Muse Spark 1.2 and Muse Code: Meta's Third Model in Four Months Buys a Tie with Grok 4.5

Meta released Muse Spark 1.2 on August 5, 2026, four weeks after Muse Spark 1.1 and roughly four months after the first Muse Spark. This one arrived with something the previous two did not have: a coding agent of Meta's own, Muse Code, shipped in beta and co-trained with the model it runs on. And on the one scoreboard neither Meta nor its rivals control, the release moves Meta into a dead heat with SpaceXAI — Muse Spark 1.2 and Grok 4.5 now both score 54 on the Artificial Analysis Intelligence Index.

Two things are worth separating before any of the numbers below mean anything. The Intelligence Index score, the latency figures and the token counts come from Artificial Analysis, which runs its own evaluations and publishes the bill; those are independent. The coding results Meta led its announcement with — Terminal-Bench 2.1, DeepSWE v1.1 and an internal benchmark — were run by Meta, on Meta's harness, with each competing model paired to its own agent product. Both kinds of number are useful. They are not the same kind of evidence, and this article keeps them apart.

What Meta actually shipped

The announcement, posted to Meta's AI research blog and dated August 5, covers two products at once.

• Muse Spark 1.2 — a coding-focused update to the Muse Spark family. Meta says it scaled up training compute on coding tasks and widened the diversity of training environments while holding onto the general-agent ability the 1.1 release was sold on. The stated training targets are long-horizon: whole-repository generation, large end-to-end projects, and what Meta calls auto-research, with planning to sequence work, goal conditioning to hold direction across many steps, and context compaction to keep the relevant state alive as a session runs long.

• Muse Code — a terminal coding agent in beta, installed with a one-line shell script from Meta's developer domain. It runs persistent async background agents that survive across a session, on a local event-log runtime Meta describes as replay-exact and restart-safe, and it coordinates multiple subagents per task in isolated worktrees. It ships with three bundled skills: /plan for approval-gated planning, /grill for stress-testing work already done, and /goal for driving a task to completion.

The pairing is not incidental. Meta says the two were co-trained — the model was tuned on rejection-sampled trajectories from the harness, with recipe optimisations for goals, compaction and subagents. That is the same co-training play that produced Claude Code and Codex, and it has a consequence for anyone reading benchmark numbers: the model's headline coding scores were produced inside the harness it was trained against. That is not cheating, it is how agentic evaluation now works. It does mean a 1.2 score obtained through some other scaffold is an open question rather than a safe assumption.

The rest of the spec is unchanged from 1.1, which is itself informative. Context stays at 1,048,576 tokens. Reasoning remains mandatory across five effort levels — minimal, low, medium, high, xhigh — with medium the default; there is no configuration where you get the weights without paying for some deliberation. The weights are closed, with no Hugging Face repository, and Meta's own Model API remains the only first-party place to call it.

43, then 51, then 54

Artificial Analysis scores Muse Spark 1.2 at 54 on its Intelligence Index at the xhigh reasoning setting, ranking it #13 of 185 models against a class median of 32. The index is a weighted composite of nine evaluations — GDPval-AA v2, τ³-Banking, Terminal-Bench v2.1, SciCode, Humanity's Last Exam, GPQA Diamond, CritPt, AA-Omniscience and AA-LCR — split evenly across agents, coding, general capability and scientific reasoning.

Read as a series rather than a snapshot, the Meta trajectory is the actual story. The original Muse Spark scored 43 in April. Muse Spark 1.1 reached 51 on July 9, an eight-point gain in a quarter. Muse Spark 1.2 adds three more in under a month. Meta has moved 11 index points in four months and is now shipping faster than the evaluation ecosystem can keep up with — there is still no published per-benchmark breakdown for 1.2, and no Design Arena record for it at all, while 1.1 has both.

Fifty-four places Meta level with Grok 4.5, the model SpaceXAI shipped a day before Muse Spark 1.1, and behind a tight frontier group: Claude Opus 5 at 61, Claude Fable 5 at 60, GPT-5.6 Sol at 59. The gap to the top is six or seven points. The gap from where Meta started the year is eleven. Whether that closes depends on whether the cadence holds, and Meta is on a four-week release clock at the moment.

A tie on a composite index is not a tie on a job, though, and it is worth saying what the 54 does and does not settle. It settles that Muse Spark 1.2 belongs in the same conversation as Grok 4.5 on aggregate capability. It does not tell you which one writes better patches, because the coding sub-index that would answer that has not been published for 1.2 yet.

Meta's coding numbers, read carefully

Meta's announcement leads on three charts. On its own runs, reported as pass@1 over five attempts with each competing model paired to its own agent product:

• Terminal-Bench 2.1 — 82.9% for Muse Spark 1.2, up from 76.2% for 1.1. Claude Opus 5 is shown ahead at 86.7%.

• DeepSWE v1.1 — 59.3%, up from 53.0%.

• Meta's internal coding benchmark — 70.6%, up from 68.3%.

The deltas are the credible part. A 6.7-point gain on Terminal-Bench and 6.3 on DeepSWE, measured the same way on the same harness by the same team a month apart, is a reasonable read of how much 1.2 improved over 1.1 at the thing Meta was optimising. The cross-vendor placement is the part to hold loosely: harness pairing is a large free variable, and a lab tuning its own harness alongside its own model is not in a position to tune everyone else's equally well. Meta was second on its own slide, which is at least not the usual shape of a vendor benchmark, but no third party has reproduced any of it as of this writing.

The second price list

The most consequential thing in this release is not on the benchmark charts. Muse Spark 1.2 ships with two price lists.

• Standard tier — $1.25 per million input tokens, $0.15 per million on a cache hit, $4.25 per million output. Unchanged from 1.1, to the cent. Rate cap around 3,000 requests per minute. Web-search grounding is billed separately at $2.50 per thousand queries.

• Contributor tier — $0.10 input, $0.002 cached input, $0.20 output. That is 12.5× cheaper on input and 21.25× cheaper on output. The price of admission is permission for Meta to train future models on your traffic, and a rate cap of 60 requests per minute — 2% of the standard budget.

At $0.10 and $0.20, Muse Spark 1.2 would undercut nearly every frontier-adjacent model on the market, including the open-weight budget tier it otherwise loses to on price. That is the interesting number in this launch, and it is not really a pricing decision. Sixty requests per minute rules out most production fan-out patterns on its own, so the tier is aimed at experimentation and small-team work rather than serving. And "we may train on your traffic" is a question for whoever signs off on data governance in your organisation, not for whoever picks models. Treat it as a legal review with a discount attached, not a cheaper SKU.

What the 54 costs to produce

Artificial Analysis publishes what its own evaluation runs cost, and that column is where Muse Spark 1.2's shape shows up. Getting through the nine-benchmark suite cost $637.85 and burned 95 million output tokens, against $548.07 and 94 million for Muse Spark 1.1 — on identical per-token pricing. The extra 16% is pure deliberation.

The latency figures move the same way. Measured at xhigh, time to first token is 26.12 seconds for 1.2 against 2.90 seconds for 1.1, and output speed drops from 213.5 to 165.0 tokens per second. Meta bought three index points with thinking time, and the mandatory-reasoning design means you cannot decline the purchase — only choose how much of it you make.

That 26-second number will be misquoted, so here is the correction in advance: it is a measurement at the most expensive effort level the model exposes, and the API defaults to medium. As a reference point from real traffic rather than a benchmark harness, Muse Spark 1.1 served through OrcaRouter — at whatever effort callers actually request — shows a 7-day p50 time to first token of 1.84 seconds and a p95 of 6.00 seconds, at 519 tokens per second and a zero error rate. Benchmark latency at maximum effort and production latency at default effort are different quantities, and they usually differ by an order of magnitude. Read 26.12s as the ceiling on deep reasoning, not as what a request will feel like.

Where you can call it today

Muse Spark 1.2 is served by Meta's own Model API and, at the time of writing, nowhere else — it was not routed on OpenRouter at launch, there is no Hugging Face repository, and it is not in the OrcaRouter catalog. Muse Spark 1.1 is, at $1.25 and $4.25 — the same numbers Meta charges, because OrcaRouter takes 0% markup and passes provider list price straight through.

That matters more for the comparison than for the model itself. If you are building a cost model for this release, the models you would benchmark it against — Claude Opus 5, Claude Fable 5, GPT-5.6 Sol, Grok 4.5, DeepSeek V4 Flash — sit behind a single key at list price, so the figure in your spreadsheet is the figure on the invoice, and a provider price cut lands the same day rather than after a renewal. For Muse Spark 1.2 specifically, today the answer is Meta's endpoint, a second contract, and a separate bill.

What the announcement did not answer

• No independent coding number. Artificial Analysis publishes a composite for 1.2 but not yet the coding and agentic sub-indices it published for 1.1 (71.3 and 37.5 respectively). Since agentic work was 1.1's weakest dimension by a wide margin, and agentic coding is exactly what 1.2 claims to have fixed, this is the number that would settle the release.

• No reproduction of the harness results. Every coding figure in the announcement is Meta-run. Nobody has yet published Muse Spark 1.2 scores from a neutral scaffold.

• No multimodal verification. The Muse Spark listing advertises text, image, video, audio and PDF input. The independent evaluation covers text and image. Meta's 1.2 messaging has moved almost entirely to software work, and the mixed-media use case 1.1 was explicitly sold for now has neither marketing nor measurement behind it.

• No throughput history. Volume on the endpoint is still too low for the usual rolling latency statistics to populate.

What to watch over the next four weeks

Three things will decide whether this release is what Meta says it is. The first is an independent agentic score for 1.2 — if the sub-index that was 37.5 on 1.1 has moved substantially, the coding pitch is real. The second is whether anyone reproduces the Terminal-Bench result outside Muse Code; a model that only performs inside its own harness is a product, not a capability. The third is what happens to the contributor tier: if the 60-request cap loosens, a frontier-adjacent model at $0.10 in and $0.20 out changes the budget tier's arithmetic in a way that a three-point index gain never would.

And if the cadence holds, Muse Spark 1.3 is due in early September. On this evidence, the number to watch when it lands is not the index score. It is whether Meta can add capability without adding another twenty seconds of thinking to the front of every request.