Meta Muse Spark 1.1 Is a Budget Agentic Coding Model
Meta's first paid model, Muse Spark 1.1, launched July 9 at \$1.25 / \$4.25 per million tokens with a 1M-token context and a computer-use mode.
Meta Enters the Coding-Agent Race With Aggressive Pricing
Meta has spent years giving its best models away for free, so the most surprising thing about Muse Spark 1.1 is the price tag. Released July 9, 2026, it is Meta Superintelligence Labs' first paid model — and the pricing is set to undercut, not match, the competition. For developers building on top of the latest AI coding tools, it lands as a genuinely cheap way to run long, autonomous agent sessions.
- Launched July 9, 2026 as Meta's strongest model yet for agentic and coding work
- API pricing: \$1.25 per million input tokens and \$4.25 per million output tokens, with \$20 in free credits for new accounts
- 1 million-token context window for large repository edits and multi-hour agent runs
- Ships with a computer-use mode that drives a real desktop from a plain-language goal
How Cheap Is Muse Spark 1.1, Really?
Pricing is the headline. At \$1.25 in and \$4.25 out per million tokens, Muse Spark 1.1 sits above entry-level tiers like GPT-5 mini and Claude Haiku 4.5, yet comfortably below mid-range models such as Claude Sonnet 4.6 — and far beneath flagships like Claude Opus 4.8 at \$5 / \$25. For teams running agents that read and rewrite whole codebases, output tokens dominate the bill, so that \$4.25 output rate is where the savings compound.
The \$20 in free credits for every new API account is a quiet but clever touch. It lowers the barrier for a solo developer to actually try repository-scale automation before committing a card, which is exactly how good habits — and platform loyalty — form.
What Can It Do Beyond Writing Code?
Muse Spark 1.1 is a reasoning model tuned for end-to-end agents. It handles repository-level coding, native multimodal perception, and a computer-use mode that steers a real desktop from a natural-language instruction. The 1-million-token context window is the practical enabler here: it is what makes long agentic sessions and sprawling repository edits feasible rather than a memory-management headache.
That puts it squarely in the same conversation as the wave of agentic coders we have covered recently, from Poolside's open-weight Laguna XS 2.1 to Grok 4.5's budget coding pitch. The field is crowded — and that is great news for the people paying the bills.
Why Competition at the Low End Is Good News
It is easy to fixate on which lab holds the top benchmark, but the more meaningful race in mid-2026 is happening at the affordable tier. When a company with Meta's reach decides to compete on price for agentic work, everyone building automation benefits: cheaper experiments, lower production costs, and more room for small teams to punch above their weight.
Muse Spark 1.1 is currently a preview limited to US developers, so global availability will take time. But the signal is clear and encouraging — capable, autonomous coding help is getting steadily cheaper, and the tools that used to require a flagship budget are sliding within reach of anyone with a good idea and a weekend.
Sources: TechCrunch — July 9, 2026; DataCamp — July 9, 2026.
More AI Stories
ChatGPT Work and GPT-5.6 Turn AI Agents Into Coworkers
OpenAI's July 9 launch pairs GPT-5.6 — split into Sol, Terra, and Luna tiers — with ChatGPT Work, an agent built to finish whole jobs across your apps.
Grok 4.5 Arrives as an Opus-Class Coding Model at Budget Pricing
xAI's Grok 4.5 lands as an Opus-class coding and agentic model — trained alongside Cursor, priced low, and available today across Grok, Cursor, and the console.
Liquid AI's Antidoom Open-Sources a Fix for Reasoning-Model Doom Loops
Liquid AI open-sourced Antidoom, a training method that stops reasoning models from looping on repeated tokens — cutting doom loops from over 20% to near 1%.