Mistral large 4 api price is now $0.70 input and $2.10 output per million tokens, about half the preview rate. Reasoning defaults to high. Mistral's own chart puts medium at 35 and high at 48, with fewer tokens per task on medium. Test both on your agent.
Gemini 4 carbon is not open to developers yet. Google starts with a small group of cyber defenders and says wider access comes later. The intro price matches argon and rises after the intro period, so plan costs around the later rate.
Copilot now has a fusion option in the model picker that chooses and combines models for each prompt. Github says fusion max scores 60 on the artificial analysis coding agent index, within two points of claude code with opus 5.5 max, at about half the cost per task.
Jetbrains IDEs now run their terminal on libghostty, starting with 2026.3. In claude code and codex, shift+enter starts a new line instead of sending a half-written prompt. Tools like opencode and neovim can also follow the ide's dark or light theme.
Opera one now ships webai in early preview. Sites and extensions can call a local model like gemma through the browser, so developers add ai features without paying per request and the data stays on the device. It's opt-in and runs on 64-bit windows and apple silicon macs.
Codex can now import your setup from claude code or cursor. It copies instruction files, skills, plugins and recent chats, and leaves the original setup untouched. Type /import in the cli or open settings in the desktop app to try the same project in both.
Google open sourced the plugin its engineers use to review a design doc before the agent writes code. It installs in claude code, codex and antigravity. Fresh subagents with no context read your spec and flag gaps, so you fix them before any code exists.
A new study ran the same models through claude code, opencode and a 100-line harness. Pass rates stayed within a few points on the large task set. Cost per task changed up to 3x, mostly from the fixed prompt resent every step. The chart shows input tokens per task.
Kimi k3.1 pro lands fourth on artificial analysis' intelligence index, behind two claude models and muse wonder full. Its token price is a third lower than k3, but it uses more tokens, so cost per task stays about the same. Moonshot says open weights arrive by nov 4.
Mercury 3 runs at over 1k tokens per second, roughly 4 to 8 times faster than gpt-6 luna and claude haiku 5.5 on inception's own table.
The api is openai compatible and launch pricing is 50% off, so a slow agent loop is cheap to test on it.
The new x subscription changes what your cursor plan pays for. The included usage now goes to grok models only, and claude or gpt bill at api rates from the first request, even on ultra.
Grok-heavy users come out ahead. Anyone who mostly codes with claude should keep the current pro plan while it lasts, since cursor says existing users will only be moved later and with advance notice.
You can now launch apps you built in AI Studio straight from Google Play with just a Google account. They stay private to you.
Until now you had to publish publicly or plug your phone into a computer, so many people built web apps instead. It's US only and 18+ for now.
Google is making it easier to build and use personal Android apps created with AI, without having to make them available to everyone on the Play Store.
spr.ly/6019BH1EZd
Anthropic is doing about half of OpenRouter's tokens right now but pulling in two thirds of the spend. Most of that is Opus 5.5.
So people are paying way more per token and still picking it. I guess the price stops mattering when the agent gets it right the first time.
Artificial Analysis just launched its leaderboard for FrontierCode 1.1, Cognition's private coding benchmark.
Every task is a real change to an open source repo, built with maintainers who spent over 40 hours per task. A failed blocking check scores zero, so it grades what a maintainer would actually merge.
Leading models still fail more than 40% of the time. Claude Opus 5.5 sits on top at 52%, but GPT-6.1 Sol is at 49% for $0.81 per task. MiMo-V2.6-Pro leads open weights at 45%.
Nine models from six labs land between 43% and 52%, so the next release can take first place.
Meta is back in the frontier race.
Muse Wonder Full just scored 56 on the Artificial Analysis Intelligence Index, second only to Claude Opus 5.5 at 58 and ahead of GPT-6 Astra, Claude Fable 5.1 and Gemini 4 Argon, all tied at 53.
It also tops Meta's eval table on long coding work, with 57.5 on SWE-Marathon and 79.0 on DeepSWE.
Opus 5.5 still holds a clear lead on ProgramBench at 70.5 vs 59.0, and wins Terminal-Bench and HealthBench Pro too.
Meta's last model, Muse Spark 1.3, scored 48 on the same index. That's an 8 point jump in one release.
Meta just released Muse Wonder and Muse Wonder Full, two frontier models built for agents and coding, 14 months after rebuilding its AI work from the ground up.
Muse Wonder is out now on the Meta Model API and in Muse Code. Wonder Full is only for Muse Code subscribers, and Wonder will soon power Meta's Muse agent, subscribers first.
We’re excited to introduce the Muse Wonder series with Muse Wonder and Muse Wonder Full, our most capable models yet. Both are frontier models, built to deliver state-of-the-art performance for personal and professional agentic tasks.
Muse Wonder is available today via the Meta Model API and Muse Code. Muse Wonder Full is available today for Muse Code subscribers. Muse Wonder will also soon power experiences in our personal agent product, Muse, rolling out first to subscribers.
I think this DigitalOcean launch is a big deal for anyone shipping AI agents to real users.
AI agents often fail without throwing a single error. The tool call succeeds, the agent replies politely, the logs look clean, and the user still ends up stuck. Most teams only find out once support tickets start piling up.
Signals watches live agent conversations and flags the ones that went wrong, even when nothing errored. Simulations then turns that flagged conversation into a test and replays it with the fix across 100 simulated users before anything ships.
Their demo shows a fintech onboarding agent telling a new user to "try again later" because a bank had switched off its old callback URL. Zero errors in the logs. DigitalOcean says a bug like that used to reach about 412 users before anyone noticed. Signals flagged it after the first user, in 4 minutes.
As more companies put agents in front of customers, I expect this kind of monitoring to become as standard as crash reporting is for regular apps.
Nvidia's best customers keep turning into its competitors. Google has TPUs, Amazon has Trainium, and now Maia is set to run Microsoft's own products.
Training stays on Nvidia for a while. Inference is where the money moves first.
Today we are announcing our biggest bet yet on our own silicon.
Starting next calendar year, we will scale our Maia family to multi-gigawatt deployments, making it the primary inference system for our own product and services.
And we’re introducing Maia 300, which delivers 40% more tokens per watt compared to leading GPU-based systems.
We are already deploying Maia 200 across the US, Europe, and Australia, delivering up to 1.7x more tokens per watt across experiences in Teams and Copilot.
Our goal is to build on the full-stack efficiency gains we’ve already seen, while continuing to bring the latest innovations from across the ecosystem into our fleet.
I actually like this way of launching. No benchmarks, no hype thread, just a free model with a Pokémon name and people left to figure it out.
If it really is Beam, Reflection gets thousands of honest reviews before the weights drop. Not all of them are nice, and that's the useful part.
