The Executor Model Arrives: Gemini 4 Hydrogen, Muse Ember, and Multi-Model Agent Orchestration

Frontier labs are splitting agent stacks: a model that judges, and a faster model that executes the loop. The default way to run an agent…

hero-executor-models

The default way to run an agent has been simple: pick your strongest foundation model, give it tools, and let it think through every step. That habit is getting expensive. Agents spend most of their tokens on short, repetitive moves—reading a file, calling a tool, running a test, parsing a document—not on the rare decision that actually needs a frontier model.

In mid-November 2026, three labs shipped products that treat that split as a design principle. On 12 November, Google DeepMind introduced Gemini 4 Hydrogen, a model it positions for rapid execution in everyday coding and agentic workflows, alongside Orchestrator in Google Antigravity, which can switch models inside one session. On 9 November, Meta launched Muse Ember, which it calls its smallest and most efficient model, built for the work agents repeat all day. The same day, Anthropic released Claude Fable 5.2, with explicit Advisor and Orchestrator roles and an optional fast mode documented in the Claude Platform release notes.

That layer is the executor model, and coordinating it with a stronger planner is multi-model orchestration. The scores, speed-ups, and dollar figures behind these launches are vendor claims, each tied to the dated page it comes from. Agents that drive desktops and browsers in production raise separate questions, covered in Computer-Use Agents Enter Production: What Actually Works in the Enterprise.

The split that showed up in one week

An executor model is a product choice: a model priced and tuned to run the middle of an agent loop quickly, while a more expensive model keeps the hard calls. Google’s 12 November post states the division in plain language. Gemini 4 Carbon and Gemini 4 Argon, it says, “excel at deep, long-horizon reasoning on your hardest problems,” while Hydrogen “is built for rapid execution across diverse agentic workflows.” Hydrogen, in Google’s framing, can stand on its own for everyday work or sit beside Carbon and Argon as “the high-efficiency execution engine in multi-model workflows.”

That is a different story from “one flagship model does everything,” and from the always-on teammate products covered earlier in Always-On Agents Arrive: Claude Duet, Muse Wonder, Perplexity Teammates, and Grok Bot. The news this week is about which model runs which step.

Gemini 4 Hydrogen and Orchestrator

Google dated the Hydrogen announcement 12 November 2026 and last updated the post on 13 November. The company says Hydrogen is rolling out across Google AI Studio, the Gemini API, Gemini for business, the Gemini app, Google Search, and paid Google AI Pro and Ultra subscriptions in Google Antigravity. Free Gemini app users, the post says, get Hydrogen through automatic model selection when it fits the prompt. Eligible business subscriptions can use it in Gemini in Google Sheets.

On pricing, Google’s claim is unusually clean: Hydrogen is “priced flat across its entire 1 million-token context window at $1.50 per million input tokens and $7.50 per million output tokens.” The same post says Hydrogen has a native 1 million-token input window and “up to 1 million output tokens.” It also describes four thinking levels—low, medium, high (the default), and a new xhigh tier—so teams can trade latency and cost against reasoning depth.

The benchmark claims should be read as Google’s, dated 12 November. At its default high thinking level, Google says Hydrogen scores 51 on the Artificial Analysis Intelligence Index against 50 for Argon, and 54 at the xhigh setting, “while often completing tasks faster than models in its class.” As a standalone upgrade from Gemini 3.8 Flash, Google claims Hydrogen finishes tasks 2.5 times faster at its high default setting, more than triples Terminal-Bench 4.1 scores (57% versus 18.2%), and cuts fabricated claims by more than 5x. In early partner testing on complex graphs, SQL, and data-heavy enterprise workflows, Google says partners have seen Hydrogen run 2x to 2.8x faster than 3.8 Flash while reducing token usage by 12% to 23%. The same post lists further vendor-reported placements on Agent Arena, HLE-Diamond, Vals tax and legal benches, a Multimodal Knowledge Agent benchmark, and Agents’ Last Exam v1.1. None of those figures were independently re-run for this article.

A translucent prism splitting a golden reasoning stream from a swift silver execution corridor

The more important product move may be Orchestrator. Google says it is introducing Orchestrator in Google Antigravity so that a session can “dynamically switch between models like Carbon and Hydrogen” instead of using a single model from start to finish. On Terminal-Bench 4.0, Google claims Orchestrator “reached near-Carbon task success rates while, on average, consuming 24% less quota and completing 19% faster.” For parallel agent work, the post says Carbon can act as the coordinator while Hydrogen powers worker and verifier agents “to build and test changes simultaneously across isolated workspaces.”

That is a planner-and-executor architecture described in an official launch post, with the vendor’s own success, quota, and speed claims attached. Google also published safety claims for a faster model going to a broad audience. In the 12 November post, it says Hydrogen reduces overreach in agentic tasks by up to 81% and under-service by up to 79%, cuts misaligned actions and task shortcutting across tool-use and coding evaluations by 12x, and is more than 2.3x more resilient on the Gray Swan IPI benchmark for indirect prompt injection. Those are Google’s figures, on Google’s evaluations; the post points to a model card for more detail. They are not a substitute for a buyer’s own tests.

Muse Ember: an executor priced for the loop

Meta’s Muse Ember post is dated 9 November 2026. The company introduces Ember as “our smallest and most efficient model designed for complex reasoning execution and the work agents repeat all day: tool calls, coding subtasks, routing, extraction, and long document parsing.” That sentence is the product definition. Ember is not being sold as a new frontier flagship. It is being sold as the model that should sit in the loop.

Meta says Muse Ember is available on the Meta Model API and on Muse Code as muse-ember-1.0, supports the Responses API, and has a 1 million-token context window with up to 512K output tokens. The list price Meta publishes is $0.10 per million input tokens, $0.004 per million cached input tokens, and $0.50 per million output tokens. A footnote in the same post changes those rates for longer prompts: $0.30 per million input tokens, $0.012 per million cached input tokens, and $1.50 per million output tokens. Meta also says Muse Ember, Muse Spark, and Muse Wonder are now available on Google Cloud and Oracle Cloud.

The efficiency claims are Meta’s, and they are dated 9 November. Meta says Ember “delivers more intelligence per dollar than Claude Haiku 5.5 or GPT-6 Luna at any reasoning setting,” that it “matches Haiku’s best score at about a third of the cost,” and that it “sits on the efficiency frontier of the full Artificial Analysis leaderboard: no model scores higher for less per task.” A table in the same post, at maximum reasoning effort, reports Ember ahead of the Haiku 5.5 and GPT-6 Luna scores Meta lists on several agent and coding evaluations, and tied with Haiku 5.5 on the Artificial Analysis Index (36 versus 28 for Luna). On FrontierCode 2.0-AA, Meta’s own table shows Haiku 5.5 ahead (34.9 versus Ember’s 27.2). Those numbers are Meta’s presentation of Artificial Analysis and internal results, with harnesses that differ by model, as the footnotes on the post disclose.

For teams already running agents, the useful question is not whether Ember “wins” a leaderboard. It is whether a $0.10 / $0.50 model is good enough for the calls that dominate token volume, so that a more expensive model is reserved for the steps that actually need it.

Claude Fable 5.2: advisor, orchestrator, and a faster mode

Anthropic’s launch post for Claude Fable 5.2 and Claude Mythos 5.2 is dated 9 November 2026. Anthropic says the two names are “the same underlying model.” Fable 5.2 is the generally available SKU; Mythos 5.2 is limited to organizations verified through programs such as the Life Sciences and Cyber Verification Programs.

What matters for orchestration is how Anthropic tells customers to use the model. The post says Fable and Mythos are “best for the most ambitious, open-ended technical work.” For most coding and knowledge tasks, Anthropic recommends starting with Claude Opus 5.5 and then switching “upward to Fable or Mythos, or downward to Claude Sonnet 5.5 or Claude Haiku 5.5,” depending on scope and complexity. Fable 5.2, it adds, “can also play the role of Advisor (where Haiku, Sonnet, or Opus models consult it to help solve harder problems) or Orchestrator (where it delegates simpler tasks to smaller models).”

That is an official planner-and-executor pattern, even if Anthropic does not use those exact words. The Claude Platform release notes for 9 November 2026 add the API constraints. Both models have a 1 million-token context window, 128K maximum output tokens, and always-on adaptive thinking. They cost $8 per million input tokens and $40 per million output tokens, down from $10 and $50 for the 5.1 generation. Anthropic’s launch post also claims the new models “generate outputs up to 30% faster” than Fable 5.1 and Mythos 5.1. The same release notes say fast mode for Claude Fable 5.2 is available in Messages API requests on the Claude API: it “delivers up to 2.5x higher output tokens per second, at $16 USD per MTok of input and $80 USD per MTok of output,” enabled by setting speed: "fast" with no beta header.

The 9 November release notes also tighten the beta advisor tool. A Claude Fable 5.2 or Claude Mythos 5.2 executor, the notes say, “accepts only Claude Fable 5.2 or Claude Mythos 5.2 as its advisor”; any other advisor returns a 400 error. Every executor model, however, “accepts Claude Fable 5.2 as its advisor.” In other words, Fable 5.2 can advise a smaller executor, but a Fable 5.2 executor is not allowed to take advice from a weaker model. Anthropic is encoding the hierarchy in the API, not only in a blog-post diagram.

One related detail from the same notes is worth flagging for anyone routing mid-conversation: Fable 5.2 and Mythos 5.2 can read thinking blocks from Claude Opus 5.5, Claude Sonnet 5.5, and Claude Haiku 5.5 on the Claude API and on Claude Models on Google Cloud (Google managed). No other model reads thinking blocks produced by Fable 5.2 or Mythos 5.2; if a conversation moves off that pair, those blocks are dropped. Multi-model routing is no longer just a cost optimization. It has state-transfer rules.

SpaceXSI this week, and in September

SpaceXSI did not announce a dedicated execution model in the week of 9–16 November 2026. That is a description of the public record on its news index, not a judgment about the company’s roadmap.

Its most recent official frontier-model post, Introducing Grok 4.7, is dated 21 September 2026 on SpaceXSI’s news index (the announcement page itself did not display a calendar date in the version fetched for this article). The post’s standfirst is a vendor claim, quoted here in full: “Twice as fast, at half the price of comparable models.” In the body, SpaceXSI says Grok 4.7 is “Served at the same price and speed as Grok 4.6,” is “highly competitive in its class,” and is “priced starting at $2 per million input tokens and $6 per million output tokens.” The same post says the company also serves “a fast variant with twice the output speed at twice the price.” As described there, that fast variant is a higher-priced way to run the same generation of model more quickly—not a smaller, cheaper executor SKU of the Hydrogen or Ember type.

On 2 November 2026, according to the same news index, SpaceXSI published Introducing Grok Bot 1.0, a major update to its AI teammates. On efficiency, the announcement says only that SpaceXSI has made Bots “more efficient, reducing the resources required to complete tasks.” It does not attach figures to that sentence, and none are invented here.

What this changes in a production stack

The practical shift is operational. Teams that still send every agent step to a single frontier model are paying frontier rates for work that Google, Meta, and Anthropic are now carving out for faster models—or for a mid-session switch between the two. The vendor claims in this article (near-flagship success at lower quota, more intelligence per dollar, a 2.5x fast mode) have to be tested on each team’s own harness.

A short checklist is more useful than a leaderboard. Routing lives in different places: Google’s Orchestrator inside Antigravity, Anthropic’s Advisor and Orchestrator roles and advisor-tool API rules, and Meta’s invitation to point repetitive calls at Ember. State does not always survive a switch: Anthropic’s release notes already say thinking blocks do not travel equally in both directions. List prices are not comparable without a harness—Ember’s footnotes change rates above a prompt-length threshold, Hydrogen’s price is flat across a 1 million-token window, and Anthropic’s fast mode doubles the token price for a claimed speed gain. And a cheap or faster model that calls the wrong tool is not a bargain; Google’s Hydrogen post emphasizes cleaner behavior when a tool call fails or a search returns empty, which is the right kind of claim to test.

The executor model is not a replacement for frontier judgment, and the mid-November launches do not pretend otherwise. It is an admission, from three official posts in four days, that most of an agent’s day is execution—and that execution now has its own SKUs, prices, and routers.

References

  1. Introducing Gemini 4 Hydrogen — Google DeepMind, 12 November 2026 (last updated 13 November 2026); Hydrogen positioning, pricing, Orchestrator claims, availability, safety figures.
  2. Introducing Our Smallest, Most Efficient Model: Muse Ember — Meta Research, 9 November 2026; Ember pricing, context window, Artificial Analysis and evaluation tables, cloud availability.
  3. Claude Fable 5.2 and Mythos 5.2 — Anthropic, 9 November 2026; pricing, optional fast mode, Advisor and Orchestrator roles, recommended routing with Opus / Sonnet / Haiku.
  4. Claude Platform release notes — Anthropic, entries dated 9 November 2026; Fable 5.2 / Mythos 5.2 context and pricing, fast mode (speed: "fast", $16 / $80, up to 2.5x), advisor-tool compatibility, thinking-block transfer rules.
  5. Introducing Grok 4.7 — SpaceXSI; standfirst vendor claim “Twice as fast, at half the price of comparable models,” pricing, fast variant at twice the speed and twice the price. Calendar date 21 September 2026 taken from SpaceXSI’s news index; the fetched announcement page did not display a date in the body.
  6. Introducing Grok Bot 1.0 — SpaceXSI; Team Bots, Slack / Microsoft Teams / video calls, efficiency sentence without figures. Calendar date 2 November 2026 taken from SpaceXSI’s news index; the fetched announcement page did not display a date in the body.
Share this article

Images: original illustrations generated for AndBeyond.Tech; no logos or real people depicted.