FULL STORY

Meta's Muse Spark 1.3: Launch, Benchmarks and Max Upgrade

Meta launched Muse Spark 1.3 on Sept 3, touting Fable 5-level performance at a fraction of the price. Third-party benchmarks largely confirmed the claims, followed by a stronger 1.3 Max release days later.

2026-09-03 ~ 2026-09-06 · 6 episodes · 76 posts

Episode 1 · Meta Launches Muse Spark 1.3: Near-Frontier Performance at ~1/8 the Price (2026-09-03, 53 posts)

Meta officially released Muse Spark 1.3 on 09-03, calling it its biggest leap yet in coding and agentic tasks, with "frontier performance at prices almost too cheap to meter." The model is live on Muse Code and the Meta Model API, with a max reasoning tier pending safety testing; Zuckerberg confirmed the launch and teased an upcoming model plus an open-weights version of Muse Spark. This marks Meta's fourth iteration in five months and the third consecutive upgrade from 1.1 to 1.3 in two months.

Confirmed

  • Pricing is $0.10 per million input tokens and $0.20 per million output tokens — roughly 1/8 of Claude Fable 5's price (per alexandrwang)
  • On the Artificial Analysis Intelligence Index, the xhigh version scores 61-62, up from 57 (1.2) and 53 (1.1), tying Claude Fable 5 and beating GPT-5.6 Sol, Grok 4.6, Kimi K3 and Gemini (per m3, m5); Reddit user DistanceSolar1449 also shared a screenshot showing it surpassing GPT-5.6 Sol
  • On Artificial Analysis's Coding Agent Index, the xhigh version sits on the Pareto frontier at $1.72 per task; the max tier can't be scored as Meta hasn't disclosed its pricing
  • Official deltas vs 1.2: 20% fewer tool calls, 25% lower token consumption, stronger long-horizon and visual coding; supports longer task chains across workflows in a single thread, asks clarifying questions, flags when stuck, and confirms before critical actions
  • Third-party benchmarks (relayed by armandruiz) show coding above GPT-5.6 and Opus 5; ChrisGPT's comparison with the same-day Gemini 3.8 Flash claims long-context scores above GPT-5.6
  • Muse Code offers one-line install: curl -fsSL <url> | bash

Unconfirmed

  • Benchmark data in m10 is sourced from Google DeepMind and not independently verified; details of the leaderboard screenshot in m2 are not public

Why it matters

  • Meta is delivering near-Claude-Fable-5 intelligence at 1/8 the price and holding a spot on the coding-agent cost-performance Pareto frontier, directly pressuring frontier-model pricing
  • The four-releases-in-five-months cadence shows Meta's momentum, and the teased open-weights version could extend its impact

33 more related posts →

Episode 2 · Meta Releases Muse Spark 1.3 as Its Strongest Model (2026-09-03, 2 posts)

Meta has released Muse Spark 1.3, its strongest model yet, with Alexandr Wang claiming it beats GPT-5.6 Sol on coding. Polymarket traders remain skeptical, giving Meta only a 12% chance of holding the top AI model by year-end.

Episode 3 · Meta's Muse Spark 1.3 Nears Fable 5.1 in Community Benchmarks at Fraction of Cost (2026-09-03, 5 posts)

Meta's new Muse Spark 1.3 has impressed across several third-party tests, nearly matching Fable 5.1 while running faster and at far lower cost; Scale AI CEO Alexandr Wang repeatedly amplified these results with high praise. All findings so far come from community-built benchmarks rather than official data, but the model is drawing wide attention.

Confirmed

  • User SPAC89 tested Muse Spark 1.3 Ultra against Fable 5.1 with identical self-improvement prompts (retry if an independent judge scores below 9.5/10): over 2 hours each, Muse Spark ran 20 self-improvement rounds and 60+ agent calls for under $1 total, which Alexandr Wang called astonishing.
  • Another comparison forwarded by Alexandr Wang used the same prompt at max reasoning settings: Muse Spark 1.3 finished in one minute, while Fable 5.1 took 70 minutes and $13.
  • A developer's self-built Counter Strike benchmark showed Muse Spark 1.3 nearly tied with Fable 5.1 at much higher speed, recreating a Counter Strike game for just $1.75.
  • User elshayib (via Scobleizer's repost) said Muse Spark 1.3 may already outperform the model behind his resident Hermes agent, calling it very good, very fast, and ridiculously cheap for frontier-level intelligence.

Unconfirmed

  • All results are personal benchmarks and subjective impressions, not seen in authoritative reviews or Meta's official figures; the model's exact positioning and capability ceiling remain to be verified.

Why it matters

  • If Muse Spark 1.3 truly approaches top-tier models at minimal cost, it could dramatically lower the barrier for long-horizon agent tasks and self-improvement loops; the Scale AI CEO's repeated reposts have amplified its community impact.

Episode 4 · Meta Chief AI Officer Praises muse spark 1.3 Sub-Agent Demo (2026-09-03, 2 posts)

Meta's Chief AI Officer Alexandr Wang shared a developer's hands-on test of muse spark 1.3, calling it "pretty cool." The developer deployed parallel sub-agents that completed the task in about an hour.

Episode 5 · Meta Muse Spark 1.3 Ranks Second on Agentic Benchmark at a Third of the Cost (2026-09-04, 2 posts)

Signal65's PINNACLE benchmark shows Meta Muse Spark 1.3 as the second-strongest agentic model, with 38% fewer weighted errors than Claude Opus and per-task cost of $0.85—about a third of the leader—potentially Meta's next "Llama moment."

Episode 6 · Meta Launches Muse Spark 1.3 Max with Major Coding and Agent Gains (2026-09-05, 12 posts)

On September 5, Meta's Chief AI Officer and Scale AI founder Alexandr Wang publicly released Muse Spark 1.3 Max, with reasoning level set to the maximum tier and safety testing completed before launch. He said the new version is markedly stronger at coding and agentic tasks, and strongly recommended that users who had tried 1.3 High or XHigh try Max again. The release came only about two days after version 1.3.

Confirmed

  • The model is available via the Muse Code CLI tool and the Meta Model API, with a curl install command provided by Wang for direct terminal installation.
  • Meta's official account announced the max reasoning mode, pitched as "frontier performance, non-frontier pricing."
  • Meta's Superintelligence Labs showcased demos including 3D content generation with the model; Wang reshared it saying "We're so back."
  • Resharer @baaadas highlighted progress in visual coding.
  • Per a third-party comparison relayed by @Lianhuiq, rebuilding a 3D model of Helsinki Cathedral across 1.1→1.2→1.3→1.3 Max showed a clear leap in detail fidelity for 1.3 Max, with lower total generation cost.
  • On September 6, user @tinkerersanky praised the model as Opus-tier; Wang reshared and recommended trying it, an official-side amplification of positive word of mouth.

Why it matters

  • This is Meta's latest move in the coding and agent race, directly targeting competitors known for strong reasoning and coding, and already benchmarked by users against Claude Opus.
  • The pricing emphasizes cost-effectiveness, with third-party tests showing better detail at lower total cost, potentially intensifying price competition among frontier models.
  • Shipping a Max tier only two days after 1.3, plus repeated official recommendations to High/XHigh users, suggests the reasoning gains are seen as a generational improvement and reflects Meta's rapid iteration pace.