Meta's Muse Spark 1.1 shows how fast the catch-up game is now

Meta released Muse Spark 1.1, a coding agent model that scores 77% on SWE-bench Verified and 61% on the full test. That puts it within striking distance of the current frontier — competitive, not leading, but solid enough to matter. Pricing is $1.25 per million input tokens and $4.25 per million output for the public preview. That’s cheap compared to premium models like Claude Opus or GPT-5.5, though not the absurd fractions of a cent I saw some numbers claiming.

The interesting part isn’t just the benchmarks. It’s the timeline. Meta wasn’t even in this race six months ago. They shipped a model that’s now benchmarking against the best from a standing start. If a well-resourced lab can catch up that fast, Chinese labs like DeepSeek or Qwen — who already have infrastructure and talent — will likely do the same. Probably faster.

I remember when catching up to the frontier meant a 12-month R&D cycle. Now it’s a quarter, maybe less. The moat isn’t some unique architecture or secret training dataset. It’s raw compute access and a willingness to burn cash at negative margin until the others fold. That’s a brutal dynamic if you’re running a wrapper startup on someone else’s API.

What I’ll actually test: I have a set of real-world coding tasks — refactoring a Magento module, writing a trading bot agent with tool calls, debugging a cron failure from logs. I’ll run those through the API with the exact same prompts I use for Claude. If it handles multi-step changes without hallucinating function signatures, and that pricing holds, it becomes my default for non-critical agentic loops immediately. The benchmark gap to the top is small enough now that cost plus capability make it hard to ignore.