Muse Spark 1.2 vs DeepSeek V4 Flash: Four Index Points for Nine Times the Bill

$72.02 and $637.85. Those are the amounts Artificial Analysis spent running DeepSeek V4 Flash and Muse Spark 1.2 through the same nine benchmarks, and the models finished four points apart — 50 against 54 on the Intelligence Index. Meta's model is better. It is also 8.9 times more expensive to put through identical work, closed-weight, served by exactly one provider, and five days younger than DeepSeek V4 Flash. Whether four points is worth that depends entirely on what you are building, and this comparison is mostly about making that question answerable.

Both models arrived within a week of each other — DeepSeek V4 Flash (build 0731) on July 31, 2026, Muse Spark 1.2 on August 5 — and both are marketed at long-horizon agent work. Every index score, latency figure and evaluation cost below is from Artificial Analysis, which runs the benchmarks itself and publishes the bill. Where a number comes from Meta or from the vendor's own documentation rather than an independent run, the sentence says so.

The four numbers that decide this

• Intelligence Index — Muse Spark 1.2 54 (#13 of 185), DeepSeek V4 Flash 50 (#3 of 101 in its class, against a class median of 25).

• Blended price per million tokens — $0.78 vs $0.06. Thirteen times.

• Cost of the same nine-benchmark suite — $637.85 vs $72.02.

• Where you can run it — one provider, closed weights vs six API providers plus MIT-licensed weights you can host yourself.

Same suite, two very different invoices

The evaluation-cost column is the most honest cost comparison available for two models, because it is the same work, priced at each model's own rates, with each model spending as many tokens as it decides to spend. Muse Spark 1.2 needed $637.85 to complete the Intelligence Index. DeepSeek V4 Flash needed $72.02 — cheaper than every other well-known model Artificial Analysis has measured, a finding that got its own press cycle in early August.

What makes that gap interesting is that it happens despite the DeepSeek V4 Flash model being the more verbose one, not because of it. Running the suite, DeepSeek V4 Flash emitted 210 million output tokens against Muse Spark 1.2's 95 million — more than twice as many. Meta's model is meaningfully more token-efficient at the same task, and it does not matter at all, because DeepSeek V4 Flash charges $0.28 per million output against Meta's $4.25. A 15× price advantage swallows a 2.2× verbosity disadvantage without noticing.

This is the thing to carry into your own cost model. Token efficiency is a real property and reasoning models differ on it by large factors, but at current market spreads it is a second-order term. The rate card decides, and only tips over when two models are within roughly 3× of each other on price.

Where the four points are — and what nobody has published

Here is the uncomfortable part of this matchup: the Intelligence Index is a composite of nine evaluations across agents, coding, general capability and scientific reasoning, and Artificial Analysis has not yet published the per-benchmark breakdown for Muse Spark 1.2. So "54 versus 50" is currently a headline without a decomposition. Four points could be a coding gap, a long-context gap, a hallucination-resistance gap, or four small gaps that cancel into nothing on your workload.

Each vendor's own numbers fill part of the void, and neither can be checked against the other. Meta reports Terminal-Bench 2.1 at 82.9% for Muse Spark 1.2 (up from 76.2% for 1.1) and DeepSWE v1.1 at 59.3% (up from 53.0%) — run on Meta's harness, pass@1 over five attempts, with each competing model paired to its own agent product. The V4 Flash release positioned the model as a post-training upgrade tuned for agent and terminal work on a 284B-total / 13B-active mixture of experts. There is no benchmark on which both labs have published a comparable figure, and no third party has reproduced either set.

What is independently established is narrow and worth stating plainly: on one composite index, run by one evaluator, Muse Spark 1.2 finished four points ahead, at 8.9 times the cost. Everything finer-grained than that is still vendor material.

Three ways to pay for Muse Spark 1.2, one and a half for DeepSeek

Price comparisons here are unusually slippery, because Meta ships two rate cards and the vendor ships the weights itself.

• Meta standard tier — $1.25 input, $0.15 cached input, $4.25 output per million, with a rate ceiling around 3,000 requests per minute.

• Meta contributor tier — $0.10 input, $0.002 cached input, $0.20 output, in exchange for permission to train future Meta models on your traffic, capped at 60 requests per minute.

• DeepSeek V4 Flash list — $0.14 input, $0.28 output, $0.003 on a cache hit (a 98% discount, the most aggressive caching rate of any model in this bracket), from six competing API providers.

• DeepSeek V4 Flash self-hosted — MIT-licensed open weights, so the price is your own hardware and the ceiling is your own capacity.

Read the second and third lines together and the ranking inverts: on the contributor tier, Muse Spark 1.2 is cheaper per token than DeepSeek V4 Flash, at $0.10/$0.20 against $0.14/$0.28, while scoring four points higher. That is the most aggressive pricing in this entire comparison and almost nobody will be able to use it, because 60 requests per minute is 2% of the standard budget and rules out essentially any production fan-out. It is priced for evaluation and small-team work, and the currency is your prompts. Whether that trade is available to you is a data-governance decision that belongs with legal, not a line item you swap in a config file.

The open-weights side runs the other way. MIT weights mean the price floor is not set by the vendor at all — if your volume justifies the GPUs, nobody can raise your rate, deprecate your model, or change the terms. That optionality has no equivalent on the Meta side at any tier.

One provider versus six

Muse Spark 1.2 is served by Meta's Model API and nowhere else. No third-party hosts, no quantisation choice, no fallback path, and — at the time of writing — no OpenRouter listing, no Hugging Face repository and no entry in the OrcaRouter catalog. A single-provider closed model is a single point of failure by construction, and for a model five days old there is no uptime history to reassure you.

DeepSeek V4 Flash is the opposite case. Six API providers serve it today, the weights are MIT, and it is in the OrcaRouter catalog at $0.15 in and $0.29 out — provider list price, passed through at 0% markup — with automatic failover between providers, so a single host degrading does not take the workload down with it. That is the practical argument for the cheaper model that has nothing to do with its score: you can move it, mirror it, or leave entirely, and none of those options require a new integration.

For Muse Spark 1.2 today, an evaluation means a second contract, a second bill and a second SDK path. Meta's model is worth testing on its merits, but be clear that the operational cost of running it is not only the token price.

Latency, measured on real traffic

In the benchmark harness, Artificial Analysis records time to first token at 1.33 seconds for DeepSeek V4 Flash and 26.12 seconds for Muse Spark 1.2 at its xhigh reasoning setting, with output speeds of 103.3 and 165.0 tokens per second respectively. Meta's model generates faster once it starts and takes a very long time to start, which is exactly what mandatory deliberation at maximum effort looks like.

Production numbers tell a gentler version of the same story. Across seven days of live routing on OrcaRouter, DeepSeek V4 Flash shows a p50 time to first token of 439 milliseconds and a p95 of 1.96 seconds, at 111 tokens per second with a 1.6% error rate. There is no equivalent production figure for Muse Spark 1.2 because it is not routed anywhere yet; Muse Spark 1.1, the closest available proxy, runs a p50 of 1.84 seconds and a p95 of 6.00 seconds. Sub-second median response is a category DeepSeek V4 Flash is in and no Muse Spark version has entered.

Both models declare roughly a million tokens of context — 1,048,576 for both, with DeepSeek V4 Flash capping completions at 384,000 tokens and the Muse Spark endpoint declaring no completion ceiling at all. On context, this matchup is a tie on paper and unverified in practice for both.

Three questions worth asking before you switch

Is self-hosting DeepSeek V4 Flash actually cheaper than $0.15 per million?

Usually not, and that is the point. A 284B-parameter mixture of experts with 13B active needs serious multi-GPU capacity to serve at reasonable latency, and at $0.15 per million input tokens the hosted rate is hard to beat on anything but very high, very steady volume. The value of the MIT licence for most teams is not that they will self-host — it is that they credibly could, which caps what any provider can charge and removes deprecation risk. Treat it as insurance, and buy the hosted tokens.

Does the contributor tier make Muse Spark 1.2 the cheaper model?

On the rate card, yes; in practice, only for workloads under 60 requests per minute whose data you are permitted to donate. If both conditions hold — a research spike, an internal tool, a benchmark harness on synthetic inputs — then a model four index points ahead at a lower per-token price is genuinely the better buy and you should take it. If either fails, the standard tier is the real price, and the standard tier is 13× the V4 Flash model's blended rate.

Four index points sounds small. When does it matter?

When errors compound. On a single-shot classification or summarisation call, a four-point composite gap is often invisible, and paying 9× for it is indefensible. On a fifty-step agent loop, a per-step success difference that reads as small multiplies into a large difference in how often the whole run has to be restarted — and a restart costs the entire trajectory, not one step. That is the case where Meta's price is arguable. It is also the case where you should measure on your own task rather than trusting a composite, because nobody has published the agentic sub-index that would settle it for 1.2.

The verdict, and the condition attached to it

For almost every workload where cost is a live constraint, DeepSeek V4 Flash is the correct default: four points behind on one composite index, thirteen times cheaper blended, sub-second median latency in production, six providers, MIT weights, and no vendor able to change the terms on you. It is currently the cheapest well-known model to run through a full benchmark suite, and it did not get there by being bad.

Muse Spark 1.2 earns its price in one specific shape of work: long-horizon agentic coding where a failed trajectory is expensive, where the extra deliberation is amortised over minutes of work rather than felt by a waiting user, and where you can live with a single provider and no independent breakdown of what the four points are made of. Meta has shipped three models in four months and moved eleven index points in that time, so the case may strengthen quickly — but today it rests on vendor-run coding benchmarks and one composite score.

If you are choosing this week, run both on fifty steps of your own agent loop and compare completed trajectories per dollar rather than tokens per dollar. That single measurement answers this comparison better than every published benchmark in it.