On 2 September Meta shipped Muse Spark 1.3 — an update to its top model, tuned for code and long agentic runs. Zuckerberg described the release as "frontier performance almost too cheap to meter", and this time the marketing line roughly matches the numbers: per completed task, Spark really is the cheapest model in its weight class.
The catch is that Meta's own figures run higher than independent measurements, and half the coverage benchmarks a configuration that is not in the API. Here is where that line falls.
The minimal request
The GPTunneL API is OpenAI-compatible, so wiring it up means a base_url and a
model name:
curl https://gptunnel.ru/v1/chat/completions \
-H "Authorization: <YOUR_API_KEY>" \
-H "Content-Type: application/json" \
-d '{
"model": "muse-spark-1.3",
"messages": [
{"role": "user", "content": "Find the cause of the crash in this stack trace and suggest a patch"}
]
}'The same thing through the official OpenAI client:
from openai import OpenAI
client = OpenAI(
api_key="<YOUR_API_KEY>",
base_url="https://gptunnel.ru/v1",
)
resp = client.chat.completions.create(
model="muse-spark-1.3",
messages=[{"role": "user", "content": "Find the cause of the crash and suggest a patch"}],
stream=True,
)
for chunk in resp:
print(chunk.choices[0].delta.content or "", end="")The full parameter list is in the documentation and on the /docs page.
Benchmarks
Meta ships Spark 1.3 in two configurations, and that matters for reading any table. xhigh is what everyone gets through Muse Code and the Model API. max is in limited partner preview pending extra safety testing; Meta has not said when it opens up.
| Benchmark | Spark 1.3 xhigh | Spark 1.3 max | Reference point |
|---|---|---|---|
| AA Intelligence Index | 61 | 62 | Claude Fable 5.1 — 66 |
| Tau3-Bench Banking | 47% | 52% | best result in the table |
| Terminal-Bench 2.1 | 85% | 86% | Meta claims 88.8% |
| GDPval-AA v2 | 1709 Elo | 1754 Elo | — |
| GPQA Diamond | 94% | 94% | — |
| Humanity's Last Exam | 47% | 47% | — |
| AA-LCR (long context) | 79% | 79% | 1.2 scored 83% |
The xhigh/max columns come from independent runs by Artificial Analysis; the 75.4% on DeepSWE 1.1 and 88.8% on Terminal-Bench 2.1 come from Meta's own scorecard.
The three-point gap on Terminal-Bench is not a fudge, it is a different harness — but for a purchasing decision, take the lower number: you are not going to tune your own scaffold and timeouts to fit a benchmark.
The public xhigh sits level with GPT-5.6 Sol and Grok 4.6 and lands five index points short of Claude Fable 5.1. "Almost Fable" is about the agentic rows and the price, not about overall intelligence.
The catch: the max you don't have
Half the release coverage measures Spark by the max variant — which is not in the API. The gap on the index is small (61 against 62), but on agentic work it shows: 47% against 52% on Tau3-Bench Banking, 1709 against 1754 Elo on GDPval-AA v2. Max pays for it in reasoning tokens: 62% more on GDPval, 28% more on the banking tasks.
If a number in someone's review looks too good, check which configuration it refers to.
Price: where Spark actually wins
At Meta the model costs $1.25 per million input tokens and $4.25 per million output ($0.15 for cached input) — unchanged from 1.2. But the figure that makes the release worth a look is a different one: $0.55 per Intelligence Index task. That is the best result among every model scoring 59 or above; GPT-5.6 Sol (max) costs $0.95 and Grok 4.6 (high) $0.94, roughly 70% more.
The saving does not come from the token rate alone. By Meta's measurements, 1.3 closes engineering comparisons with about 20% fewer tool calls and 25% fewer tokens than 1.2. For an agent looping for hours, that is the bulk of the bill.
In GPTunneL Spark 1.3 costs $2.50 per million input tokens and $8.50 per million output. For comparison, Claude Fable 5.1 in the same catalogue runs $20 and $100 per million: input is eight times cheaper, output almost twelve. Current figures are on the pricing page, along with a per-request calculator.
One more note on Meta's own Contributor tier: $0.10 and $0.20 per million, up to twenty times cheaper than the base rate. The price of that discount is Meta's right to train on your prompts and the model's completions. For a product that runs other people's data through the model, that is not a pricing choice, it is a question for legal.
Where Spark falls short
- Long context got worse, not better. AA-LCR is 79%, down from 83% on 1.2. The model holds a million tokens in the window, but pulls facts out of it a little less reliably. If you run RAG over a large window, measure on your own documents before you move traffic.
- Factual recall. AA-Omniscience is 42% for xhigh and 44% for max. This is a model for code and tools, not an encyclopedia; knowledge questions are better served by web search.
- The agent rows. On Meta's own scorecard Spark 1.3 takes DeepSWE v1.1, SWEAtlas CodeBase QnA and both MRCR bands, ties Terminal-Bench 2.1 — and loses all six agent rows to Claude Opus 5 or GPT-5.6 Sol.
What to check before you migrate
Streaming. A model with adaptive reasoning answers unevenly: the pause before
the first token is noticeably longer than with Flash-class models. Without
stream: true that runs into your proxy timeout, not into any model limit.
The cost of a long prompt. A million-token window tempts you to put everything in context. At $2.50 per million input tokens, a full window is $2.50 for a single request. Input caching takes part of that back, but do the maths up front.
What the model accepts. The catalogue lists reasoning, image input and PDF parsing for Spark 1.3. If your pipeline leans on tool calling, verify the behaviour against your own function set before moving production over.
FAQ about Muse Spark 1.3
How is Muse Spark different from LLaMA? They are different branches. LLaMA is the open line with published weights that you can deploy yourself. Muse Spark is the closed flagship: no weights, but a higher level on code and agentic runs. Both lines are collected on the Meta models page.
What does xhigh mean? It is the reasoning effort level. xhigh is the configuration available through the API; max is the partner build, not publicly available yet.
Should I move off Claude Fable 5.1? If the job is code and agents with a defined tool set, the price difference justifies a test: five index points against an eightfold difference in input cost. If long context and factual precision matter, stay on Fable and compare on your own data.
Which languages does it handle? The line is multilingual and answers confidently in dozens of languages, though system instructions for agents are more reliable in English.
The bottom line
Muse Spark 1.3 is not a "Fable killer" — it is the cheapest entry into the top tier: index 61 at half the per-task cost of its rivals. Take it for agents and code, where the cost of a loop matters more than the last five benchmark points, and re-check long context on your own data, because that is where the model regressed against its predecessor.
Switching is one line: open Muse Spark 1.3 in GPTunneL, change the model name in your config and run your own request set — the balance is shared with every other model in the catalogue, no separate subscription needed.