Claude's New Flagship Lands as the Industry Builds Out a Budget Tier — Enter the Decision API

This digest covers model releases and updates from 2026-11-07 through 2026-11-14 that actually change capability, pricing, or availability for engineers building on these platforms.

🔥 Highlights

1. Claude Fable 5.2 and Claude Mythos 5.2 — Anthropic's new flagship, faster and cheaper per token than 5.1.

2. Gemini 4 Carbon is now available — previously cyber-defender-only frontier model opens to everyone.
   

3. Introducing Gemini 4 Hydrogen — a cheap "daily driver" that nearly matches frontier coding scores.
   

4. Introducing Our Smallest, Most Efficient Model: Muse Ember — Meta's cheapest Muse model, built for repetitive agentic tasks.

5. Decision Models in Hugging Face — a non-generative interface for structured yes/no/score outputs.
   

Anthropic

Claude Fable 5.2 and Claude Mythos 5.2 — 2026-11-09
Anthropic's 5.2 generation (same model, two access tiers) reportedly beats Claude Opus 5.5 across most evaluated capabilities, with coding/reasoning/vision gains, outputs up to 30% faster than 5.1, and an optional "fast mode" (up to 2.5x faster token generation). Mythos 5.2 specifically improves vulnerability-discovery scores and had far fewer safeguard blocks in early testing than Fable 5.1. Fable 5.2 is generally available (API id claude-fable-5-2) at $8/M input and $40/M output tokens ($16/$80 in fast mode) — a straightforward drop-in upgrade for anyone on the 5.1 line, with meaningfully lower latency at the same price structure.

OpenAI

Nothing model-release-relevant this week — openai.com/news items in this window (ChatGPT Wallet, Businesses in ChatGPT, ChatGPT Enterprise on AWS Marketplace, internal "Control Room" safety infra, and several community/education posts) are product or business announcements with no new model, capability, or pricing change.

Google DeepMind

Gemini 4 Carbon is now available — 2026-11-10
Carbon was previously gated to cybersecurity defenders via Google's Fairwind Program; it's now open to developers, enterprises, and consumers through the API, AI Studio, Gemini Enterprise, and Antigravity. Introductory pricing is $2/M input and $10/M output tokens, rising to $4/$20 on 2026-11-30 — about a three-week promotional window for anyone evaluating it.

Introducing Gemini 4 Hydrogen — 2026-11-12
A "daily driver" positioned below the frontier Carbon/Argon tier, with four thinking levels (low/medium/high/xhigh), 1M-token input and output context, and flat pricing of $1.50/M input and $7.50/M output. On Terminal-Bench 4.0 it reportedly lands close to Carbon's score while using ~24% less quota and running ~19% faster — a strong cost/latency option for high-volume agentic coding workloads.

Gemini Robotics 2.5 accelerates real-world general dexterity — 2026-11-10
DeepMind's most capable vision-language-action model yet: 30-100% higher task success and 20-30% faster execution than the previous baseline across 200+ tasks, including a 96% success rate in a Mercedes-Benz factory pilot. It also ships the open ASIMOV-Jailbreak safety benchmark. Access is currently limited to robotics partners, with no public API yet.

AlphaGenome 2: Deciphering how cells read DNA to better understand disease — 2026-11-11
A new DNA model that predicts effects conditioned on cell/tissue type, rather than generic parallel outputs like the original AlphaGenome, and reportedly nearly doubles the causal variants identified on the TenK10K dataset. It's available via API for non-commercial research now, with commercial access in private preview — a specialized release, but a genuine new model launch from DeepMind's research side.

Meta AI

Introducing Our Smallest, Most Efficient Model: Muse Ember — 2026-11-09
Muse Ember is the smallest and cheapest model in Meta's Muse family, distilled from Muse Wonder Full and aimed at repetitive agentic work (tool calls, code subtasks, extraction). It has a 1M-token context window (up to 512K output), native multimodal input/output, and costs $0.10/M input and $0.50/M output tokens — an order of magnitude cheaper than Muse Spark 1.3 while scoring nearly as well on the Artificial Analysis Intelligence Index (36 vs 37). It's already live via Muse Code, the Meta Model API (muse-ember-1.0), Google Cloud, and Oracle Cloud.

Mistral

Nothing model-release-relevant published this week — the most recent mistral.ai/news post in this window is about the Studio platform, not a new model.

Hugging Face

BotPsy: On-Device Personal Assistant with a Sub-Billion-Parameter Agentic Language Model — 2026-11-11
A new 0.8B-parameter model built on Qwen3.5-0.8B, tuned for function calling and agentic tasks (search, travel, shopping, calendar) suited to on-device personal assistants. It's Apache 2.0 with bf16/Q8/Q4 weights — a solid starting point for cheap, private, local agents, though the published benchmarks are self-reported.

Decision Models in Hugging Face — 2026-11-10
Not a single model but a new interface category: "decision" models take a state plus a question (choice/score/yes-no) and return structured probabilities directly, instead of asking a general LLM to produce and parse JSON. It ships with a DecisionPipeline in Transformers, a /v1/systemone endpoint, llama.cpp support, and a decision-model Hub tag — useful for classification/routing pipelines that currently lean on fragile prompt engineering to extract structured output.

LLMON: An LLM-Native Markup Language, now in Granite 4.1 3B and 8B — 2026-11-12
IBM released two preview models (granite-4.1-3b-llmon-preview, granite-4.1-8b-llmon-preview) trained on a markup language that separates instruction, data, and execution inside the context window. They report a strong resistance improvement against indirect prompt injection on IBM's OPI benchmark, with overall performance comparable to the base Granite 4.1 Instruct models — relevant for anyone running RAG or agents over untrusted content, though these are experimental previews, not production releases.

llama.cpp v1.0.0: A stronger foundation for the local AI ecosystem — 2026-11-13
llama.cpp adopts real semantic versioning starting at v1.0.0: breaking API/ABI changes only land in major releases, new model/feature support in minors, bugfixes in patches. No new feature on its own, but it's a meaningful stability commitment from a tool nearly everyone running local models depends on directly or indirectly.

Simon Willison

Introducing Our Smallest, Most Efficient Model: Muse Ember — 2026-11-10
Simon's linkpost on Meta's Muse Ember launch, including his own pelican-on-a-bicycle test; he found output quality comparable to Meta's much larger (and ~40x more expensive) Muse Wonder for everyday agentic tasks.

llm-gemini 0.36 — 2026-11-10
Plugin release documenting Gemini 4 Carbon's move to general availability (previously restricted to a small group of cyber-defense testers), with the $2/$10-per-million-token launch pricing that rises to $4/$20 in December.

llm-anthropic 0.31 — 2026-11-09
Plugin release adding support for the freshly launched Claude Fable 5.2 and Claude Mythos 5.2 (llm -m claude-fable-5.2) — same-day tooling support for anyone already using the llm CLI.

Gemini Decisions API — 2026-11-11
Linkpost on Google's new "Gemini Decision 1" API — a decision-model endpoint supporting yes/no, choice, and rubric scoring, with multimodal input (text, image, audio, video, PDF), unlike competing decision APIs that are text/JSON only. Pricing is $0.10/M input tokens with free output — notable for multimodal classification/moderation without the overhead of full text generation.

llm-gemini 0.37 — 2026-11-12
Plugin release covering Gemini 4 Hydrogen, described as the cheapest member of the Gemini 4 family ($1.50/$7.50 per million tokens), and confirming the Carbon/Argon price increase to $4/$20 effective 2026-12-01.

Through-line

Two patterns repeat across otherwise unrelated sources this week. First, every major lab is now shipping a deliberately cheap "sibling" alongside its flagship — Gemini 4 Hydrogen next to Carbon, Meta's Muse Ember next to Wonder — explicitly trading a sliver of frontier capability for an order-of-magnitude cost cut on high-volume agentic workloads. Second, two independent teams (Hugging Face and Google) launched dedicated "decision model" APIs in the same window, both betting that a lot of production LLM calls aren't really open-ended generation at all — they're classification, scoring, or routing wearing a chat-completion costume, and purpose-built endpoints can serve that faster and more reliably than prompting a general model for JSON.

What's your team actually using this week — a frontier model, one of these new budget siblings, or has a decision-model API already replaced some of your structured-output prompts? Drop a comment below.