SpaceX's Grok 4.5 and Meta's Muse Spark 1.1 are the latest signals that major artificial intelligence companies are racing into the developer market. As later entrants trailing OpenAI and Anthropic, both companies are leaning heavily on coding and agentic capabilities, along with sharply lower token pricing, to draw developers to their platforms.
The centerpiece of Muse Spark 1.1, which Meta unveiled on the 9th, is what the company calls agentic coding: the ability to manage an entire software project on a developer's behalf. The release also marks the first time Meta has offered a commercial A.I. application programming interface, or A.P.I., to outside developers.
Meta had previously made its Llama models available externally, but only as open-weight models — developers downloaded the model weights and ran them on their own hardware or cloud infrastructure, fine-tuning as needed. Meta itself never operated an inference service the way OpenAI does; instead, third parties like Together A.I., Fireworks A.I. and Hugging Face supplied Llama access through their own A.P.I.s.
Muse Spark 1.1 represents Meta's first direct entry into the A.I. A.P.I. business, a shift analysts see as significant. The company has poured hundreds of billions of dollars into A.I. development, but has generated comparatively little direct revenue from its models — its core business remains advertising sharpened by A.I.-driven targeting.
The move into paid A.P.I. access signals that Meta now intends to sell its A.I. models directly to corporate customers. Meta's stock rose on the news, climbing 4.6 percent on the 9th after having slid on a Reuters report that the company would begin producing its own chips starting in September.
For the first time, Meta is operating its own A.I. infrastructure, charging developers per token, and offering an A.P.I. compatible with OpenAI's format. Developers can now call the Muse architecture directly through the Meta Model A.P.I., currently available in public preview, to build their own applications.
The push reflects the fact that Muse Spark 1.1 is both a general-purpose model and a capable coding agent. It was built by Meta Superintelligence Labs under Alexander Wang, who joined Meta after the company acquired a 49 percent stake in his company, Scale A.I., for $14.3 billion last June.
Meta has said it rebuilt its infrastructure, data pipelines and model architecture from the ground up for Muse Spark — a distinct lineage from Llama, not an iteration of it. Industry observers expect the two model families to diverge going forward, with Llama continuing as an open-weight offering while Muse Spark anchors Meta's commercial services. Meta may also keep releasing smaller and midsize Llama 4 models, such as Scout (109 billion parameters) and Maverick (400 billion parameters), while restricting its largest model, Behemoth (2 trillion parameters), to research uses such as serving as a teacher model for distillation.
According to Meta, Muse Spark 1.1 can autonomously manage large codebases — parsing millions of lines of existing code, diagnosing bugs on its own, implementing new features, and even carrying out large-scale migrations from older code versions to newer ones without human direction.
The model also relies on a main agent that coordinates parallel work among subagents. When given a coding task, the main agent devises a plan and assigns roles — front-end development, back-end development, debugging — to subagents that then work simultaneously. Meta says this parallel structure has significantly improved both the speed and efficiency of coding tasks.
In a post on its A.I. blog, Meta quoted the chief executive of Cline, an A.I. coding agent platform and an early Muse Spark 1.1 partner, who described the model as a cost-effective option capable of running serious agentic coding work at scale.
The chief executive of Replit, another prominent A.I. coding tool, offered a similarly positive assessment, pointing to the model's support for context windows of up to a million tokens, full multimodal capability across images, video and PDFs, built-in search with source citations, strong reasoning, top-tier coding performance — particularly in front-end and design work — along with structured output, parallel tool calling and an OpenAI-compatible A.P.I., calling it a complete foundation for building agentic A.I.
Despite its emphasis on coding and agents, Muse Spark 1.1 remains fundamentally a general-purpose, multimodal model, capable of processing images and video and handling ordinary conversation. Like OpenAI's GPT models and Anthropic's Claude, it is accessible directly through Meta's official A.I. platform, currently at no cost.
For everyday users, the model's million-token context window allows for long, complex exchanges — uploading a lengthy PDF report, a dense multi-page contract or an academic paper and asking the model to flag inconsistencies or summarize key points.
Muse Spark 1.1 also works in tandem with Meta's new image-generation model, Muse Image, allowing users to describe a website concept in natural language and receive both a visual mockup and working HTML code. Meta is rolling out the Muse Spark 1.1 engine to power A.I. chatbots across its own services, including Facebook, Instagram and WhatsApp.
A day earlier, on the 8th, SpaceX unveiled Grok 4.5, likewise emphasizing coding and agentic performance. The model was co-developed with Cursor, the coding tool SpaceX acquired last month for $60 billion, with the two companies collaborating since May, ahead of the formal acquisition.
Grok 4.5 has 1.5 trillion parameters, three times the size of its predecessor, Grok 4. SpaceX says the model runs on a new "V9" architecture and posted top-tier results on benchmarks such as SWE Marathon, which test performance on large, real-world codebases.
The company said it strengthened Grok 4.5's practical coding-agent abilities by incorporating supplemental training on real usage data drawn from Cursor, the widely used A.I. code editor.
According to the trade publication AI Chronicle, Grok 4.5 was trained on billions of lines of code optimized through the Cursor platform, giving it an understanding of complex system architecture rather than surface-level syntax.
SpaceX says the model can manage entire code repositories, catch logical errors that conventional debuggers miss, and even suggest ways to optimize cloud resource usage.
Analysts point to three broad reasons Meta, SpaceX and other major A.I. companies are focusing so heavily on coding: the revenue potential of enterprise A.P.I. sales, the way coding data appears to strengthen a model's broader reasoning abilities, and coding's role as a leading use case for A.I. agents. The lucrative developer market and coding's apparent links to stronger general reasoning and agent performance are seen as reinforcing one another.
Courting developers, in particular, is seen as a gateway into the broader corporate market. Because most developers already use A.I. coding tools inside their organizations, adoption at the individual level tends to embed a company's models deep within enterprise workflows.
Because a model's reasoning and coding abilities are closely intertwined, strengthening one tends to support the other — laying the groundwork for future enterprise A.I. agents and agentic services.
That connection stems partly from the fact that code, unlike most natural-language reasoning, is objectively verifiable. Where it can be difficult to judge whether an answer in plain language is correct, code either compiles and runs or it doesn't — a judgment made instantly by compilers, tests and runtime environments. Unit tests can be automatically graded, and outputs can be checked precisely against expected results.
That verifiability makes code particularly well suited to a training method known as Reinforcement Learning with Verifiable Rewards, or R.L.V.R.
R.L.V.R. swaps out human judgment for an automated grader. Traditional A.I. training, known as reinforcement learning from human feedback, relies on people to evaluate which answers are correct. R.L.V.R. instead uses the results of compiling or testing code as the reward signal.
Under R.L.V.R., rule-based systems — compilers, calculators, answer keys — do the grading. Code that runs correctly earns a reward; code that fails does not, creating a feedback loop that lets a model refine itself through repeated trial and error, sometimes across thousands or tens of thousands of attempts, gradually converging on correct logical solutions.
That process is thought to sharpen a model's ability to reason step by step and solve problems logically.
Writing code requires breaking a problem into smaller components, tracking intermediate states such as variables and conditions, connecting logical cause and effect, and tracing errors back to their source when something goes wrong — a process structurally similar to solving math problems, working through logic puzzles, or general planning tasks. Training on large code datasets, researchers say, tends to strengthen both step-by-step problem solving and broader reasoning at once.
Several factors reinforce one another: code's suitability for automated grading makes it useful raw material for large-scale reinforcement learning; that training appears to build genuine reasoning ability; and the multistep patterns involved in coding agents — calling tools, planning across steps, verifying intermediate results, retrying after failure — closely resemble the work done by research agents or office-automation agents more broadly.
Because coding and reasoning reinforce each other, a model's coding strength becomes a foundation for future enterprise agentic tools — one reason Meta, SpaceX, OpenAI, Anthropic and Google have all doubled down on coding performance even as they promote their agentic capabilities.
That emphasis is reflected in recent benchmarks like OSWorld, SWE-bench and TerminalBench, most of which center on coding and tool-use ability. Coding benchmarks such as SWE-bench are increasingly treated as proxies for how capable a model is as a general-purpose agent.
Speaking to CNBC about the Muse Spark 1.1 release, Alexander Wang, who leads Meta Superintelligence Labs, said, "You kind of have to build coding capabilities as part of that in service of overall agentic capabilities" — framing coding skill as a core building block of agentic performance.
A related trend is emerging among the newer entrants: unlike OpenAI and Anthropic, which maintain separate coding-focused products (Codex alongside GPT, Claude Code alongside Claude), Meta and SpaceX are folding general-purpose and coding capabilities into a single model.
Codex and Claude Code are themselves still built on top of GPT and Claude, respectively, with a coding-specific harness layered on — tools for file editing, running terminal commands and integrating with Git bolted onto the underlying general-purpose model.
As recently as 2024, it was common for companies to run separate general-purpose and coding-specific models. But as evidence mounted that coding and reasoning ability are closely linked, and as general-purpose models themselves grew more capable, the industry has increasingly converged on offering coding through a single, general-purpose model. OpenAI and Anthropic still package their coding agents, Codex and Claude Code, as distinct products layered on GPT and Claude.
Meta and SpaceX, by contrast, are integrating general-purpose and coding capability directly into Muse Spark 1.1 and Grok 4.5. Meta has also emphasized that it trained its model to work smoothly with third-party agent harnesses developers already use, an effort to gain traction quickly in the developer market.
Of the two, Grok 4.5 stands out as the largest model that makes no architectural distinction between general-purpose and coding capability. It remains a full frontier model capable of real-time search, document writing and data analysis, but its training incorporated trillions of interactions between real developers and the Cursor coding editor. SpaceX has said that because strong coding ability tends to correlate with strong reasoning, it deliberately infused Grok 4.5 with what it describes as high-quality "coding genes."
Newer entrants are also uniformly emphasizing lower token costs relative to established frontier models. Meta AI remains free for casual users of Muse Spark, and the company has not announced paid consumer tiers comparable to ChatGPT Plus or Claude's subscription plans.
Businesses and developers integrating Muse Spark through the Meta Model A.P.I. do pay, though Meta has set aggressive pricing: $1.25 per million input tokens and $4.25 per million output tokens, with U.S. developers also receiving $20 in initial testing credits.
SpaceX has likewise cut Grok 4.5's A.P.I. pricing well below earlier levels, even as it keeps its existing subscription structure. The new pricing is $2 per million input tokens and $6 per million output tokens, with cached input priced at just 50 cents per million tokens.
Notably, Grok 4.5 costs more than its predecessor, Grok 4.3. SpaceX attributes the increase to higher performance and greater token efficiency, arguing that the cost per completed task remains competitive. Consumer subscriptions remain unchanged, with SuperGrok Lite at $10 a month, SuperGrok at $30, and SuperGrok Heavy at $300.
Among current frontier models, Google offers the lowest A.P.I. pricing, charging $2 and $12 per million input and output tokens, respectively, for Gemini 3.1 Pro. OpenAI's top model in its latest 5.6 series, called Sol, is priced at $5 and $30, while Anthropic's Claude Fable 5 runs $10 and $50. Because OpenAI and Anthropic price based on their most capable models, with smaller models offered far more cheaply, the gap narrows considerably lower in each company's lineup.
Meta's pricing undercuts OpenAI's by roughly 86 percent and Anthropic's by about 92 percent. Unlike OpenAI and Anthropic, for whom A.P.I. revenue is central to the business, Meta draws substantial cash flow from advertising and SpaceX from its Starlink business — giving both companies room to prioritize rapid developer and enterprise adoption over near-term profitability from their A.I. models.