Gemini Decision 1: $0.10 per 1M input tokens, 0 output tokens.
Google's new model scores yes/no, multiple-choice and rubric questions with calibrated probabilities. Up to 1,000 per call.
I'd start with routing and LLM-as-judge calls.
Would you swap your classifier prompts?
Qwen4-Max is out in preview with a 1M-token context, and Alibaba retires it after Nov 25.
The model ID is qwen4-max-preview. API access is through Alibaba's Token Plan subscription only. No pay-as-you-go.
• Up to 393,216 output tokens, shared by thinking and the final answer.
• Thinking is on by default at xhigh. reasoning_effort lets you drop it to low or medium.
• Text, image and video input, function calling, web search and JSON Schema outputs.
• Compatible with OpenAI Chat Completions, Responses and Anthropic Messages APIs.
• No changes during the preview, which runs Nov 11 10:00 to Nov 25 24:00 (UTC+8).
My take: the line I'd read twice is in the usage notes. From this model on, max_tokens in Chat Completions caps thinking plus the answer together. A limit sized for answers alone can cut the output short. If you point your coding tool at it, raise that limit first.
Caveat: the model page lists no benchmark scores and no parameter count, so any comparison is one you run yourself.
Would you test a model with an expiry date, yes or no?
Gemini 4 Carbon is now the top model on CursorBench at 60.4% (high effort, per Cursor). Google's own advice: keep Argon for most work.
Carbon went public today, two weeks after a release only for cyber defenders.
• API: paid tier in Google AI Studio and Android Studio, plus Gemini for business. Not on the free tier.
• Intro price through Nov 30: $2 per 1M input tokens, $10 per 1M output. From Dec 1: $4 per 1M input tokens, $20 per 1M output.
• Google says it ranks #1 on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench at more than 2x lower cost per task than the next best external model.
• It spends more tokens than Argon to explore and run longer verification loops.
• Google AI Ultra subscribers also get it in Antigravity, which is previewing /swarm: one prompt becomes a team of agents.
My take: same per-token price as Argon, more tokens per task, so the bill depends on the job. I'd send only the hard work to Carbon (multi-package migrations, root-cause hunts across subsystems) and keep Argon as the default.
Caveat: the rankings and cost claims come from Google and Cursor, not an independent run, and the price doubles on Dec 1.
Would you make Carbon your default coding model, yes or no?
AI agents wrote over 6 million lines of Lean to prove key security properties of Signal's protocol code. Humans got a second look at about 1%.
Google DeepMind's AlphaProof Code proved key security properties of the actual compiled code behind the Signal protocol. Most protocol proofs stop at the design. DeepMind says it took four weeks.
• Researchers and models turned Signal's security goals into formal specs. Human experts reviewed every spec.
• Agents powered by Gemini Carbon wrote the proofs. Lean checks them mechanically.
• The agents flagged about 1% of the codebase for a human second look.
• Sanity check: two real, already-fixed bugs were put back into an isolated copy. Both times the system produced a machine-checked proof that the buggy version was insecure.
• DeepMind is open-sourcing the verification artifacts.
My take: the agents did the volume, but humans decided what "secure" means. If you want this for your own code, the human hours go into writing the specs. I'd start with the one module that would hurt most to get wrong.
Caveat: this covers Signal's core cryptographic code, and every claim above is DeepMind's.
Would you trust a proof no human read in full?
Muse Ember: 40.9 on Terminal-Bench 4.0. Claude Haiku 5.5: 39.2. Max effort, own harnesses.
Meta's smallest model, $0.10 per 1M input tokens, $0.50 per 1M output. On the Meta Model API today. Vendor-reported.
I'd test it first as a coding subagent.
Would you swap Haiku for it?
Google just put two TPU 8i chips in one package: 2x the compute, on-chip memory and memory bandwidth of TPU 8i. It's built for million-token agent contexts.
TPU 8i Max and TPU 8tx, announced today, make four eighth-gen TPU systems in one year. What's in the launch:
• TPU 8i Max fits twice the silicon in the same rack footprint, aimed at dozens of agents working across a whole codebase or hours of video.
• TPU 8tx keeps data and model weights protected even while in use, on AMD EPYC hosts, with up to 2x lower latency than TPU 8t (Google-reported).
• TPU Sync, being open-sourced, moves the model's long-context memory off the chip and cuts its transfer latency and time-to-first-token by more than 90% (Google-reported).
• An open-source TPU agent suite, in preview, plugs into Claude Code, Cursor and Antigravity to onboard, find bottlenecks and autotune kernels in plain language.
My take: the chips are for cloud buyers, but the software is the part devs will actually touch. Since last month's TorchTPU release, PyTorch code and Hugging Face models run on TPUs without rewrites, and vLLM and SGLang have TPU backends.
Caveat: the post gives no pricing or availability dates, and every number above is Google-reported.
Would you try a TPU if your PyTorch code ran unchanged?
Half of Qwen3.5-35B-A3B's experts deleted. SWE-Bench Pro still matched the full model: 42.3 vs 41.8.
That's HOPE, an expert-pruning method from Amazon researchers, written up today on the Hugging Face blog. Most pruning methods score each expert on its own. HOPE also scores which experts fire together, then picks a prune set that keeps those groups intact.
• Same 50% cut with REAP, the state-of-the-art method that scores each expert alone: 36.2 on SWE-Bench Pro. HOPE: 42.3 (team-reported).
• Best average rank across 3 models, 2 calibration sets and 10-50% pruning rates: 2.07 among 5 methods.
• The extra solve step takes 1-2 seconds per layer on top of the usual calibration pass.
• Output is a standard Hugging Face checkpoint, no vLLM or SGLang changes. Tested on Qwen3, Qwen3-Next, Qwen3.5 and GLM-4.5.
My take: in MoE models the experts are most of the memory (95% of Qwen3.5-122B's ~250 GB, per the post). Pruning stacks with quantization, so this is a cheap way to fit a bigger model on the GPU you already have. I'd calibrate on my own agent traces, not generic code.
Caveat: at 10-20% pruning, simpler methods like REAP are sometimes just as good, and every number above comes from the team.
Would you run a 50%-pruned model in your coding agent?
90% vs 76% on Active Memory: Meta shipped Muse Wonder-trained models to Muse and Meta AI.
Vendor-reported. Active Focus in Muse: +67% finishing tasks when things break.
I'd treat memory and recovery as two harness skills, not one smarter knob.
Ship memory or recovery first?
Claude Code put signed-in users' email into User-Agent headers on web and curl calls. HN is at 1,290 points.
An Anthropic engineer on Claude Code (lydia_hallie on HN) said they added the account email to model context so author tags would stop hallucinating. The model then reused that email as a contact in User-Agent when sites ask for one.
What's shipping:
• v2.1.317+ ships a built-in privacy mod (on by default) that asks before a web or curl call containing your email.
• It still misses some cases, like when Claude writes a script and runs it alone.
• Stable channel is still on 2.1.311. Update to 2.1.317 or later today.
• Permanent fix: email leaves the model context in the next release, early next week.
My take: I'd update now and add a PreToolUse hook that blocks my own email in User-Agent until the context change lands. Don't wait for stable to catch up.
Are you on 2.1.317+ already, or still on stable?