Muse Spark vs Gemini 4 Argon: the Price Doubles on December 1

Almost every M​use Spark versus G​emini 4 Argon comparison you can find today compares $1.25 against $2.00 per million input tokens, and that comparison expires in six weeks. G​oogle's own pricing page prices G​emini 4 Argon at $2.00 in and $10.00 out with cached input at $0.10 on an introductory basis "through November 30, 2026"; from December 1, 2026 the same model is listed at $4.00 in, $20.00 out and $0.20 cached. M​eta's M​use Spark family has not moved off $1.25 in and $4.25 out across Muse Spark 1.1, 1.2 and 1.3. So the interesting question about this matchup is not which model is smarter in October — it is what the arithmetic looks like on the first of December, and whether whatever you build on G​emini 4 Argon this month survives that date.

Two models, two very different publication patterns

G​emini 4 Argon is G​oogle's newest frontier model and it went generally available on October 13, 2026 — the changelog entry is a single line, and the developer documentation is dated the same week. Artificial Analysis' listing carries a September 2026 release note, which we read as a pre-GA or preview origin rather than a second launch; we could not find a formal preview announcement to confirm that, so the GA date is the one we treat as solid. G​oogle's model page marks it "Stable".

M​use Spark 1.3 is the current checkpoint of M​eta Superintelligence Labs' M​use Spark line, released September 2, 2026, with its "max" reasoning level opened to developers on September 4. The unversioned "M​use Spark" name refers to the original April 2026 model in the same family; version numbers are doing real work here, and 1.3 is the one to compare.

The independent scoreboard

Both listings come from Artificial Analysis and both measure a high-compute configuration, so this is one of the few matchups where the reasoner-versus-reasoner comparison is apples to apples on the index itself. Everything in this section is Artificial Analysis' measurement, not a vendor claim.

• Intelligence Index — G​emini 4 Argon (High): 53, rank 8 of 217. M​use Spark 1.3 (max): 48, rank 24 of 217.

• Cost per index task — G​emini 4 Argon: $1.99. M​use Spark 1.3: $1.60.

• Output speed — G​emini 4 Argon: not measured; the listing shows no figure. M​use Spark 1.3: 121.9 tokens per second, ranked 43 of 186.

• Verbosity — G​emini 4 Argon: 110M output tokens during the index. M​use Spark 1.3: 170M. Argon is the more concise of the two.

• List pricing as measured — Argon at the introductory $2.00 / $10.00; M​use Spark 1.3 at $1.25 / $4.25 with cache reads at $0.15.

The five-point intelligence gap is real and modest. The cost gap is the one that inverts, and it inverts hard: at the introductory rate Argon costs about 24% more per completed task; at the December rate, the same comparison on input and output pricing roughly doubles the input side and doubles the output side, so the per-task premium stops being marginal. That is the single most decision-relevant fact in this comparison and it is not in any of the launch coverage.

What each one will actually accept

Here the two sources of record disagree, and it is worth showing the disagreement rather than picking a side.

G​oogle's own model page for G​emini 4 Argon lists text, image, video, audio and PDF as inputs with text output and thinking effort at low, medium or high, defaulting to medium. Artificial Analysis' summary sentence for the same model says it "supports text and image input" and outputs text. We checked both because that gap is exactly the kind of thing that gets copied forward; the vendor's own page is the broader and more specific list, and it is the one to test against. Treat the vendor list as the claim and verify it on your own files.

M​use Spark 1.3 is the narrower model on paper: text, image and video in, text out, with no audio or PDF path in the documented modality row and a reasoning ladder that runs to max.

• Inputs — G​emini 4 Argon: text, image, video, audio, PDF per G​oogle; text and image per the independent listing. M​use Spark 1.3: text, image, video.

• Output — text only, both.

• Context window — 1,048,576 tokens for both, described by G​oogle as a unified window shared across input, thinking and output.

• Max output — G​emini 4 Argon: 65,536 tokens on default synchronous calls, extensible toward the full window with long decode. M​eta does not publish a max-output figure for M​use Spark 1.3 that we could verify.

• Reasoning control — Argon: low / medium / high, no off switch. M​use Spark 1.3: minimal through max, and M​eta's tooling notes reasoning cannot be turned off either.

Vendor claims, kept on the vendor's side of the line

Neither company published a comparable table, so this is a list of claims rather than a scoreboard, and none of it is reproduced by us.

G​oogle's published figures for G​emini 4 Argon run 77.9% on DeepSWE v1.1, 57.4% on Terminal-bench 4.0, 69.2% on the OSWorld 2.0 offline subset, 51.3% on AutomationBench, 68.9% on the Vals Index, 71.6% on Chartography and 91.7% on LVBench, among others. M​eta's published figures for M​use Spark 1.3 centre on DeepSWE v1.1 at 75.4, OSWorld 2.0 at 66.9, Terminal-Bench 2.1 at 76.4 with the Terminus 2 agent, MCP-Atlas at 89.2 and GDPval-AA v2 at 1,754. Note where the two overlap and where they do not: DeepSWE is shared and close, OSWorld is shared and about two points apart on different revisions, and Terminal-Bench is reported by M​eta on an older version with a named agent harness. Subtracting across those differences is not arithmetic, it is fortune-telling.

One independent signal worth carrying alongside the vendor rows: LMArena's text leaderboard currently places a G​emini 4 Argon configuration first overall at 1524 Elo, marked preliminary. M​use Spark 1.3 has no entry on that board. Preliminary or not, that is the strongest evidence in the record that Argon's five-point index lead understates its standing with human voters.

The engineering details that decide it faster than benchmarks

• Sampling parameters are ignored — G​oogle's documentation states that temperature, top_p and top_k have no effect on G​emini 4 Argon, and that explicit caching is not supported, leaving implicit caching only. If your stack tunes temperature, that knob silently stops working rather than erroring. M​use Spark 1.3 exposes a reasoning-effort parameter but no equivalent published restriction on sampling.

• Timeouts — G​oogle recommends client timeouts of at least fifteen minutes for this model. If you are on a 60-second gateway default, you will see failures that look like model errors and are not.

• Conversation state — Argon uses a server-side interaction identifier for multi-turn state; M​use Spark's family is documented around the standard message-array pattern. Porting between them is not a find-and-replace.

• Cache reads — M​use Spark 1.3 returns cached input at $0.15 against $1.25 standard, an 88% discount. Argon's cached input is $0.10 against $2.00 introductory, rising to $0.20 against $4.00. In a long agent loop with heavy reuse this is the line that matters most, and it is one of the few places where the two stay close after December.

Where OrcaRouter fits, and where it does not

Being precise about our own coverage: OrcaRouter routes Muse Spark 1.2 and Muse Spark 1.1 at Meta's list price with 0% markup passed through, alongside 200+ other models on one API with automatic failover and a routing DSL for splitting traffic between them. M​use Spark 1.3 and G​emini 4 Argon are both absent from our catalogue today — Argon is reachable through G​oogle's own API and several third-party platforms, and we are not one of them. So the honest framing is not "run both here". It is that the M​use Spark side of this comparison is available to you on a key you may already hold, which makes a cost test against Argon on your own prompts a same-week exercise rather than a procurement one, and that M​eta's pricing has been static across three checkpoints while G​oogle's has a published expiry date.

What to do before December

If you are already building on G​emini 4 Argon, the useful thing to do this month is not a benchmark run — it is a cost model at $4.00 and $20.00, because that is the price that will apply to everything you ship and leave running. If the answer at that price still works, Argon is the better model on the current evidence: higher index, more concise, broader documented input list, and the only one of the two with a top-of-board human-preference placement. If it does not, the substitution to evaluate is M​use Spark 1.3 at a flat $1.25 / $4.25, which trades five index points for a price that has not moved in three releases and roughly three times the output speed. The one thing we would not do is treat October's advertised rates as the rates — in this matchup, the number in the title is the one with an end date.

What the numbers cannot tell you is whether any of it has gone wrong in the field: the documented AI incidents and agent failures around models in this class are collected in the OrcaRouter AI Incident Archive, where each entry carries its own sources, severity, confidence grade and disputed flag.