Meta launched Muse Spark 1.3 on 3 September, its fourth flagship model release in five months, and within hours the new model had leapfrogged the Gemini 3.8 Flash that Google had unveiled the same day. The duel marks the fastest cadence of frontier releases the industry has seen, with model rankings now turning over in hours rather than quarters.
A 62-point score, a same-day reversal
On Artificial Analysis, Muse Spark 1.3 scored 62, behind only Claude Fable 5.1 at 66 and Claude Opus 5 at 63. Google’s Gemini 3.8 Flash, released just hours earlier, registered 59. Meta chief executive Mark Zuckerberg called the model’s performance “beyond imagination”, while Meta’s chief AI officer Alexander Wang taunted Google on social media with “gemini who?” and suggested Google could still cancel the launch.
Meta has now shipped four Muse Spark iterations since April, moving from the original to 1.1 in July, 1.2 in August and 1.3 in September. The company says Muse Spark 1.3 tops five benchmark categories and beats GPT-5.6 Sol and Claude Opus 5 on most long-context and coding tests, tying GPT-5.6 Sol on Terminal-Bench. Its weaker spots remain agentic work such as web-search agents and general-knowledge tasks.
Pricing below the Chinese frontier
The more striking number is cost. API pricing holds at USD 1.25 per million input tokens on a cache miss, USD 0.15 on a cache hit and USD 4.25 per million output tokens. That is slightly above Gemini 3.8 Flash but below DeepSeek V4-Pro at peak pricing, a deliberate squeeze on the Chinese models that reset the market’s price expectations earlier this year.
Developers testing the model report fast, cheap runs with solid but not spectacular output. One built a Minecraft game for 10 US cents. Another ran more than 60 agentic tasks across two hours for under a dollar. A third compared Muse Spark 1.3 with Qwen 3.8 Max, GLM 5.3 and GPT-5.6 Sol, judging GLM 5.3 the best all-rounder and Qwen 3.8 Max the highest in output quality, while praising Muse Spark’s cost and speed.
Built for long tasks and coding
Meta frames Muse Spark 1.3 around long-horizon work. The model can juggle multiple workflows inside one long conversation, pull context from messy and contradictory sources, call tools autonomously and correct its own plan. Compared with Muse Spark 1.2, Meta engineers measured 20 per cent fewer tool calls and 25 per cent fewer tokens on coding tasks. Safety work was upgraded around agentic robustness and prompt-injection resistance, with the model now judging risk before executing irreversible actions.
The takeaway for enterprises is less about any single benchmark than about route selection. Meta chose not to chase every dimension at once, concentrating instead on long-context and coding strengths where it beats frontier rivals on real engineering tasks. The pattern points toward a future of multi-model routing, with companies dispatching each task to the model built for it rather than betting everything on one all-rounder.
Editor’s note: This is an adapted translation of the original Zhidx report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://www.zhidx.com/p/590181.html.