Codara Solutions

AI news, coding agents, and live vibe-coding sessions. Building Codara in public.

[Codara Solutions @CodaraSolutions]

Google just shipped Gemini 3.5 Live Translate. 70+ languages, near real-time speech-to-speech translation, and public preview for developers via the Gemini Live API and Google AI Studio. Much more usable than the usual turn-by-turn voice stack.

[Codara Solutions @CodaraSolutions]

Anthropic just shipped Claude Fable 5. 1M context, 128K max output, adaptive thinking always on, and $10/$50 per 1M tokens. First Claude release in a while that looks built for long agent runs instead of short coding bursts.

[Codara Solutions @CodaraSolutions]

Ideogram just open-sourced Ideogram 4.0. 9.3B Diffusion Transformer, frozen 8B VLM text encoder, and the nf4 checkpoint runs on a single 24GB consumer GPU. Strongest open-weight image-model drop in a while if the arena rankings hold.

[Codara Solutions @CodaraSolutions]

Microsoft just shipped MAI-Thinking-1. 35B active params, ~1T total, 97.0% on AIME 2025, 94.5% on AIME 2026, and toe-to-toe with Claude Opus 4.6 on SWE-Bench Pro. Serious first swing at a frontier reasoning model.

[Codara Solutions @CodaraSolutions]

Holo3.1 just shipped. 79.3% on AndroidWorld for the 35B-A3B model, plus 0.8B/4B/9B/35B sizes and FP8, Q4 GGUF, and NVFP4 checkpoints for local inference. Strong open computer-use drop.

[Codara Solutions @CodaraSolutions]

Qwen just shipped Qwen3.7-Plus. It is live in Alibaba Cloud Model Studio, runs full GUI + CLI agent loops, and posts 70.3 on Terminal-Bench 2.0. Feels more like a real multimodal coding agent than another VLM refresh.

[Codara Solutions @CodaraSolutions]

NVIDIA just shipped Cosmos 3. Open models are live as Nano 16B and Super 64B, and the model card lists 256K context plus text/image/video/audio/action IO. Serious open-source swing at robotics and physical AI.

[Codara Solutions @CodaraSolutions]

Qwen just shipped Qwen-VLA. One generalist robot model: 97.9 on LIBERO, 73.7 on Simpler-WidowX, 86.1/87.2 on RoboTwin, and 69.0 on R2R OS. More interesting than another VLM demo because this one is built to act.

[Codara Solutions @CodaraSolutions]

Google Cloud just put Nano Banana 2 and Nano Banana Pro into GA. 1K and 2K output are live, 4K stays in preview, and Nano Banana 2 now takes video as an input prompt. That is a real multimodal image workflow upgrade, not just another image model rename.

[Codara Solutions @CodaraSolutions]

StepFun just shipped Step 3.7 Flash. Apache-2.0 weights are live, with 256K context, up to 400 tok/s, SWE-Bench Pro 56.3, and $0.20/M in, $1.15/M out. Serious new open-weight agent model if the tool-use evals hold.

[Codara Solutions @CodaraSolutions]

xAI just put Grok Build 0.1 on the API. 256K context, $1/M input, $0.20/M cached input, $2/M output, and xAI says it runs past 100 tok/s. Real price/perf shot at the coding-model lane, not just another CLI wrapper.

[Codara Solutions @CodaraSolutions]

Anthropic just shipped Claude Opus 4.8. 1M context by default on API, Bedrock, and Vertex, plus 128K max output and mid-conversation system messages. First Opus release that feels built for long agent runs, not just one-shot coding.

[Codara Solutions @CodaraSolutions]

ANTHROPIC 🔥: Mythos 1, "claude-mythos-1-preview", is being prepared for release in Claude Code and Claude Security.
> claude-mythos-1-preview
Doesn't mean public access yet — Anthropic has staged internal models before.

[Codara Solutions @CodaraSolutions]

NVIDIA just shipped Nemotron-Labs Diffusion. 3B, 8B, and 14B text models, plus an 8B VLM. NVIDIA says the 8B beats Qwen3 8B by 1.2% average accuracy and hits 2.6x higher TPF, up to 6.4x with self-speculation. Wild open-model speed push.

[Codara Solutions @CodaraSolutions]

Cohere just shipped Command A+ open-source. Apache 2.0, 218B total / 25B active, 128K context, multimodal + tool use, and it runs on 2 H100s. Serious open-weight agent model if the benchmarks hold up.

[Codara Solutions @CodaraSolutions]

xAI just put Grok in OpenCode. SuperGrok and X Premium now work inside /connect, including a headless OAuth flow for remote boxes, and the model is Grok Build. Clean step toward using Grok in real coding-agent runs.

[Codara Solutions @CodaraSolutions]

Google just rolled out AI Ultra. From €99.99/mo gets 20 TB, up to 20x Gemini limits vs Pro, plus higher AI Studio, Antigravity, and Jules access. That’s Google finally turning the agent stack into a real paid tier.

[Codara Solutions @CodaraSolutions]

Gemini 3.5 Flash is not cheap Flash anymore.

Official pricing: $1.50 in / $9 out per 1M tokens — 6x Gemini 3 Flash Preview. Vals says it ranked 10th on Vibe Code Bench.

Fast, sure. But the “cheap coding model” story needs receipts.

[Codara Solutions @CodaraSolutions]

Qwen just shipped Qwen3.7-Max. Built for writing and debugging code, automating office workflows, and keeping long autonomous runs alive across hundreds or thousands of steps. Serious swing at the long-run agent loop.

[Codara Solutions @CodaraSolutions]

Google just shipped Gemini Omni. Gemini Omni Flash takes image, audio, video, and text as input, starts with video generation, and lets you edit clips through conversation. Rolling out now in the Gemini app, Flow, and YouTube Shorts.