Meta's Muse Spark 1.3 scores 75.4% on DeepSWE, ahead of Opus 5

Meta says its new coding model has passed Claude Opus 5 and GPT-5.6 Sol on the benchmark the industry currently argues about most.

The numbers Meta reports

Muse Spark 1.3 was released on September 2, 2026. Meta reports 75.4% on DeepSWE v1.1, the long-horizon software engineering test, which it says puts the model ahead of Anthropic’s Claude Opus 5 and OpenAI’s GPT-5.6 Sol. On Terminal-Bench 2.1 it reports 88.8, and on the MRCR long-context tests 98.5 and 98.1 at 256K to 512K and 512K to 1M tokens. Meta describes the release as its largest improvement in coding and agentic work to date.

Efficiency is the other claim. According to Meta, 1.3 completes the same tasks with about 25% fewer tokens and 20% fewer tool calls than Muse Spark 1.2, which matters for agent workloads where the bill is driven by how many steps a model takes rather than by a single answer.

Availability

The model is live in Muse Code and through the Meta API. Meta has said open weights are coming for the Muse Spark line, but has not committed to a date for 1.3, and reporting suggests the decision on this version is still open.

How to read the leaderboard

Every frontier lab now publishes its own DeepSWE number on launch day, and each one leads on the day it is published. The useful comparison is the token count per completed task, which Meta is reporting and which the closed labs mostly are not. If the 25% efficiency figure holds up in independent runs, it is the more durable claim in this release.