Meta started charging developers for its own model on July 9, and the number everyone latched onto was $1.25 in, $4.25 out per million tokens. Zuckerberg’s framing: about a quarter of what Anthropic and OpenAI charge. Against Claude Opus 4.8 at $5/$25 and GPT-5.5 at $5/$30, that’s a 4x to 7x gap on the output side, which is the side that actually hurts.
So I did the thing you’re supposed to do before moving a production coding agent onto a preview API: I read Meta’s own evaluation report instead of the launch coverage. The pricing is real. The 4x is not. And the category Meta is loudest about — coding — is the one where Muse Spark 1.1 looks worst.
Here’s the part that should give you pause. Meta pitched this at the coding-agent market. Meta did not publish a SWE-bench Verified score.
What actually shipped
The Meta Model API went into public US preview on July 9, 2026. It’s the first time Meta has charged for access to its own model rather than dumping weights and letting you self-host. Muse Spark 1.1 is the launch model: a reasoning model with a 1M-token context window, parallel tool calling, subagent delegation, and multimodal input covering images, video, and PDFs.
The endpoint speaks both OpenAI-compatible and Anthropic-compatible request formats. In practice that means swapping a base_url and an API key, not rewriting your tool schemas. New accounts get $20 in free credits. That $20 is enough to answer the only question that matters, which is whether the model finishes your tasks — spend it before you plan a migration.
The full rate card, which is more interesting than the headline:
| Input | Cached input | Output | |
|---|---|---|---|
| Muse Spark 1.1 | $1.25 | $0.15 | $4.25 |
| Claude Opus 4.8 | $5.00 | $0.50 | $25.00 |
| Claude Sonnet 4.6 | $3.00 | $0.30 | $15.00 |
| GPT-5.5 | $5.00 | $0.50 | $30.00 |
| Grok 4.5 | $2.00 | $0.50 | $6.00 |
Prices per million tokens. Sonnet 4.6 is running introductory pricing of $2/$10 through August 31, 2026, which narrows the gap more than the table suggests — worth remembering if your comparison is Sonnet rather than Opus.
Two line items don’t appear on that table and both will show up on your bill. Web search grounding runs $2.50 per 1,000 queries on top of the tokens for the request. And rate limits are per team, not per key: 60 RPM and 2M TPM on free, 3,000 RPM and 4M TPM on paid. If you’re running a fan-out agent architecture with a dozen subagents hammering one team quota, 3,000 RPM is a ceiling you will find.
The reasoning-token problem
Muse Spark 1.1 is a reasoning model, and its internal chain-of-thought tokens bill at the output rate. Not a discount rate. The full $4.25.
This is where the headline multiplier starts eroding, because output tokens are already the expensive half and reasoning models are exactly the models that produce a lot of them. A model that costs 6x less per output token but emits 3x more output tokens is only 2x cheaper, and you won’t see that in the rate card. You’ll see it at the end of the month.
Let me put actual numbers on a single agentic coding task. Assume 200K input tokens, of which 180K hit the cache (a realistic ratio once your system prompt and repo context stabilize), plus 30K output tokens:
- Muse Spark: 20K × $1.25 + 180K × $0.15 + 30K × $4.25 = about $0.18
- Opus 4.8: 20K × $5.00 + 180K × $0.50 + 30K × $25.00 = about $0.94
That’s the 5x everyone is quoting. Now let the reasoning model reason. If Muse Spark emits 3x the output tokens on the same task — 90K instead of 30K, which is not an aggressive assumption for a chain-of-thought model doing multi-step tool work — its cost goes to roughly $0.43. The gap collapses from 5x to about 2.2x.
Then account for the tasks it doesn’t finish. On SWE-Bench Pro, Meta’s own report puts Muse Spark at 61.5 against Opus 4.8 at 69.2. Expected attempts to land one completed task: 1.63 versus 1.45. Cost per completed task lands around $0.70 for Muse Spark and $1.36 for Opus.
Still cheaper. Roughly half, not a quarter. That’s a real saving and I’d take it in a lot of contexts — but it is a different decision than the one the press release invites you to make, and it’s the difference between “obviously switch” and “depends what you’re doing.”
The cached-input rate is the genuinely strong part of this pricing and nobody covered it. At $0.15/M, Meta’s cache read is an 88% discount off its own input rate, and it’s 3.3x cheaper than Anthropic’s cache read in absolute terms. If your workload is a long stable prefix hit over and over — a big system prompt, a fixed codebase context, a document you’re answering many questions against — the cache rate is where your money actually goes, and Muse Spark wins that lane more convincingly than it wins the headline.
Where Muse Spark 1.1 is legitimately good
Read Meta’s eval report by category and a clear shape emerges. This is an agentic tool-use model that happens to code, not a coding model.
On tool use and agentic orchestration it beats Opus 4.8 outright: MCP Atlas 88.1 vs 82.2, JobBench 54.7 vs 48.4, Finance Agent v2 57.2 vs 53.9, HealthBench Professional 59.3 vs 55.8, and Humanity’s Last Exam with tools 62.1 vs 57.9. Those aren’t rounding errors. If your workload is “call twelve tools in the right order, recover when one fails, keep the thread straight” — MCP-heavy agents, research pipelines, workflow automation — this model is competitive with frontier pricing at a third of the cost.
That’s the honest pitch, and it’s a good one. It’s just not the pitch Meta made.
Where the story falls apart
Coding. The thing Meta aimed this at.
Terminal-Bench 2.1: 80.0, against GPT-5.5 at 83.4. SWE-Bench Pro: 61.5, against Opus 4.8 at 69.2. DeepSWE 1.1: 53.3, against GPT-5.5 at 67.0 — that one is a rout, a 13.7-point gap on an agentic software engineering benchmark.
And the absence: no SWE-bench Verified number. It’s the most-cited coding benchmark in the industry, Opus 4.8 posts 88.6 and GPT-5.5 posts 88.7, and Meta launched a model explicitly positioned against them in coding without publishing a comparable figure. Companies publish the benchmarks they win. Draw your own conclusion, but don’t assume the missing number is a good one.
The 1M context has a similar problem. On MRCR long-context retrieval it scores 54.1 against GPT-5.5’s 74.0. A 1M window that retrieves poorly is a 1M window you can’t trust to actually use 1M tokens — and if you’re paying $1.25/M to stuff a giant context in, retrieval quality is the whole point. Long context is a capacity claim. Retrieval is the capability, and these are 20 points apart.
Computer use is close but behind: OSWorld-Verified 80.8 vs 83.4 for Opus. Multimodal, same pattern — CharXiv 88.4 vs 89.9, BabyVision 76.3 vs 83.6. Respectable. Not leading.
One more caveat on all of the above: these are Meta’s own numbers, from Meta’s own harness, and independent verification hasn’t landed yet. Vendor-run evals are marketing artifacts with error bars. The fact that Meta’s own numbers show coding as the weak spot is what makes them credible here — nobody talks themselves down by accident.
The migration is easy, which is the trap
Changing a base URL is a two-line diff. That’s the seduction. The work isn’t the swap, it’s everything you find out afterward.
Prompt caching semantics differ. Anthropic wants explicit cache_control breakpoints; OpenAI-style caching is automatic prefix matching. Your carefully-placed breakpoints don’t carry over, and if you assumed they did, your cache hit rate — and therefore the $0.15 rate that makes this whole thing worth it — quietly falls off a cliff.
Tool-call behavior differs even when the schema validates. Meta’s docs themselves tell you to test for invalid tool calls, repeated calls, unnecessary tool use, and recovery from tool failures. That’s an unusually candid list to find in vendor documentation, and it’s a good test plan. Run it against your actual tools before you trust it with anything that writes.
Then the preview problems. US-only. Rate limits per team. Public preview means the model can change under you, pricing can change under you, and there’s no SLA to point at when it does. If you’re routing production traffic through a preview endpoint, that’s a decision, not an oversight — make it deliberately.
Analysts are already pricing in the endgame. Amit Jena at Kanerika expects the classic playbook: aggressive entry pricing, then a 30–50% increase in 18–24 months once share solidifies. Pareekh Jain of Pareekh Consulting put the adoption logic more usefully — cost becomes the differentiator only after a model is judged good enough. Or as Muskan Bandta of ZopDev framed it, price is why people show up, capability is why they stay.
Which means the smartest use of Muse Spark for a lot of teams isn’t migrating at all. It’s walking into your Anthropic or OpenAI renewal with a credible alternative and asking for a volume discount. That’s free money and it costs you a base_url in a test harness.
What I’d actually do
If you’re running MCP-heavy agents, research pipelines, or tool-orchestration workloads with a big stable prompt prefix — pilot it this week. That’s where the benchmarks and the cache rate both point, and the $20 in credits covers a real evaluation. Route it behind a gateway so you can fall back in one config change, and put observability on it so you can see the token bloat rather than infer it from the invoice.
If you’re running production coding agents — long refactors, anything where a wrong edit is expensive — don’t move yet. The coding benchmarks are Meta’s weakest published category, the headline benchmark is missing entirely, and a cheap model that fails a refactor halfway through costs you far more than the token delta saved. If cheap coding is the goal, Grok 4.5 at $2/$6 is the better-evidenced bet right now.
And if you’re anywhere in between, dual-route by task rather than picking a winner. Tool orchestration and high-volume cached workloads to Muse Spark, hard code to Opus or GPT-5.5. That’s more plumbing than a single-provider setup, but it’s also the only configuration where you capture the discount without eating the failure rate.
Before you touch any of this: measure your current output-token-to-input-token ratio. If output is a small fraction of your spend, the 6x output discount you’re excited about is worth almost nothing to you, and you’d have found that out in ten minutes instead of a sprint.