Three things builders should know
- Meta is charging developers for the first time. The Meta Model API launched in public preview on 9 July 2026 with Muse Spark 1.1 — closed-weight, hosted, metered — at $1.25 per million input tokens and $4.25 per million output tokens, with $20 in free credits for new accounts.
- It speaks both dialects. The API accepts OpenAI Chat Completions and Responses formats and the Anthropic Messages format. Pointing an existing agent at
https://api.meta.ai/v1with a new key and model ID is the whole migration. - The preview is US-first. Self-serve access is limited to US developers; everyone else joins a waitlist. Indian and UK teams cannot bill against it yet — but there is useful preparation work to do now, and we cover it below.
For years Meta's developer pitch was simple: the models are free, take the weights and go. Llama built an entire ecosystem on that promise — we covered Llama 4's open-weight multimodal MoE when it reset the cost equation earlier this year. Muse Spark 1.1 is the opposite bet. It is the first model from Alexandr Wang's Meta Superintelligence Labs to be sold rather than given away, per Fortune, and it slots into a wider pattern of Meta monetising its AI build-out — the same company that is renting out surplus GPUs in the $300B cloud war. Whether or not Muse Spark wins your workload, the API's design choices matter to every builder, because they signal where the provider market is heading: cheaper tokens, bigger contexts and drop-in compatibility as a weapon.
What Meta actually shipped
Muse Spark 1.1 is a natively multimodal reasoning model built for agentic work. Per Meta's announcement and the MarkTechPost breakdown, it takes text, images, video and documents as input, outputs text, and ships with parallel tool calling, structured output, prompt caching, an integrated web-search tool with citations, and computer use — screenshot analysis, script generation and click automation. The context window is 1,048,576 tokens, and Meta says the model actively compacts its own context across long sessions rather than simply truncating.
On capability, treat the launch numbers with the usual caution: they are Meta-reported and independent verification is still thin eight days in. Meta's figures put Muse Spark 1.1 ahead of Claude Opus 4.8 on tool-use and reasoning benchmarks (88.1 vs 82.2 on MCP Atlas; 62.1 vs 57.9 on Humanity's Last Exam) but behind it on computer use (80.8 vs 83.4 on OSWorld-Verified) and clearly behind on coding, where Meta's own chart shows 61.5 vs 69.2 on SWE-Bench Pro. A vendor publishing a benchmark it loses is mildly reassuring, but the honest summary is: strong claimed tool-use, competitive-not-leading coding, and you should run your own evals before moving anything that matters.
Equally important is what it is not. Muse Spark 1.1 is not open-weight. There is no download, no self-hosting, no fine-tuning your own copy — a genuine strategy break that CNBC frames as Meta chasing Anthropic and OpenAI's API revenue directly.
How to actually try it
If you are US-based (or the moment the waitlist admits you), the setup is deliberately boring. Per Meta's developer docs, you need three values: the base URL, an API key, and the model ID muse-spark-1.1. New accounts get $20 in free credits — at these prices, that is several million tokens of experimentation before you spend a rupee or a pound.
# OpenAI-style (Chat Completions)
from openai import OpenAI
client = OpenAI(
base_url="https://api.meta.ai/v1",
api_key=os.environ["MODEL_API_KEY"],
)
resp = client.chat.completions.create(
model="muse-spark-1.1",
messages=[{"role": "user", "content": "Summarise this PDF..."}],
)
# Anthropic-style (Messages)
import anthropic
client = anthropic.Anthropic(
base_url="https://api.meta.ai",
api_key=os.environ["MODEL_API_KEY"],
)
msg = client.messages.create(
model="muse-spark-1.1",
max_tokens=1024,
messages=[{"role": "user", "content": "Summarise this PDF..."}],
)That dual-format support extends beyond raw SDKs. Meta's docs list compatibility with OpenAI-format stacks (Codex, LangChain, LlamaIndex, Vercel AI SDK, OpenCode) and Anthropic Messages-format tools, which per early developer guides includes agent harnesses like Claude Code pointed at the Meta endpoint. We saw this playbook from the other direction when Ollama added Anthropic API compatibility so local models could slot into Claude-native tooling. The industry is converging on two wire formats, and Meta has simply decided to speak both on day one rather than fight for a third.
Pricing: where Muse Spark sits, as of July 2026
Sticker prices move quickly, so treat this table as a snapshot — all figures are per million tokens, standard tier, as of mid-July 2026, from each provider's published pricing.
| Model | Input $/M | Output $/M | Context |
|---|---|---|---|
| Muse Spark 1.1 (Meta) | $1.25 | $4.25 | 1M |
| Claude Opus 4.8 (Anthropic) | $5.00 | $25.00 | 200K+ |
| Claude Sonnet 5 (Anthropic) | $2.00 intro | $10.00 intro | 200K+ |
| Claude Haiku 4.5 (Anthropic) | $1.00 | $5.00 | 200K |
| GPT-5.6 Sol (OpenAI) | $5.00 | $30.00 | 1M |
| GPT-5.6 Terra (OpenAI) | $2.50 | $15.00 | 1M |
| Gemini 3.1 Pro (Google, ≤200K input) | $2.00 | $12.00 | 1M |
Read the column that matters for your workload. Against the frontier tier — Opus 4.8 at $5/$25 (Anthropic pricing) and GPT-5.6 Sol at $5/$30 (OpenAI pricing) — Muse Spark's output tokens are five to seven times cheaper, which is why coverage like The Decoder's frames this as a price-war move. Against the mid tier, the gap narrows: Sonnet 5's introductory $2/$10 (rising to $3/$15 after 31 August 2026), GPT-5.6 Terra at $2.50/$15 and Gemini 3.1 Pro at $2/$12 are all in the same neighbourhood, and Gemini's long-context surcharge ($4/$18 above 200K input) is a reminder to model your actual token distribution, not the headline rate. Also note Meta's docs flag preview pricing as subject to change.
Muse Spark 1.1 is a reasoning model, and per early pricing analyses its chain-of-thought tokens are billed at the $4.25 output rate. On reasoning-heavy agentic loops your effective cost per visible output token can be a multiple of the sticker price. Benchmark cost per completed task, not cost per million tokens.
The dual-SDK play: gateways, failover and A/B tests
The most consequential design decision here is not the price — it is the compatibility. By speaking both OpenAI and Anthropic formats natively, Meta has reduced the cost of trialling Muse Spark to roughly zero for anyone with a sanely architected stack. If your model calls already flow through a gateway layer, adding a new provider is a config entry: our comparison of LiteLLM, OpenRouter and Portkey walks through exactly this pattern, and the moment those gateways add a meta/muse-spark-1.1 route, A/B testing it against your incumbent becomes a one-line change.
The same property makes Muse Spark interesting as a failover target. A resilient setup — retries, rate-limit handling, and an ordered fallback chain across providers, as we detailed in our guide to building a resilient LLM gateway — gets strictly better when a third first-party provider speaks your existing wire format. A chain of GPT-5.6 Terra, then Sonnet 5, then Muse Spark 1.1 needs no adapter code at all. The strategic read: wire-format compatibility has become table stakes, and lock-in is shifting up the stack to things formats do not cover — caching semantics, tool-execution behaviour, evals and observability.
Before trusting any drop-in provider swap, run a contract test: same prompt set, temperature 0, through both your OpenAI-format and Anthropic-format paths, and diff tool-call behaviour specifically. "Compatible" wire formats still differ on stop sequences, parallel tool-call ordering and streaming edge cases — find that out in CI, not in production.
Every article here is written by a Verified Builder. Want your name on the next one?
AI Tech Connect lists AI engineers, founders and researchers across India and the UK — and the people hiring browse it to find them. Adding your profile is free.
Become a Verified Builder →The US-first catch — and what IN and UK builders can do now
Here is the frustrating part. The public preview is US-only. Per Meta's statements to press, some early partners already have access and everyone else "will be able to add themselves to a waitlist and be added from there over time" — with no announced timeline for international availability. That leaves builders in Bengaluru, Mumbai, London and Manchester in the same position: you can read the docs, but you cannot bill against the API today.
That does not mean there is nothing to do. Four concrete moves for Indian and UK teams this week:
- Join the waitlist now. Meta has said admission happens over time; being early in the queue costs nothing.
- Test the model itself free. Muse Spark 1.1 powers the Meta AI app's Thinking mode and meta.ai, so you can qualitatively evaluate reasoning and multimodal behaviour on your domain before the API reaches you.
- Make your stack provider-agnostic today. If adding a provider currently means touching application code, that is the real blocker — fix it with a gateway layer now, so that Muse Spark (or whatever ships next quarter) becomes a config change.
- Watch the aggregators. US-first previews have historically reached international builders earlier via resellers and gateway platforms than via direct sign-up; the same may or may not happen here, but it is worth tracking.
There is also a precedent worth remembering: US-first has been the default rollout shape for nearly every frontier API preview, and both India and the UK have typically followed within months once billing and compliance were sorted. No provider chasing developer revenue can ignore India's developer population or the UK's enterprise market for long. Plan for access; do not architect around its absence.
The bottom line
Muse Spark 1.1 is two announcements wearing one launch post. The model is a credible, aggressively priced agentic contender whose benchmarks you should verify yourself. The API is the bigger story: Meta abandoning free-weights-only distribution, undercutting frontier output pricing by 5–7x, and treating its rivals' wire formats as a feature rather than a threat. For builders, the action list is short — join the waitlist, keep your stack gateway-shaped, contract-test before you switch, and price your workloads on reasoning-token reality rather than sticker rates. The cheapest experiment in AI right now is being ready to run experiments.