For two years the frontier-API market had exactly three names you could put on a purchase order: OpenAI, Anthropic, Google. On July 9, Meta made it four.

The Meta Model API opened in public preview, and for the first time Meta is selling one of its in-house foundation models — Muse Spark 1.1 — to outside developers. It's US-only to start, there's no published SLA, and Meta is very clearly entering on price rather than on the top of the leaderboard. That combination is more interesting for founders than a simple "new model dropped," because it changes what your options cost, not just what they can do.

The two facts that matter for your stack#

It's OpenAI-compatible. The API speaks the same protocol as OpenAI's, so an existing client should work with a base-URL and API-key change — no rewrite. That single detail is the difference between "a thing I'll evaluate someday" and "a thing I can A/B against my current provider this afternoon."

# Same client, three lines of config. That's the whole migration to *trial* it.
export OPENAI_BASE_URL="https://api.meta.ai/v1"   # Meta Model API endpoint
export OPENAI_API_KEY="$META_MODEL_API_KEY"
# model: "muse-spark-1.1"

(Confirm the exact base URL and model string against Meta's live docs — this is a public-preview product and those strings move.)

It manages its own context. Muse Spark 1.1 carries a 1M-token window and, per Meta, actively manages it — remembering earlier actions, retrieving from much earlier in a run, and compacting to keep what it needs later. For anyone who has hand-built summarization, retrieval, and compaction to keep a long-horizon agent from drowning in its own history, that's a chunk of context-engineering plumbing the model now claims to own. If it holds up on your workload, that's real leverage.

The honesty check: Meta's own eval report#

Here's what keeps this from being hype: the sober number comes from Meta itself. Its published evaluation report puts Muse Spark behind the frontier leaders on the hardest coding — reported at roughly 61.5 on SWE-Bench Pro against Claude Opus 4.8's ~69.2. Meta is not claiming the crown. It's claiming a price.

A vendor that publishes the benchmark it loses on is telling you how to use it. The message here isn't "we're the best coder." It's "we're cheap and we hold context — point the right work at us."

That's the correct way to read a fourth entrant that undercuts on cost. The reported pricing — $1.25 per 1M input tokens and $4.25 per 1M output, with $20 in free credits — sits under the incumbents' top agentic tiers. Cheap tokens plus self-managed long context is a specific shape of value: it's built for high-volume, long-running agentic work where per-token cost dominates your bill and the task doesn't demand the absolute top of the coding leaderboard.

What it means for you#

The instinct on a new frontier API is to ask "is it better?" That's the wrong question for a founder. The right question is "what does it let me route differently?"

Because integration is nearly free and the model trails on the hardest coding, the move is not to switch — it's to split. Send the cost-sensitive, high-volume, and long-context agentic paths to Meta and measure the bill. Keep your critical-path and hardest reasoning/coding on the incumbent that's already earning its keep. Then re-run your own evals on your own tasks, because a benchmark is a proxy and your workload is the truth.

There's a second-order win even if you never move a single request: a fourth credible vendor pricing aggressively is leverage in every renewal conversation you have with the other three. It's the same dynamic we traced in the demand-side AI price war — and this is the entrant with the deepest pockets and the least need to make money on inference. The AI price war just got another combatant. Watch the incumbents' agentic-tier pricing over the next quarter — this is the kind of entry that moves it.

Just don't put a public-preview, US-only, no-SLA endpoint on your critical path yet. Trial it now for the leverage and the routing; wait for GA before you bet a customer-facing flow on it.