Meta

Meta has positioned itself as the champion of open-source AI. Their Llama 4 family includes Scout (with a record 10 million token context window) and Maverick (which competes with proprietary mid-tier models on benchmarks), all free to download and self-host under the Llama Community License. On July 9, 2026 Meta shifted strategy with Muse Spark 1.1, its first paid API model, opened to US developers at $1.25 input / $4.25 output per 1M tokens, roughly a quarter of OpenAI and Anthropic list rates. On August 5, 2026 it followed with Muse Code, a terminal coding agent for macOS and Linux, and Muse Spark 1.2, a coding-focused update co-trained inside the Muse Code harness. Standard 1.2 pricing holds at $1.25/$4.25, but the new muse-spark-1.2-contributor tier serves the same weights at $0.10/$0.20 in exchange for training rights on your data, currently the cheapest frontier-adjacent coding path on the market. Muse Spark 1.3 followed on September 2, 2026 with every rate held from 1.2, Standard at $1.25/$4.25 with $0.15 cached input and Contributor at $0.10/$0.20, so the pitch moved from price to consumption: roughly 20 percent fewer tool calls and 25 percent fewer tokens on the same agentic work, a discount that reaches the invoice without touching the price list. Meta reports DeepSWE 1.1 at 75.4, SWEAtlas CodeBase QnA at 59.4, Terminal-Bench tied with GPT-5.6 Sol at 88.8, and MRCR at 98.5 and 98.1 across the long-context bands, with the caveat that the 1.3 column runs at max reasoning against 1.2 at xhigh. Worth stating plainly given the company's reputation: Muse Spark 1.3 is not open weights, and Meta has promised a Muse Spark open-weight release without naming a date, a variant, or a license. Meta remains the most important player for teams that need on-premise deployments, fine-tuning, or zero marginal inference cost, and now also competes hard on low-cost hosted agentic inference.