect that whatever @OpenAI model version makes it to the leaderboard will be miles above everything else. Nectarine and Starfish around Kiki K2 level.

Why hasn't Microsoft trained an actually working PowerPoint and Excel agent? They have full software access, data, environments, compute - and importantly, unlike the SF tech companies, they realise how important PowerPoint and Excel actually are.

As we get ready for GPT-5, it's useful to look back at how often labs featured in the Top 5 of @lmarena_ai over the last 1.5 years. The competition is primarily between OpenAI and Google. Average appearances overall and specifically in 2025:.- OpenAI: Overall: 2.2; in 2025: 1.7

The killer feature of @OpenAI latest releases (Agent Mode, o3-pro, Codex) is async. Being able to set a task and walk away without worrying about keeping the tab open is exactly how it should be.

Could GPT-5 be claimed to be 'AGI'? The funny thing about the OpenAI 'AGI clause' with Microsoft is that OpenAI needs to show it has developed 'systems' that have the 'capability' to generate $100bn in profits, not to actually generate the $100bn. I am not saying they will do.

"Generate an SVG of Mona Lisa"

Something I'm slightly worried about is that GPT-5 would have a router and these would be the models that it would decide to route to. For example Zenith (based on others' examples) seems to be amazing at SVGs, while Summit/Lobster not so much.

Feels like Qwen Coder was heavily trained on Claude Sonnet.

It is worth your while spending a bit of time on the Web Dev Arena if you want to get the glimpse of GPT-5.

Anthropic is on track to make $4bn revenue this year, projecting to reach $12bn (or even up to $35bn) in revenue by 2027. This largely depends on its continued success in coding, with, as far as I can tell, every single coding assistant defaulting to Claude 4 Sonnet. This could

o3-alpha.Summit.Lobster.Nectarine.Starfish.

Interesting that 9 months is exactly how long babies develop before being born.

@lmarena_ai @FeatureCrewPod @btibor91 @testingcatalog @chetaslua @kimmonismus @apples_jimmy Prompt link:

@lmarena_ai Prompt from @FeatureCrewPod .Try it on: CC: @btibor91 @testingcatalog @chetaslua @kimmonismus @apples_jimmy.

New models on the @lmarena_ai WebDev arena: . - Lobster. - Nectarine. - Starfish (not in this video). In the video compared to the 'Anonymous Chatbot' (aka o3-Alpha) from 17th July. Observations:. - Lobster is closest to the o3-Alpha, but nowhere near as good. - Nectarine was

This chart shows the best image generation model at any given time, based on the @ArtificialAnlys arena and the model release dates. A few points stand out:. - Massive gains from Dall-e 2 up to Midjourney 6. - Arguably a slowdown in progress for diffusion models since then -

Unlike with LLMs, the image generation market is a bit less competitive - there are many players, but they are not constantly breaking new ground, and some vendors haven't released a competitive model in many months. In this data from @ArtificialAnlys image arena, we can see

I'm old enough to remember when a year ago, 'more than $100m' was considered an 'extreme cost' to train a model

RT @ArtificialAnlys: 🎵 Announcing Artificial Analysis Music Arena! Vote for songs generated by leading music models across genres from pop….