Claude Sonnet 5.5 on max absolutely destroyed Opus 5.5 on my Lamborghini SVG test. It even drew the scissor door open with the interior showing. Then look at the tokens. 648M in. Opus used 167M. 5 and a half hours and $279. Opus did it in under 3 hours for $102. The cheap model cost 2.7x more.

Claude Sonnet 5.5 just broke my NYC test and I did not expect this from a Sonnet model. It might have even destroyed Opus 5.5. It flies its own tour. Central Park packed with trees, traffic running along the park, the Empire State, Chrysler, One World Trade, 432 Park, the Brooklyn Bridge, the Statue of Liberty, then down to street level and over the rooftops. Golden hour the whole way. In June Sonnet 5 spent 4 hours on this exact prompt and I couldnt even move the camera. Sonnet 5.5 max on top, Sonnet 5 max below. Full run, nothing cut. If this is Sonnet I cant wait to see what Haiku 5.5 does.

Opus 5.5 has once again destroyed my expectations for games and 3D modeling. A full survival game like The Forest, in three.js, in 1 day of continual work. A 2 km island carved by simulated erosion. 44,000 trees, every one generated in code and wrapped in real photoscanned bark. Six real Pacific Northwest species. The sun and moon sit where they really would at 49 degrees north in late September. Thunder arrives late because it travels at the speed of sound. Wolves stalk at dusk, circle the fire at night and back off from a torch. Claude Code split the world between waves of agents, one per domain, and critic agents that never touched the build checked the renders against real photos. 62 seconds of 4K, 1,881 frames, rendered one by one out of the game itself. The wolf is still the weakest thing in it and it doesnt hold 60 fps yet. I ran out of Claude usage before the realism pass finished. Its not even done.

Opus 5 has completely destroyed my expectations for what an LLM can do. 26 billion tokens. 4 weeks of feeding it Claude accounts nonstop. 124 seconds of 4K, 2978 frames, one camera, zero cuts. Nothing imported. Nothing sculpted by hand. Zero texture files. Every surface is code it wrote itself. 17,707,774,735 triangles evaluated every single frame. 52.7 trillion across the film. It rented 3 RTX 5090s on its own through middleware another Claude agent built. 231 GPU hours for the final render. The render cost $229. The thinking cost $20,740. The 4K film was the cheap part by 90x. It also rendered the car inside out for weeks and none of us caught it. A back facing solid fills the same silhouette, so it still looked like a car. Audio is the one thing it never got. Rebuilt five to seven times and I reverted to the original every time. A year ago none of this was possible.

60 frontier models have already shipped this year. OpenAI 14. Anthropic 9. Google 8. Qwen 7. xAI 4. Kimi 4. Meta 3. GLM 3. MiniMax 3. DeepSeek 3. April had 14 on its own. July 9 had four in a single day. Google shipped seven Gemini models. Five were Flash. No Pro since February 19. Still coming: Grok 4.6, Grok 4.7, Gemini 3.5 Pro, Gemini 4, Kimi K4, Grok 5, Qwen 4, Meta Avocado. Plus OpenAI Astra, still unnamed. GPT-6 or 5.7, they havent said. Seven months in and we are still expecting A LOT more to come.

Wanted to see if grok 4.5 can recreate this procedural dungeon generator. Took only 5 minutes and 1 reprompt due to shadows. Its not as detail and colorful with the lightning and shadows but basically one shotted the idea with 118k tokens and took 1/10th of the time.

Fable 5 + Claude Code just built what might be the most detailed solar system sim I’ve seen from any AI model. 92 celestial bodies. 122 walkable surfaces. Real physics (eclipses, barycenter wobble, seasonal caps, plasma tori, sodium tails). 17 time-travel events. Live spacecraft. Even predicted real future transits on its own. 1 prompt. One codebase. This is actually insane.

Warp just went open source One of the best agentic terminals available is now community-driven with a public roadmap and open contribution process OpenAI is the founding sponsor and the repo is managed by Oz agents handling the coding, planning, and testing so contributors can focus on direction and ideas The terminal that already uses AI to build your code is now using AI to build itself github.com/warpdotdev/warp

If these benchmarks hold in real world testing, deepseek v4 is the cheapest model for the performance we have ever seen V4 pro: $1.74 in / $3.48 out per MTok V4 flash: $0.14 in / $0.28 out per MTok 8x cheaper than gpt-5.5 on output alone 93.5% livecodebanch, codeforces 3206, 1M context, MIT license, open weights The price argument just got a lot harder to ignore

DeepSeek-V4 just dropped on Hugging Face. no announcement. no press release. just live. Pro: 1.6 trillion parameters, 49B activated. Flash: 284B, 13B activated. both with 1 million token context. MIT license. open weights. trained on 32 trillion tokens. 3.7x lower FLOPs than V3.2. Codeforces rating hitting 3206 in benchmarks which if accurate is completely unreal. DeepSeek does not do press conferences. they just ship. huggingface.co/deepseek-ai/D…