GPT-6.1 Terra vs Muse Wonder: Four Index Points, Eight Times the Bill

GPT-6.1 Terra and Muse Wonder are four points apart on the public scoreboard that has measured both, and about eight times apart on what it costs to finish a task on that scoreboard. OpenAI released Terra on October 26, 2026 at $1 per million input tokens and $5 per million output; it scores 48 on the Artificial Analysis Intelligence Index and costs $0.40 per index task. Meta released Muse Wonder on October 12 at $4 and $20; it scores 52 and costs $3.27 per task. Both list a one-million-token context window. A leaderboard sort would end the decision in one line, which is why the leaderboard is the wrong place to end it: the model with the higher score and the higher bill is also the faster one, and the cheap model's advantage depends entirely on what a wrong answer costs you.

The figures below come from Artificial Analysis, read on October 27, 2026, and from the two vendors' launch documentation. Vendor claims are marked as claims. The one task-level benchmark in this piece was run by DualEntry, a company that sells accounting automation, so it is third-party evidence with a commercial interest, not a neutral audit; it is treated that way.

Two models built for different jobs

Terra is OpenAI's volume model, the replacement for GPT-5.6 Terra and the cheapest way to get the current OpenAI family's general competence. It accepts text and images, returns text, documents a 1,050,000-token context with a 128,000-token output ceiling, and exposes reasoning effort from none through max. OpenAI also shipped it to Amazon Bedrock on launch day and into ChatGPT Work and Codex, though not into the consumer ChatGPT chat box. It is API-only in the sense that matters for procurement: no fine-tuning, no Realtime, no Assistants.

Muse Wonder is Meta's first model that belongs in the frontier conversation, and it arrives with a split that most comparison pages blur. The API model is Wonder, at $4 input and $20 output. The higher-scoring checkpoint, Muse Wonder Full at 56 on the same index, is sold only inside the Muse Code subscription at $50 or $200 a month and is not served through the API. The headline benchmark rows often describe Full. The model a developer can actually call is four points above Terra, not eight.

• Price — GPT-6.1 Terra $1.00 input / $5.00 output per million; Muse Wonder $4.00 / $20.00.

• Cached input — Terra $0.05 at a 95% discount; Wonder $0.20 at a 95% discount.

• Context window — Terra 1,050,000 documented; Wonder 1,000,000 documented.

• Maximum output — Terra 128,000 documented; Meta has not published a ceiling for Wonder, and third-party trackers report figures near 943,000.

• Independent Intelligence Index — Terra 48, 25th of 221; Wonder 52, 10th of 221.

• Cost per index task — Terra $0.40; Wonder $3.27.

• Output speed — Terra 90.9 tokens per second; Wonder 207.8.

• Time to first token — Terra 77.96 seconds; Wonder 31.62 seconds.

• Inputs — Terra text and images; Wonder text, images, video and audio according to trackers, with Meta's own launch material emphasizing text and images.

• Weights — closed on both sides. Meta's open-weight option is Muse Spark, a different model.

The bill is eight times the index gap

The cost-per-task figure is the one that should change how the four-point gap reads. Artificial Analysis does not just multiply list rates by a fixed token count; it measures how many tokens each model spends completing the same set of index tasks, then prices that consumption. Terra averages $0.40 per task and Wonder $3.27, an 8.2x difference on a capability gap of less than 8%.

The list rates alone do not explain it. Take a workload of 100 million input tokens and 10 million output tokens: Terra bills $150 and Wonder $600, a 4x difference. The per-task gap is twice as wide, which means the tokens the index evaluations actually consumed did not split 10-to-1; Artificial Analysis prices the measured input and output volume for each model, and that per-task figure reflects the evaluations rather than a flat example. A per-token price is a fuel price, not a journey cost.

That distinction matters most at the volume where Terra is aimed. A pipeline that runs 100,000 tasks a month spends $40,000 on Terra and $327,000 on Wonder at Artificial Analysis's measured rates. The capability difference between them does not change by a factor of eight when the workload does; the bill does. If your task is classification, extraction, form reading or a bounded agent step, the burden of proof sits with the expensive model.

One caution on the arithmetic: the index is periodically renormalized, and cost-per-task figures move with it. Compare models from the same reading, as these figures are, rather than subtracting a number from an older article.

Wonder is faster, and Terra's speed claim does not survive the comparison

OpenAI's launch material describes Terra as substantially faster than GPT-6 Astra and GPT-6.1 Sol. Within the OpenAI family that appears to be true, and it is the reason the model exists: a cheap, quick generalist for visual and document work. Against Meta's model it is not close. Wonder produces 207.8 output tokens per second to Terra's 90.9, and it returns its first token in 31.62 seconds against Terra's 77.96.

Both are slow starters in absolute terms, which is a property of large reasoning models rather than a scandal. But the ordering is the opposite of the price ordering, and that has a practical consequence. For a batch job that runs overnight, Terra's latency is irrelevant and its cost is decisive. For an interactive product where a user waits for the first token, a 78-second start is a product decision, not a benchmark line, and Wonder's 32 seconds is the better experience even at five times the input rate.

The speed figures also undercut a common assumption about cheap models: that they are the fast option by default. Terra is cheap per token and cheap per task. It is not fast per second. Those are different claims, and only one of them is true.

A task-level test where the cheap model lost

The broadest third-party check available at launch is DualEntry's accounting benchmark: 101 synthetic cases across transaction classification, journal entries, payables, receivables, reconciliation, reporting, month-end close and accounting knowledge, run three times per model with fixed reference answers rather than an AI grader. Twenty-two models were tested. DualEntry sells accounting automation, so the benchmark is not neutral, but the method is more concrete than a composite index.

Terra finished at 78.2% overall, tied for 18th of 22. Its profile was uneven in a way the composite hides. Transaction classification was perfect, 13 of 13. Accounts payable was 53.8%, tied for last. Bank reconciliation was 83.3%, behind both Kimi K3.1 Pro and Muse Wonder, though the release does not publish Wonder's exact reconciliation score. Terra was the fastest OpenAI model on the benchmark and third fastest overall at 16.1 seconds per task, at $0.042 per task — roughly half of GPT-6.1 Sol and a tenth of GPT-6 Astra.

A separate evaluation platform, Vals AI, places Terra at 60.17% and 10th of 28 on its own suite, with coding and terminal work as strengths and legal, tax and finance as weaknesses. Different suite, different scale, so the number is not comparable to the Artificial Analysis index; the shape is. Terra is a fast, inexpensive generalist with a wide variance across task families, not a uniformly good one.

That variance is the argument against buying on cost per task alone. In a workflow where a 53.8% sub-score means a human reworks half the payables queue, the $2.87 saved per task is not a saving; it is a transfer of cost from the token budget to the payroll budget. Muse Wonder's four-point index advantage does not prove it would clear that bar — no shared benchmark in public hands establishes it — but on the one task-level test both appear in, it is ahead where Terra is weakest.

Modalities, output length, and what Meta has not documented

Wonder's input row is the clearest structural difference. Trackers list text, images, video and audio in, text out; Meta's launch material is narrower in places, but video and audio input are the reason a team would consider it over Terra at all. Terra takes text and images. If your pipeline ingests recorded calls, screen recordings or product video, Terra cannot be the only model in it.

Output length is the second. Terra documents 128,000 output tokens. Meta has not published Wonder's ceiling, and the ~943,000 figure circulating comes from third-party observation rather than a specification. Treat it as a lead to verify, not a number to design around. The same caution applies to knowledge cutoffs: Terra documents June 1, 2026; Meta has not published one for Wonder.

Neither model is on our catalogue. Here is what is

One honest constraint before any product note: as of October 27, neither GPT-6.1 Terra nor Muse Wonder is in OrcaRouter's 214-model catalogue. Terra is available through OpenAI's own API and, since launch day, through Amazon Bedrock. Wonder is sold on Meta's Model API. Anyone claiming to route either of them through us today is describing a roadmap, not a catalogue.

What the catalogue does carry is the nearest hosted arm on each side: GPT-6.1 Sol at $2 input and $10 output, which scores 52 on the index and costs $0.72 per task, and Muse Spark 1.2, Meta's Spark line. If the Terra-shaped workload is the one you need to measure — high-volume text and image work where cost per completed task decides the purchase — Sol is the hosted proxy with more capability and roughly twice the per-task cost, and the comparison tells you whether Terra's saving is worth the migration. Both are served at provider list price with no markup, on one key with automatic failover, so a provider outage does not become a product outage.

When either model lands in the catalogue, the switch is a model string rather than a second contract. That is the practical reason to benchmark the workload now instead of waiting for the launch post that settles the matchup.

Who should pick which

Pick Terra when the work is high-volume, well-specified and dominated by text or images: document extraction, classification, form and screenshot reading, drafting, bounded agent steps. The $0.40 per task is the lowest in this comparison by a wide margin, the 128K output is documented rather than rumored, and the uneven subtask scores are acceptable when a human reviews the output anyway.

Pick Muse Wonder when the input is richer than text and images, when an agent has to return a long artifact in one call, or when a wrong answer costs far more than $2.87. The 207.8 tokens per second and the 32-second first token make it the better interactive model, and the four-point index advantage is real even if it is smaller than the gap between Muse Wonder Full and Terra. Budget for the $20 output rate and confirm which checkpoint the benchmark you are reading actually measured.

The four points are the least interesting number in the matchup. The bill, the time to first token and the task-level variance are what a production decision turns on — and on those, the cheaper model is not simply the lesser one.