Meta’s Muse Spark 1.3 for UK builders: what to test before you switch

Meta released Muse Spark 1.3 on 2 September 2026. It costs less per token than GPT-6 Sol, Grok 4.7 or Claude Opus 5.5, and Meta says it finishes agent work in fewer steps. Both are reasons to test it. Neither is a reason to switch until you have tested it.

Prices, model details and access rules checked against each vendor’s own pages on 26 and 27 September 2026 (prices on 27 September). All prices are in US dollars. They change without notice, so check again before you commit.

The answer in brief

• What it is: Meta’s latest agent and coding model, muse-spark-1.3, in Muse Code and the Meta Model API.
• Price: $1.25 per million input tokens, $4.25 output and $0.15 cached input, unchanged from 1.2. That is the lowest list price of the four models compared below.
• What Meta claims: about 20% fewer tool calls and 25% fewer tokens than 1.2, from comparisons by Meta’s own engineers. We found no independent test.
• UK access: yes, under Meta’s published policy, and we confirmed it from the UK on 27 September 2026. The UK is not a restricted territory.
• Watch out for: Muse Code 1.4.0 and the dashboard’s ready-made setup snippets both default to the Contributor tier, which trains on your data, and prompt caching can move an agent’s bill more than the model switch.
• Not included: no change to Ads Manager or Advantage+, and the consumer Muse agent is available in the US, Canada and Mexico, with no UK date.
• How to judge it: cost per finished job on your own tasks, not the token price.

Which Muse product does this cover?

This post is about the model you call from code. If you came for Muse Image in ads, our guide What is Meta Muse? covers that, along with disclosure and what UK businesses can use today.

What did Meta ship in Muse Spark 1.3?

• Where it runs. Muse Code and the Meta Model API, as muse-spark-1.3, marked “recommended for new work” on Meta’s models page. It is also listed on OpenRouter, according to Meta’s model page, and the Connect 2026 recap adds Oracle Cloud AI Platform, with Google Cloud in private preview.
• How it behaves. It is trained for longer agent tasks: it asks clarifying questions when a prompt is ambiguous, asks for help when stuck, confirms before consequential actions, and keeps separate tasks apart within one messy thread.
• Efficiency. In comparisons by Meta engineers, reported in the launch post, it used about 20% fewer tool calls and 25% fewer tokens than 1.2, with fewer unnecessary turns and less verbose output. These are Meta’s figures; the coverage we read repeats them, and none of it measured them.
• Inputs and context. A 1,048,576-token context window. Text, image, video, audio and PDF in; text out. Audio quality on 1.3 may be degraded, so Meta points audio work to 1.2 or Muse Voice Transcribe.
• Safety. Meta claims stronger resistance to prompt injection and better judgement about irreversible actions.

How does Muse Spark 1.3 compare on price with GPT-6 Sol, Grok 4.7 and Claude Opus 5.5?

On list price, Muse Spark 1.3 is the cheapest of the four for input, output and cached input.

The last column matters for agents, which carry a growing history from loop to loop. Meta’s pricing page says “There is no long-context premium” for Spark.

A lower token price is not a lower bill if a model needs more loops, more tokens or more human rescue to finish the same job. That is why the test further down measures finished jobs. We cover the other three models in our posts on GPT-6 Sol, Grok 4.7 and Claude Opus 5.5.

Is Muse Spark 1.3 worth upgrading to from 1.2?

The price is the same, so you are not buying a cheaper model. You are betting that the same finished job takes fewer loops, which means less waiting, fewer tokens and less chance of the agent wandering off task.

That bet only pays on multi-step, tool-heavy work. A single prompt will barely notice. An agent making dozens of tool calls per task is where a 20% cut would show on the invoice, if it holds for your tasks.

Which model is your setup actually running?

Check this before anything else, because the answer is often “not 1.3”.

Muse Code may not start where the docs say. The Muse Code configuration docs say “The default model is muse-spark-1.2”, but when we installed Muse Code 1.4.0 on 27 September 2026 it defaulted to muse-spark-1.3-contributor. Check the status line, then switch with muse --model muse-spark-1.3, use /models mid-session, or set the default in ~/.config/muse/settings.json. Muse Code also allows max reasoning effort on Contributor, which the API does not, so a Muse Code session and an API call can behave and bill differently on the same model ID. Our Muse Code guide covers the rest of the set-up.

Other coding agents can use it too. Meta’s coding agents guide documents Claude Code, Codex and OpenCode pointed at the Meta Model API, and recommends muse-spark-1.3 for all three. That lets a team test Muse Spark 1.3 inside the agent it already uses.

The dashboard’s setup snippets use the Contributor model. The guide recommends muse-spark-1.3, but the ready-to-copy snippets on the Meta Model API dashboard use muse-spark-1.3-contributor, as we found on 27 September 2026. The Claude Code block sets both the main model and the Opus slot to it, so choosing Opus in Claude Code runs Meta’s Contributor model, and the Codex and cURL snippets and the playground start on it too. Paste them unchanged and you are on the tier that trains on your prompts, with far lower rate limits. Change the model ID to muse-spark-1.3 before any client data goes through it.

Check the context setting in your tool. Meta lists a context window of 1,048,576 tokens for Muse Spark 1.3. Cursor offers the same model with a 300K or a 1M context setting, and it was on 300K when we checked. If your test depends on long agent histories or large repositories, set the context before you compare models, or you will be testing a smaller window than the API gives you.

What can push the bill up, and what brings it down?

Four settings move the bill more than the model ID does.

1. Reasoning effort. Reasoning tokens are billed as output. The reasoning docs list six reasoning_effort levels for Muse Spark: minimal, low, medium, high, xhigh and max. On the API, max works only on Standard-tier muse-spark-1.3, not Contributor. If you omit the setting, the model picks its own depth. Meta’s advice is to “use the lowest level that gives you acceptable results”.
2. Contributor tier. muse-spark-1.3-contributor costs $0.10 per million input tokens, $0.20 per million output and $0.002 per million cached input, in exchange for letting Meta train on your prompts and completions (more on that below). It also has far lower rate limits: 100 requests per minute against 3,000 on Standard. That suits throwaway prototypes, not parallel agents or customer data.
3. Web search. Grounding adds $2.50 per 1,000 search queries on top of token costs.
4. Prompt caching. Cached input costs $0.15 per million tokens on Standard, 88% less than uncached input. Caching is automatic: the prompt caching docs say the API caches the stable start of each prompt with no flag or key to manage. Agents resend their system prompt, tool definitions and history on every loop, so they benefit most, provided the stable content comes first and the changing content last. Check cached_tokens in the usage data to see what you are getting. At high volume, you can also send an optional prompt_cache_key so requests that start the same way hit the same cache more often.

If your team lives in Muse Code, there is also a flat-rate option: three Muse Code plans, Everyday Usage, High Usage and Power Usage. The Muse Code product page lists them at $5, $15 and $50 a month, with High Usage at five times and Power Usage at 20 times the Everyday allowance. The docs say benefits and availability “may vary by region”, so check what UK onboarding shows you before budgeting in sterling. The subscription covers only the Muse Code key; any other API key you create is billed per token.

A worked example: a weekly paid social reporting agent

Picture an agent that pulls last week’s paid social exports, checks them against the week before and drafts commentary for an account manager. The token figures are illustrative numbers we chose to show the arithmetic, priced at Meta’s Standard rates. They are not measurements. Say one run sends 2 million input tokens across its loops and produces 100,000 output tokens, reasoning included. For the 1.3 row we apply Meta’s 25% reduction to both input and output tokens.

Getting caching right saves more than the model switch. Both are worth having, but fix caching first, and warm the cache for both models before you compare them, or you will credit 1.3 with savings that came from the cache. The model charge is also the small part: the account manager’s time checking the commentary is what the test needs to count.

What should you test this week?

Pick one recurring job, such as the reporting agent above or a site audit agent that crawls a client site and drafts a fix list for the web team. Then:

1. Same harness, two models. Keep tools, system prompt, reasoning_effort, max tokens and temperature identical. Run muse-spark-1.2 and muse-spark-1.3 on the same tasks, with caches warmed for both. If you use another vendor today, add your current model as a third arm.
2. Log what matters. Tool-call count, input, cached, output and reasoning tokens, wall-clock time, failed tool rounds, and human interventions per finished job.
3. Judge finished work. Count jobs completed without rescue, not demos or leaderboard scores.
4. Tune effort afterwards. Change reasoning_effort only once the model comparison is done. Changing both at once hides which one made the difference.
5. Separate data policies. Keep Standard and Contributor routes apart so customer or confidential data never reaches Contributor by accident.
6. Sandbox tools. Require human approval for irreversible actions regardless. Meta’s safety claims do not replace your own controls.
7. Decide. Switch when 1.3 cuts loops or interventions on your jobs without hurting quality. Keep 1.2 for audio-heavy work, or for jobs where 1.3 shows no gain.

The Meta Model API speaks the OpenAI Responses format, so the usage data comes back in familiar fields. This sketch records one response’s numbers; add them up across a job’s loops:

import os
from openai import OpenAI

client = OpenAI(base_url="https://api.meta.ai/v1",
                api_key=os.environ["MODEL_API_KEY"])

def usage_row(model, response):
    u = response.usage
    return {
        "model": model,
        "tool_calls": sum(1 for item in response.output
                          if item.type == "function_call"),
        "input_tokens": u.input_tokens,
        "cached_tokens": u.input_tokens_details.cached_tokens,
        "output_tokens": u.output_tokens,
        "reasoning_tokens": u.output_tokens_details.reasoning_tokens,
    }

Can UK teams use Muse Spark 1.3?

Yes, under Meta’s published policy, and we confirmed access from the UK on 27 September 2026. Meta’s Geographic Use Policy (last updated 18 September 2026) lists Muse Spark 1.3 as a Standard Model and names the Restricted Territories where the API is unavailable; neither the UK nor any EU country is among them. Meta’s Connect 2026 developer recap (24 September 2026) says the Model API “is now generally available globally”.

The same policy limits access to “jurisdictions where Meta has enabled access”, and Meta “may enable or restrict access in any jurisdiction at any time” without publishing a country-by-country list, so treat our check as a snapshot. The consumer Muse agent is a separate product. Meta’s announcement on 8 September said it was “rolling out in the US”; Meta now lists it in the US, Canada and Mexico, and there is still no UK date. Our Meta Connect 2026 recap covers what that means for UK shops.

What should UK teams check before sending client data?

Data handling matters more than access for agencies and anyone working under UK GDPR.

• Keep client work on Standard. Meta’s Data Commitments say it “does not use your Content from Standard Services to train Meta Models”, while content from discounted services such as Contributor may be used to “train, develop, evaluate, and improve” its models. Standard content can still be processed for safety and abuse review. The Muse Code subscription docs add that how Meta uses “any code you submit” depends on the models you select. Meta’s API terms (updated 2 October 2026) go further: section 6.2 says you “must not submit sensitive, confidential, or personal information” to the discounted Contributor tier, including code you are required to keep confidential.
• Ask about Zero Data Retention. ZDR is available for qualified Meta Model API accounts, enabled per organisation through Meta’s sales team.
• Ask where data is processed. The Meta developer pages we checked do not say where requests are processed or what UK GDPR documentation, such as a data processing agreement, Meta offers. If you are putting personal data through an agent, get that answer from Meta in writing first.

If client material must not leave your own machines at all, Meta’s open-weight Muse Glimmer runs on your own hardware instead. Our Muse Glimmer test shows where it keeps up on agency work and where it falls short.

How do you keep a coding agent on a short lead?

In August Meta disclosed that a misconfigured third-party test environment let a pre-release Muse Spark 1.1 reach the live internet, where it “identified and exploited a security vulnerability in the real website”. The lesson is about the environment, not the model: check that your isolation actually holds.

The Muse Code permissions docs give you the controls. Shell commands run in an OS-enforced sandbox, and outbound network access is approved per destination by default. Two defaults deserve a second look on client repositories:

• New sessions start in Auto-review, where an automated reviewer, not a person, approves routine actions. If you want a person to approve every eligible action, choose the “Ask me” profile.
• Committed project memory loads even in untrusted workspaces. The configuration docs say to treat a repo’s MEMORY.md as a prompt-injection surface and review it on checkouts you don’t control.

AIWIZ verdict

Muse Spark 1.3 is worth an afternoon of your time if you use the Meta Model API. The price has not moved and the efficiency claim is specific enough to check.

If you already run 1.2, test 1.3 this week. The only cost is the test. Check which model and tier Muse Code starts on before you do.

If you run agents on GPT-6 Sol, Grok 4.7 or Claude Opus 5.5, the lower token price earns Spark a place in your comparison, not your production stack. Move only if it finishes your jobs as well for less.

Whichever model you choose, keep client data off any tier that trains on it, and fix your caching first. Those two decisions are likely to do more for your bill and your clients than any version number.

Next step

AIWIZ works on paid social and Muse Image creative, and on agent and Meta Model API builds. If you want a side-by-side test of 1.3 against 1.2 or your current model on your own jobs, we can set it up and give you a cost per finished job for each model. We will tell you what we would do, including when that means staying where you are. To talk it through, contact AIWIZ.