Meta launches Muse Spark 1.1 to challenge OpenAI and Anthropic in AI coding

Meta rolls out Muse Spark 1.1 with public-preview API

Meta on Thursday introduced Muse Spark 1.1, an update to the AI model it first released in April, and opened a public-preview developer portal where users can sign up to access the technology, CNBC reported. In an interview with CNBC, Meta chief AI officer Alexandr Wang called Muse Spark 1.1 Meta’s “strongest model for agentic and coding work yet.” The earlier version of Muse Spark had been limited to “select partners” through a private API preview, according to the report.

The new model is the second release under Meta’s Muse family this week: on Tuesday, Meta released Muse Image — internally code-named Mango — a model aimed at creators and advertisers. Muse Spark’s own code name was Avocado, CNBC said.

Pricing, distribution and self-imposed limits

Wang said every new API account will start with $20 in free credits. After that, Meta will charge $1.25 per million tokens for input and $4.25 per million tokens of output — pricing he characterised as “very aggressive and attractive” compared with similar offerings from Anthropic and OpenAI. “The goal is to really have attractive pricing that scales with immense consumption usage,” Wang told CNBC.

Despite the developer-portal launch, Meta is for now restricting distribution. A spokesperson told CNBC that, while some early partners already have access and new users can join a waitlist, API access is being limited to Meta’s own properties rather than being opened to third-party platforms such as OpenRouter. “This is going to be served on top of the computer infrastructure that we’ve built,” Wang said.

Coding as the wedge into “agentic” AI

CNBC reported that Meta Superintelligence Labs trained Muse Spark 1.1 specifically on coding tasks. Wang said coding is foundational to agentic systems that can autonomously perform chains of actions. “You kind of have to build coding capabilities as part of that in service of overall agentic capabilities,” he said.

Wang said Meta trained the model “to be able to work well with all of the most popular harnesses that developers use today,” referencing tools such as OpenClaw, a coding framework that gained sudden popularity earlier in 2026 and has been credited with helping developers manage AI models powering agentic assistants. He added that Muse Spark 1.1 outperformed rival models on certain tasks involving interaction with third-party coding products.

Open-source retreat, with an asterisk

Meta’s previous AI strategy centred on its Llama family of open-source releases. Wang said that posture has shifted as MSL now sells proprietary API access. He told CNBC, however, that Meta remains “committed to open source” and that an open-source variant of Muse Spark “is in development.” Wang declined to give a release date. A more powerful successor model, code-named Watermelon, is currently being trained, he said, without a release timeline either.

The competitive backdrop — and the regulator in the room

The launch lands in a market the US government has begun to actively shape. According to Channel News Asia and AFP, OpenAI on Thursday publicly released its new GPT-5.6 family — comprising Sol, Terra and Luna — after the Trump administration gave a “green light” for a broader launch, per Axios. A White House official told AFP the wider rollout was not an approval, saying OpenAI submitted the model for scrutiny voluntarily. AFP and Channel News Asia also reported that Anthropic last week began restoring access to its Fable 5 and Mythos 5 models after Washington lifted a restriction on where they could be released.

These releases follow an early-June order in which the Trump administration required Anthropic to block non-Americans from using Mythos 5 and Fable 5 — a directive Anthropic effectively implemented by pulling the models offline entirely, according to France 24. France 24 also reported that OpenAI agreed to let the government approve every customer for GPT-5.6. Both OpenAI and Anthropic have filed confidential IPO documents and are targeting public listings at valuations approaching $1 trillion, according to Channel News Asia and AFP.

Why it matters

The stakes are unusually concrete. Anthropic and OpenAI — both racing toward IPOs at near-$1-trillion valuations — have just absorbed government-imposed distribution shocks that, for several weeks in June, cut off the most powerful versions of their models. Meta is offering developers a workaround: a coding-focused model with API availability at low token prices and no foreign-user restriction, at least for now. CNBC quoted Wang as describing one of his own uses — searching the web, reading academic papers and accessing personal health data — as “one of these use cases that really encapsulates the needs of these agentic systems,” signalling the consumer-health lane Meta intends to “dog-food” with its own model.

Wall Street pressure, not just developer goodwill, is also in the frame. CNBC reported that Meta’s AI infrastructure spending is at the rate of hyperscaler peers, but the company has no cloud-infrastructure business (though it plans to start one) and “failed to keep up with OpenAI, Anthropic and Google in developing popular models and AI applications” — language the network attributed to its own reporting context rather than to any Meta executive.

Where the reporting diverges

The sources do not fully agree on what the US government’s actions amount to. Axios, cited by Channel News Asia, described a “green light” for OpenAI’s broad launch; a White House official, quoted by AFP, denied that any approval was given and said submissions were voluntary. France 24 characterised the government’s moves as a “de facto ban” on Anthropic’s top models, while Channel News Asia and AFP reported the same episode as a “restriction” rather than a ban. Each framing carries different implications: a voluntary review suggests industry-led self-regulation, while a “ban” implies coercive state action.

The sources also frame Alex Karp’s recent CNBC appearance differently. The New York Times Opinion column described Karp, chief executive of Palantir, as declaring “the jig is up” on closed-model AI in language the paper called “vivid and spastic.” CNBC, by contrast, paraphrased him as saying “something has gone completely wrong.” The Times emphasised Karp’s conflict of interest — including a Palantir-Nvidia partnership and recent French intelligence and UK NHS contract pressures — more heavily than CNBC did.

Comparisons and scale

Wang’s per-token figures — $1.25 per million input tokens and $4.25 per million output tokens — are pitched as undercutting Anthropic and OpenAI, but the reporting does not include side-by-side benchmarks. CNBC did note OpenAI has priced Terra, the mid-tier GPT-5.6 model, at half the cost of GPT-5.5, and that OpenAI chief executive Sam Altman told CNBC GPT-5.6 is 54% more token-efficient for agentic coding. Palo Alto Networks chief executive Nikesh Arora, separately on CNBC, said token efficiency needs to drop another 20% within 12 months and 90% the year after.

A separate data point from France 24 puts competitive pressure in another light: on OpenRouter, the aggregated share of usage of Google, Anthropic and OpenAI dropped from 55% to 33% between January and June 2026, with China’s open DeepSeek now leading on the platform.

Different angles and stakeholders

For developers, the launch may relieve a near-term supply shock: an alternative coding model available immediately, at low cost and without the customer-by-customer US vetting now applied to frontier Anthropic and OpenAI models. For OpenAI and Anthropic, the threat is asymmetric — Meta can grow share without resolving the open-versus-closed debate their CEOs are publicly fighting. For Palantir and its allies, including the New York Times Opinion writer, Meta’s pivot toward proprietary API sales partly vindicates Karp’s “AI sovereignty” critique even though Meta retains open-source intentions for a variant of Muse Spark.

The White House faces a different calculus: any move that visibly blesses Meta while restricting Anthropic and OpenAI risks the appearance — already flagged by the Times — that political dynamics, not technical review, are determining which labs get to ship at scale. The Times noted Karp’s contract pressures in France and the UK, suggesting that the politics of frontier AI extend well beyond US borders.

What to watch next

Several specific milestones will move this story. The first is whether Meta opens the Muse Spark 1.1 waitlist broadly in the coming days, and whether it expands from Meta-only properties to third-party platforms such as OpenRouter. The second is the release timing of the open-source Muse Spark variant Wang said is in development. The third is how GPT-5.6 Sol, Terra and Luna — publicly released on July 9 — hold up against Muse Spark 1.1 on agentic coding benchmarks, and how Anthropic prices Mythos 5 and Fable 5 now that US distribution restrictions have eased. The fourth is the wider policy track: the White House is reportedly drafting criteria for which AI models fall under new security restrictions in line with a recent executive order, which could affect whether Meta’s frontier work eventually comes under the same regime now applied to OpenAI and Anthropic.

Distribution friction and the moat it builds

Meta’s choice to host Muse Spark 1.1 exclusively on its own infrastructure, rather than route it through aggregators such as OpenRouter, is a quietly significant structural move. Aggregators are often where developers benchmark and default to new models because they reduce switching costs; staying off those rails means Meta forgoes that low-friction discovery channel. The trade-off appears intentional: by owning the serving stack end-to-end, Meta collects first-party telemetry on coding workloads, retains pricing control, and avoids ceding margin to intermediaries. The risk is thinner adoption at the very moment OpenAI is widening GPT-5.6 distribution following its government review. Whether the self-imposed limit is a temporary bridge to broader availability, or a longer-running posture mirroring the proprietary pivot away from Llama-style openness, will shape how seriously developers treat Muse Spark as a default rather than a fallback.

The coding wedge narrows the field of play

Training Muse Spark 1.1 specifically on coding tasks, and tuning it for compatibility with the OpenClaw harness, signals that Meta is contesting a narrower battleground than the general-purpose frontier model race. This is consequential because coding has become the proxy metric by which “agentic” capability is judged: investors, enterprise buyers and rival labs increasingly treat SWE-bench-style results as a shorthand for broader autonomy. If Meta can claim credible leadership there, it pressures OpenAI and Anthropic on the dimension that most influences downstream agentic product narratives, even without leading on general reasoning benchmarks. The second-order effect is a possible commoditisation of coding-only APIs, where price elasticity is high and switching is cheap — which would favour a hyperscaler-style subsidised pricing strategy and penalise labs whose margins depend on premium coding tiers.

Regulatory asymmetry as an underpriced variable

The launch is unfolding in an environment where US authorities have demonstrably altered distribution for Anthropic’s top models and conditioned OpenAI’s customer pipeline on government approval. Meta has not been the subject of comparable directives, according to the reporting, and is offering a coding API without geographic restrictions. That asymmetry creates an arbitrage window for developers whose workloads were disrupted in June, and it gives Meta leverage in procurement conversations where compliance teams are now asking vendors hard questions about availability risk. But the reporting also shows the regulatory picture is unsettled: sources frame the same government actions as a “green light,” a “de facto ban,” or a “restriction,” and a White House official explicitly rejected the approval framing. This suggests the rules of the road are being written in real time, and Meta’s current advantage could narrow quickly if Washington’s posture broadens to additional labs.