This article is part of the HOGDigest editorial series. → Explore HOGDigest
Muse Spark 1.2 carries a 1,048,576-token context window with a maximum output of 131,072 tokens — the same numbers as its predecessor had, because nothing about the window changed in this release. What changed is how the model behaves inside it, which is the more interesting story and the one we dug into when comparing it with a very different long-context model in our piece on vs Gemini 3.5 Flash-Lite on long context.
A million tokens has stopped being a headline feature — several models offer one now, and the useful question has moved on to what you can actually do with it. For a coding model the answer is specific: a million tokens is roughly the point at which you stop building a retrieval pipeline and start just handing the model the repository.
What a million tokens actually holds
Rough working conversions, useful for sizing rather than billing:
• ~750,000 words of English prose, or about 1,500 pages.
• A mid-size codebase. A 200,000-line service with its tests, configs and dependency manifests generally lands somewhere between 400,000 and 700,000 tokens depending on language and comment density.
• A codebase plus its history. Source, plus the issue thread, plus the design doc, plus the last six months of relevant commits, still fits with room left over.
• A long multimodal session. Screenshots, PDFs of specifications, video walkthroughs and text all share the same budget.
The 131,072-token output ceiling matters as much as the input side for coding work. That is enough to emit a substantial multi-file change in one response rather than stitching together a dozen partial diffs — which, in practice, is where a lot of agent workflows quietly fall apart.
The three jobs it genuinely changes
Whole-repository reasoning. The classic failure mode of a retrieval-augmented coding assistant is that it answers confidently from three files when the answer lived in a fourth it never retrieved. Give the model the whole repository and that specific failure disappears. Meta's own positioning for this release leans on exactly this — multi-file refactors, extended debugging sessions, whole-repository generation.
Long-horizon agent runs. Meta's demonstration case had the model optimising GPU kernels across more than 1,000 tool calls over 24 hours. A run like that accumulates enormous state: tool outputs, failed attempts, intermediate reasoning. A big window is what lets an agent remember why it already ruled something out instead of rediscovering it four hours later.
Document-heavy professional work. This one is under-discussed, and the independent evidence is striking. On Vals AI's common-harness evaluations, the model ranks #1 of 44 on Finance Agent (v2), #1 of 136 on TaxEval v2, and #1 of 31 on Harvey's Legal Agent Benchmark. Those are all tasks where the input is a large pile of documents and the work is synthesis across all of it. The context window is doing real work there.
Where a big window stops helping
Three honest caveats, none of which are specific to this model but all of which apply to it.
Advertised length is not measured recall. A context window is a capacity claim, not a quality guarantee. Every long-context model degrades somewhere between "holds it" and "reasons over all of it", and the position of that line is a property of the model, not the number on the spec sheet. Meta has published no needle-in-a-haystack or long-context retrieval results for version 1.2 specifically, and no third party has either. Treat the million as an upper bound you should verify at your own working length.
Long context is expensive context. At $1.25 per million input tokens, filling the window once costs about $1.31 before the model has produced a single output token. Do that on every request in a loop and the arithmetic gets ugly fast. The mitigation is real, though: cached input runs at $0.15 per million — an 88% saving — so a workload that keeps the large, stable portion of the prompt identical across calls pays roughly a seventh as much. Structuring prompts so the repository comes first and the varying instruction comes last is the single highest-leverage optimisation available here.
Long context is slow context. Muse Spark 1.2 already deliberates heavily — Artificial Analysis measured 95 million output tokens to complete its Intelligence Index against a ~70 million median, and publishes no time-to-first-token figure for it at all. OrcaRouter's own seven-day production telemetry shows a p50 first token of 7.73 seconds and p95 of 10.00 seconds, against 1.93 seconds p50 for version 1.1. Add several hundred thousand tokens of prompt to that and you are firmly in background-job territory.
Practical patterns that work
• Put the invariant first. Repository, schema, style guide, house conventions — all at the top, byte-identical between calls, so the cache can do its job. The variable instruction goes last.
• Don't fill it because you can. More context is not free accuracy. Include what the task plausibly needs; a focused 200,000-token prompt often beats an indiscriminate 900,000-token one on both cost and quality.
• Measure recall at your working length. Before committing an architecture to full-repository prompting, run a handful of questions whose answers you've deliberately buried at different depths. Ten minutes of testing beats a spec-sheet assumption.
• Reserve headroom for output. With a 131,072-token output ceiling, a prompt that consumes 1,020,000 tokens leaves the model no room to answer at length. Budget both sides.
• Keep a fallback for the long tail. Very long prompts are where timeouts and provider hiccups concentrate. Running the model through a router with automatic failover — OrcaRouter passes Meta's list price through at 0% markup and carries alternatives behind the same key — means a stalled long-context request degrades to a retry rather than an incident.
The takeaway
The context window is the least changed thing about this release and one of the most useful. A million tokens in and 131,072 out is enough to hand a model an entire mid-size repository, or a quarter's worth of filings, and ask a question that spans all of it — and the independent domain results suggest that's where this model is genuinely strong. Just don't confuse capacity with capability: verify recall at your own working length, keep the stable part of your prompt cacheable, and remember that every token you add is one the model has to read before it starts answering seven seconds later.
Sourcing note: context window, output limit and pricing are Meta's published figures. Token consumption is from Artificial Analysis; domain rankings from Vals AI; first-token latency is OrcaRouter's own seven-day production telemetry. No long-context retrieval benchmark has been published for Muse Spark 1.2 by Meta or any third party as of August 7, 2026.
