Muse Spark 1.3 cuts tool calls 20%, Meta stock jumps 4%

A model that needs fewer round trips to finish the same engineering job is the kind of proof Meta's AI skeptics have been waiting for. Muse Spark 1.3, shipped September 2 through Muse Code and the Meta Model API, consumes roughly 20 percent fewer tool calls and 25 percent fewer tokens than Muse Spark 1.2 on comparable coding work — and investors responded by pushing Meta shares up about 4 percent.

The efficiency gains give Meta a path to monetize its AI buildout without raising prices, Bank of America analyst Justin Post said, reiterating a Buy rating and an $810 price target that implies about 32 percent upside from Meta's September 3 close of $610.68.

Muse Spark 1.3 keeps the same 1,048,576-token context window and $1.25-per-million-input-token pricing as Muse Spark 1.2, but Meta engineers found it needs fewer turns and produces cleaner code on long-horizon engineering tasks. The model scores 75.4 on DeepSWE v1.1 and 88.8 on Terminal-Bench 2.1, while trailing Anthropic's Opus on the GDPVal-AA v2 agent benchmark, 1754 to 1824, according to Investing.com's coverage of the release.

Meta shares rose about 4 percent after the announcement. The stock trades near 18 times projected 2027 GAAP earnings, below its historical multiple of roughly 21 times, and BofA's $810 target assumes 24 times 2027 earnings — a bet that efficiency gains from Muse Spark and in-house MTIA chips eventually translate into margin expansion.

Efficiency becomes the new competitive axis

The release marks a deliberate shift in how Meta competes in agentic coding. Rather than chasing top scores on general agent benchmarks, the company is optimizing for cost per completed task — the metric that shows up on enterprise invoices. A 25 percent token reduction at unchanged pricing is functionally a 25 percent price cut for teams that upgrade, and a 20 percent cut in tool calls reduces latency-driven engineering time across thousands of daily agent sessions.

That framing puts Meta in a different lane from Anthropic and OpenAI. Muse Spark 1.3 trails GPT-5.6 on DeepSearchQA, 89.4 to 93.0, and on the Agentic IF Index, 57.8 to 60.5, but leads on narrow coding benchmarks. Meta is not trying to out-generalize Claude or GPT-5.6 in one release; it is competing directly inside the coding-agent category on cost and turn-efficiency.

The model ships through Muse Code, Meta's terminal-based coding agent launched in beta in early August, and through the Meta Model API. It is closed-weight, meaning no self-hosting path — a contrast with Meta's earlier open-source Llama strategy that reflects how the company treats its most commercially sensitive coding models. A contributor tier priced at $0.10 per million input tokens and $0.20 per million output tokens offers an 88 to 95 percent discount in exchange for training rights on prompts and completions.

The capex ROI question

Meta has spent much of 2026 defending its AI infrastructure spending against questions about whether investment is outrunning commercially useful progress. Muse Spark 1.3 gives bulls something concrete: a model that demonstrably lowers the cost of running agentic coding workflows at scale.

Bernstein reiterated an Outperform rating and an $800 target, arguing Meta's AI-enhanced advertising engine remains a major advantage and that the company is on track to rival or surpass Google Search in advertising revenue. KeyBanc remains constructive but more conservative, cutting its target to $760 from $855 while keeping an Overweight rating, citing "meaningful progress" at Meta Superintelligence Labs.

The burden of proof keeps rising. Investors will want evidence that Muse Spark gains adoption, that consumer agents such as Hatch become useful products, and that MTIA custom chips — which BofA estimates could represent 15 to 20 percent of Meta's total AI capacity — actually reduce computing costs. Meta does not need Muse Spark to become a standalone business on OpenAI's scale; better models can improve recommendations, ad targeting, engagement, and developer tools across its existing profit engine.

Meta shares, trading near 18 times projected 2027 earnings against a historical multiple of roughly 21 times, have priced in some skepticism about AI returns. If efficiency gains from Muse Spark 1.3 and subsequent releases translate into measurable margin improvement, the valuation gap BofA identifies could close. If AI costs keep rising faster than revenue, the multiple compression risk remains.

This article is for informational purposes only and does not constitute investment advice.