Meta's Muse Ember brings cheap, efficient reasoning to high-volume agent tasks on Vercel AI Gateway

Meta has unveiled Muse Ember, a model it describes as its smallest and most efficient, built for the repetitive structured work software agents perform all day: tool calls, coding subtasks, routing, extraction, and parsing long documents. Vercel announced the same day that the model is live on its AI Gateway under the name meta/muse-ember.

Built for the high-volume middle of an agent run

According to Meta, agents burn most of their tokens not on hard reasoning but on many small, structured steps: fetching records, invoking tools in sequence, updating systems across applications. Muse Ember is aimed squarely at that layer, and Meta positions it as a companion to its larger Muse Wonder model — Wonder handles the planning while Ember carries the high-volume execution.

The model offers a 1M-token context window and can produce up to 512K output tokens, enough to absorb an entire codebase or document set in a single call. Meta says it is natively multimodal across text, images, audio and video, so an agent can work from screenshots, charts, scanned contracts or meeting recordings without routing to a separate model.

Vercel's announcement describes the same 1M-token window with text, image and video input, adjustable reasoning effort from minimal to max, and tool calling that lets an application feed each result into the next step.

The numbers Meta is publishing

Meta claims better intelligence per dollar than Anthropic's Claude Haiku 5.5 or OpenAI's GPT-6 Luna at every reasoning setting, and says that on Artificial Analysis's leaderboard no other model delivers a higher score for a lower cost per task. On charts Meta reproduced from Artificial Analysis, Muse Ember's best setting scored 35.8 at an estimated $0.136 per task, while Haiku 5.5's best scored 36.2 at $0.411 — similar output for roughly a third of the spend.

At maximum reasoning, Meta's figures put Muse Ember ahead of both competitors on AutomationBench-AA (37.8 percent, versus 20.2 for Luna and 13.9 for Haiku 5.5), on Agents' Last Exam (61.6 percent on the Linux subset), on Terminal-Bench 4.0 (40.9 percent) and on the Coding Agent Index (46 against 37 and 31). It is not a clean sweep: Haiku 5.5 leads FrontierCode 2.0-AA with 34.9 percent to Muse Ember's 27.2. Meta notes some results come from Artificial Analysis while others are internal evaluations.

Among early testers, Box reported strong accuracy, speed and token efficiency on its Complex Work Eval, including scoring contracts clause by clause against a company's negotiation playbook.

Pricing, safety and availability

Meta is shipping the model on its Model API and in Muse Code as muse-ember-1.0, priced at $0.10 per million input tokens and $0.50 per million output tokens, with cached input listed at $0.004 per million. On the safety side, Meta says Muse Ember went through its Superintelligence Scaling Framework evaluations — covering chemical, biological, cybersecurity and loss-of-control risks — and stayed below the framework's high-risk thresholds, along with tests of how it behaves when acting on a user's behalf, such as resisting instructions hidden in documents or tool outputs.

Vercel, for its part, has added the model to AI Gateway. Developers can reach it through the AI SDK, the Chat Completions, Responses and Anthropic Messages APIs, or coding agents such as Claude Code, Codex and Cursor after running npx vercel ai-gateway setup and selecting the model. Vercel says the gateway passes provider pricing through with no markup and no platform fee on inference, including bring-your-own-key requests, and adds usage tracking, retries and failover on top.

Why it matters

The economics of agents are decided by volume: one run can fire off hundreds of small tool calls, file edits and summarisations, and the cost of those steps determines whether the workflow is viable at all. A small model tuned for exactly that workload — long context, multimodal input, aggressive pricing — attacks the unit-cost problem directly, and the Wonder-plus-Ember pairing points to a broader industry split between planning models and execution models. The caveats are real: the comparison set is narrow, the figures are Meta's own, and at least one benchmark favors a rival. But ready availability through a neutral gateway at provider pricing makes it cheap for teams to check those claims against their own agent workloads.