Meta’s Muse Spark 1.2 arrived on August 5, four weeks after 1.1, with an identical rate card, an identical context window, an identical modality list and an identical reasoning dial. It scores meaningfully higher on every independent index — and it is also roughly four times slower to first token in production, and costs about 38% more per completed task. Whether that adds up to an upgrade depends entirely on what you’re building; we worked through the matchup in detail in our full 1.1 vs 1.2 head-to-head, and this is the decision-focused version.
Version upgrades are usually easy calls: the new one is better, costs the same or less, and you move on. This one isn’t — and the reason is a number neither Meta nor the benchmark sites publish.
Table of Contents
What is genuinely identical
Before the differences, it’s worth being clear how much didn’t move, because it’s unusual:
• Price — $1.25 input, $0.15 cached input, $4.25 output on both. Meta did not reprice.
• Context window — 1,048,576 tokens on both; 131,072 maximum output on both.
• Modalities — text, image, video, file and audio in, text out, on both.
• Reasoning — mandatory on both, same `minimal`→`xhigh` dial, same `medium` default.
• Weights — closed on both.
Put the two spec sheets side by side and you cannot tell them apart. Everything that separates these models is behavioural, which means everything that separates them has to come from measurement.
What moved: the index
Artificial Analysis is the cleanest generational read, because it runs the same fixed task set across every model it tracks.
• Muse Spark 1.0 (April) — 43
• Muse Spark 1.1 (July 9) — 51
• Muse Spark 1.2 (August 5) — 57 on the live board today
One caveat that trips up most coverage: AA’s launch-day article scored 1.2 at 54, and that’s the figure still circulating in almost everything written in the past week. The live model page has since been revised to 57, ranking it #12 of 185 models. If you’re comparing numbers across articles, check the date.
Either way the direction is clear and the magnitude is real: fourteen index points across four months, and six of them in the last four weeks, on an unchanged price.
Where did the gain come from? Agentic and long-horizon work, consistently. AA’s launch measurements showed GDPval-AA v2 rising to 1631 Elo (#5) from 1371 — a 260-point jump — with Terminal-Bench v2.1 moving 78% to 80% and τ³-Banking 25% to 27%. There’s also a behavioural shift worth naming: 1.2 abstains more. Its hallucination rate fell from 38% to 28%, while raw accuracy fell from 41% to 38%. It got more careful, not only more capable.
What moved the other way: speed
This is the part that isn’t on any benchmark site, because the benchmark sites don’t have it. Artificial Analysis lists no output speed and no time-to-first-token for Muse Spark 1.2 — both fields read N/A. Production telemetry is the only source.
On OrcaRouter’s seven-day window, at the same list price on the same platform:
• Muse Spark 1.2 — p50 time to first token 7.73 s, p95 10.00 s
• Muse Spark 1.1 — p50 time to first token 1.93 s, p95 8.13 s
That is a four-fold regression at the median. Vals AI’s independent runs corroborate the direction from a different angle, averaging roughly 610 seconds per test.
This isn’t a bug and it isn’t a provider problem. It is the mechanical consequence of a model that deliberates more: AA measured 1.2 burning 95 million output tokens to complete the index, and the cost per task rose from $0.29 to $0.40 on unchanged pricing — a 38% increase for a six-point gain.
So the trade is legible. You are buying reasoning depth with latency and tokens.
The one thing 1.1 has that 1.2 doesn’t
Here’s an inversion most upgrade guides miss: on the *official* Terminal-Bench 2.1 leaderboard, version 1.1 has an independently verified entry and version 1.2 does not.
The board lists mini-SWE-agent + Muse Spark 1.1 at 76.2% ± 1.2%, xhigh, submitted by Princeton on July 9, 2026, at rank 8 with a run cost of $198.05. There is no Muse Spark 1.2 row. There is no Muse Code row either.
Meta’s claim for 1.2 is 82.9% on that benchmark, with a stated 6.7-point gain over the previous version. Subtract and you get 76.2 — exactly the board’s figure for 1.1 under a deliberately minimal third-party scaffold. Meta never published the harness behind its baseline, so this is an inference rather than a proven fact. But if the baseline is that row, the “+6.7” is comparing a bare scaffold to Meta’s own co-trained agent, and part of that gain belongs to the harness rather than the model.
The practical consequence: if you need a coding score you can point to in a review, 1.1 currently has one and 1.2 has a vendor claim.
Who should upgrade, and who shouldn’t
Upgrade if your workload is agentic, long-horizon or document-heavy — multi-step tool use, whole-repository refactors, batch analysis, overnight jobs. The six index points and the 260-Elo agentic jump are concentrated exactly there, and latency is nearly free in a background job.
Upgrade if hallucination is a bigger risk to you than a missed answer. The abstention shift is a genuine safety improvement for anything user-facing that gets fact-checked.
Stay on 1.1 if anything human-facing depends on first-token latency. Two seconds versus eight is the difference between a usable interactive experience and a spinner, and you gain nothing else by moving, because the price is the same either way.
Stay on 1.1 if you need the verified third-party coding number more than the vendor-claimed one.
Run both if you can. They cost the same, they take the same request format, and they differ only in a trade-off that varies by task. Behind a single OpenAI-compatible endpoint that carries both — OrcaRouter passes Meta’s list price through at 0% markup — routing latency-sensitive traffic to 1.1 and long-running work to 1.2 is a config decision, not an engineering project.
The takeaway
Muse Spark 1.2 is a real improvement over 1.1 on every independent measure of intelligence, at exactly the same price — and it is four times slower to answer and a third more expensive per finished task. That makes this an unusually clean decision rather than a difficult one: if your work happens in the background, take the upgrade; if a person is waiting on the response, the older version is still the right tool and costs you nothing to keep. The one thing you should not do is assume the newer version is strictly better, because on latency and on independently verified coding results, it currently isn’t.
Sourcing note: pricing, context window and the 82.9% / +6.7 benchmark claims are Meta’s own published figures, unreproduced by third parties. Index scores, cost per task, token consumption and agentic sub-results are from Artificial Analysis; the 76.2% row is from the official Terminal-Bench 2.1 leaderboard; per-test latency from Vals AI; p50 and p95 first-token figures are OrcaRouter’s own seven-day production telemetry. The 82.9 − 6.7 = 76.2 observation is our inference. Checked August 7, 2026.
