Muse Spark 1.2 Ships With Muse Code
Meta ships Muse Spark 1.2 with Muse Code at the same $1.25/$4.25 rates. helloai bumps 1.1 to 1.2 at 1498 Elo — still the cheapest agentic pick.
Meta just upgraded its cheapest frontier API without raising the price. Muse Spark 1.2 landed on the Meta Model API and OpenRouter around August 5–6 at the same $1.25 per million input tokens and $4.25 per million output tokens that made 1.1 the budget agentic pick — still with a full 1M-token context window. On Arena's text overall board dated August 6, muse-spark-1.2 (xHigh) sits at 1498 Elo, second only to Claude Fable 5 at 1507 in helloai's tracked set.
The model did not arrive alone. Meta AI chief Alexandr Wang said Spark 1.2 was developed and trained alongside Muse Code, Meta's first coding-agent product, now in preview with pay-as-you-go pricing described as similar to the Spark API. That pairing matters more than a version digit: Meta is selling a harness and a model as one stack, the same play Anthropic and OpenAI have been running with Claude Code/Cowork and Codex-style surfaces. For teams already on the Meta Model API, the path is a same-id upgrade, not a vendor switch.
helloai applies the same rule it uses for every same-provider bump: replace 1.1 with 1.2 under the existing muse id. Pricing and context do not change; the card copy and arena aliases do. Muse remains the cheapest agentic option we track, undercutting Opus 5 ($5/$25), Kimi K3 ($3/$15), and Grok 4.5 ($2/$6) on both sides of the meter for teams that want multi-agent orchestration without Mythos-class bills.
Trade-offs stay real. Arena votes for 1.2 are still early compared with Fable and the long-lived Claude Opus lines, and Meta's contributor tier — roughly $0.10/$0.20 when you opt in to training use of prompts and completions — is a different product from the standard no-train production rate. Standard tier is the default for anyone treating this as a production dependency. Coding-agent previews also move faster than public Elo; treat Muse Code claims as vendor-directed until your own harness evals land.
The wider board did not freeze while Meta shipped. Qwen3.8-Max is pressing the $2/$6 tier with high preliminary arena scores, and Gemini 3.5 Pro is still missing from Google's pricing page. For /api/recommend today, Meta's slot is 1.2: same budget, slightly higher human preference, and a coding-agent surface that did not exist when 1.1 replaced Qwen in July. The next decision is whether that stack holds once Alibaba's Max-class model and Google's delayed Pro both finish their launches.