DeepSeek V4.1 Flash 1017 is now in TheGitAI. Pick it with /model. Based on Artificial Analysis: Intelligence 45, same as GLM-5.3, at 230+ tok/s. Terminal-Bench 4.0 up from 26.8% to 34%. So it totally deserves a place in TheGitAI Coding Agent.

AI is making it ridiculously cheap to create things nobody wants. More apps. More code. More content. More startups. Creation used to be filtered by effort. That filter is disappearing. Taste, judgment, and distribution are becoming the real competitive advantages.

DeepSeek calls V4.1-Flash-1017 a "minor update" while jumping +11.1 pts on Terminal-Bench 4.0: • V4.1 Flash (0910): 31.2% • V4.1-Flash-1017: 42.3% • GLM-5.3: 37.9% • GPT-5.6 Sol: 39.9% Compared against Opus 5 instead of 5.5 but at $0.15/$0.60 but superb nonetheless.

The hype behind Grok 4.7 makes no sense at all. Grok 4.7 (xhigh) is nowhere near actual frontier models. It's wild that it scores worse than Mistral Large 4 and several other dirt-cheap models, sitting in 24th place on independent Terminal-Bench 4.0 evals. Embarrassingly low.

Where do you prefer to run your AI coding agent?

CursorBench benchmarks the latest Claude, Gemini, Grok, and even Muse Spark models. But GPT-6 Sol and GPT-6.1 Sol? Completely missing. 👀 Were they never tested, or were the results never published? 🤔 Too good, too bad, or something else? @cursor_ai, what's the explanation? cursor.com/cursorbench

Grok Build and Grok 4.7 xhigh feel unbearably sluggish compared to other flagship models at similar reasoning effort. I often give up mid-task just to run it on another model.

Tested MiMo V2.6 Pro & Flash for coding agents. They overthink every action; one Pro request hung 8m 36s before a tool call. It’s not just Xiaomi’s API either; standard hosts average 50s+ TTFT. Great on paper, but unusable in an agent loop.

Grok 4.7 is live on TheGitAI. We keep our model list lean and high-signal. Every model has to earn its spot on price, speed, or problem-solving ability. Grok 4.7 cut it close, but the reasoning benchmarks earned it a seat.

GPT-6.1 Sol is available in TheGitAI coding Agent. Pick it with /model. We also heavily tested Mistral Large 4 this week and decided not to add it. It scored below GLM 5.3 flash and Deepseek V4.1 and was the by far the slowest model.

Union Alpha (revealed as Pareto 26.9, not a new foundation model) claims 51% on Terminal-Bench 4.0. For context: Astra 58%, Fable 5.1 56% on their table. Always skeptical of vendor-run cascade benchmarks vs independent runs. Anyone hooked this up to an agent yet?

GPT-6 Sol and GPT-6 Luna are now live in TheGitAI. Both are available alongside the rest of the frontier lineup, and you can use them for free with generous limits. If you’ve been waiting to try the new GPT-6 family in a real coding agent, it’s there now.

GPT-6 Sol is a much better deal than 5.6 Sol. Terminal-Bench 4.0: Sol max: 43.9% 5.6 Sol max: 39.9% Astra xhigh: 59.6% Pricing: Sol: $2/$10 5.6 Sol: $4/$20 Astra: $10/$50 Better than 5.6 Sol, at half the price.

OpenAI just answered Opus 5.5 within hours. GPT-6 Sol: $2/$10 per 1M Opus 5.5: $4/$20 And GPT-6 Luna is just $0.10/$0.50. OpenAI says Sol also beats 5.6 Sol across coding and agentic work. The frontier model race just became a price war.