# Meta Releases Muse Spark 1.3 to Challenge GPT-5.6 in the Coding Wars > Source: > Published: 2026-09-02 23:24:35+00:00 *Meta shipped Muse Spark 1.3 on September 2, 2026, and the claim is simple: its coding agent is now close enough to OpenAI and Anthropic that you have to take it seriously.* Meta did not bury this release. Muse Spark 1.3 went into Muse Code and the Meta Model API on the same day, giving developers the new model without waiting for a separate product cycle. That speed is the point. Muse Spark first arrived in April, then 1.1 followed in July, then 1.2 on August 5. Now 1.3 is here less than a month later. That is fast. It also tells you where Meta wants the argument to move. The company has spent years trying to prove that its AI work is more than a research pipeline feeding Facebook, Instagram and WhatsApp. Coding agents give it a cleaner test: can the model take a messy job, work across files, use tools, and avoid making a dangerous mess while doing it? ## The Benchmark Claim Is Narrow, But Real Here's the scorecard. According to Meta's own September 2 release and evaluation report, Muse Spark 1.3 scored 75.4 on DeepSWE v1.1, a long-horizon software engineering benchmark covering 113 tasks across 91 repositories and five languages. Meta says that puts it ahead of Anthropic's Claude Opus 5 and OpenAI's GPT-5.6 Sol on that test. On Terminal-Bench 2.1, Meta reported 88.8. That matches GPT-5.6 Sol on Meta's comparison table. On JobBench and OSWorld 2.0, which test professional tool use and computer operation, Muse Spark 1.3 posted 64.9 and 66.9, ahead of GPT-5.6 Sol in Meta's numbers, though not ahead of Opus 5 on every agentic measure. [Meta is preparing an AI pendant for the wearable race](https://startupfortune.com/meta-is-preparing-an-ai-pendant-for-the-wearable-race/) Meta is reportedly developing an AI-powered pendant as part of a broader wearables roadmap. The move shows how the company is looking beyond glasses and headsets toward ambient AI devices that could feed its assistant, subscriptions and future consumer agents. - [Meta AI pendant wearable device announcement](https://startupfortune.com/meta-is-preparing-an-ai-pendant-for-the-wearable-race/) - [how to wear AI without glasses headset](https://startupfortune.com/meta-is-preparing-an-ai-pendant-for-the-wearable-race/) So don't read this as a clean sweep. It isn't. Meta's own table shows Opus 5 still leading on GDPVal-AA v2, and GPT-5.6 Sol ahead on DeepSearchQA and Meta's internal Agentic IF Index. The better read is more specific: Meta now has a coding and terminal-work result strong enough that rivals can't wave it away as a catch-up story. The long-context scores are the sharpest part of the release. Muse Spark 1.3 posted 98.5 on MRCR v2 at 256K to 512K tokens, and 98.1 at 512K to 1M tokens, compared with GPT-5.6 Sol's 91.5 and 73.8 in Meta's table. If you work on large codebases, that matters more than a pretty demo: losing the thread is where agents usually become expensive. ## The Safer Agent Is The Real Pitch Meta's bigger claim sits away from the leaderboard: Spark 1.3 is better calibrated on irreversible actions. It asks clarifying questions when a prompt is ambiguous, calls in the user when it gets stuck, and confirms before taking consequential steps. It also claims stronger resistance to prompt injection and adversarial inputs. Here's the thing: you do not only worry about whether a coding agent writes clean code. You worry about whether it deletes the wrong file, overwrites a branch, or follows a malicious instruction buried in an issue comment. A model that knows when to stop is more useful than one that blindly finishes the task. Meta also says 1.3 uses about 20% fewer tool calls and 25% fewer tokens than Muse Spark 1.2 in its internal comparisons. That is not glamorous, but it is exactly the kind of detail developers notice. Agentic coding gets expensive when the model loops, re-reads, re-plans and narrates every move. Less waste is a product feature. Pricing has not moved. VentureBeat reported at the Muse Spark 1.2 launch that standard access was $1.25 per million input tokens and $4.25 per million output tokens, with a contributor tier at $0.10 input and $0.20 output if customers allow Meta to use prompts and completions to improve its products. Meta and Axios both described Spark 1.3 as coming at the same price as 1.2. That creates a blunt choice for teams. Pay the normal rate and keep your data out of training, or take the cheaper tier and accept the trade. Don't dress that up. For a startup sending proprietary code, the discount is only useful if the data bargain is acceptable. [Meta Copies Snapchat's Homework Again With Instagram Plus](https://startupfortune.com/meta-copies-snapchats-homework-again-with-instagram-plus/) Meta is rolling out Instagram Plus, Facebook Plus, and WhatsApp Plus subscriptions starting at 2.99to2.99to3.99 per month, directly following Snapchat Plus's successful playbook as the company builds toward a broader Meta One subscription ecosystem that includes AI and creator tiers. - [Meta Instagram Plus subscription feature launch pricing](https://startupfortune.com/meta-copies-snapchats-homework-again-with-instagram-plus/) - [how Meta copies Snapchat's subscription model for growth](https://startupfortune.com/meta-copies-snapchats-homework-again-with-instagram-plus/) ## Meta Needed A Straightforward Win The release lands after a noisy stretch for Meta's broader agent push. TechCrunch reported that Meta bought Manus, the Singapore-based AI agent startup, in December 2025, then reported in June 2026 that Meta had begun unwinding the $2 billion deal after pressure from Beijing. That is not the kind of AI story any company wants following it into the next product cycle. Its smart glasses have brought a different problem. The Verge reported last week that Meta added a privacy fix so the camera stops recording if the LED capture light is covered during filming, building on earlier tamper protections. Business Insider reported that Meta has disabled camera functions on thousands of glasses after detecting recording-light tampering. Against that backdrop, Muse Spark 1.3 is cleaner. No consumer privacy fight. No cross-border acquisition tangle. Just a model release, a coding product, a public API, and benchmark claims that can be argued over by people who actually run these tools. That is enough. If you are choosing a coding agent today, you still should not buy the whole story from Meta's table alone. Run it on your own repository, with your own tests, and watch what it does when the task gets vague. But Meta has earned a place in that comparison now, and that is the part its competitors should notice. **Also read:** [Independent Restaurant Owners Do Not Need Another Dashboard, They Need Continuity](https://startupfortune.com/independent-restaurant-owners-do-not-need-another-dashboard-they-need-continuity/) • [Zoltán József Szabó on Értéksziget: A Good Medical Aid Is Not Enough on Its Own](https://startupfortune.com/zoltan-jozsef-szabo-erteksziget-a-good-medical-aid-is-not-enough-on-its-own/) • [OnlyFans Turned $715 Million in Profit Last Year With Just 47 Employees](https://startupfortune.com/onlyfans-turned-715-million-in-profit-last-year-with-just-47-employees/)