The AI coding market used to look like a leaderboard bakeoff with nicer fonts. Now Meta has walked in with Muse Spark 1.1 and a price sheet, which is how you know the conversation has moved from model bragging to procurement math. That is less glamorous than a demo video, but for anyone building coding agents, it is also more useful. According to Quartz, Muse Spark 1.1 is priced at $1.25 per million input tokens and $4.25 per million output tokens, with $20 in free credits to start. The headline fight is Meta versus Anthropic and OpenAI, but the practical fight is inside your product margin. Coding tools do not just autocomplete one line, they read context, propose changes, explain themselves, and sometimes retry like a Roomba trapped under a chair. ## Quartz says the price tag is the opening shot Quartz frames Muse Spark 1.1 as Meta taking on Anthropic and OpenAI in the AI coding market, and the token pricing is the part builders should actually put in a spreadsheet. Input cost matters when a model is asked to ingest files, tickets, tests, logs, and whatever README archaeology your team calls documentation. Output cost matters when the model writes patches, explanations, test plans, and apologies for the patch it wrote five seconds ago. The lesson is not that the cheapest model automatically wins. The lesson is that price can make experimentation easier, especially in workflows where switching costs are still low. If a team can test a coding model against real repository tasks without setting the finance department on fire, Meta gets a shot at becoming part of the evaluation set. That is strategy disguised as a receipt. ## CNBC says Meta is selling agentic work, not just autocomplete CNBC reported that Meta introduced Muse Spark 1.1 as it tries to compete with OpenAI and Anthropic in critical areas of the market. In the same report, Alexandr Wang described Muse Spark 1.1 as Meta's "strongest model for agentic and coding work yet." That phrase matters because agentic coding is not just predicting the next semicolon, although semicolons do remain tiny punctuation landmines. For builders, agentic work changes the cost model. A simple code completion might be one model call, but an agent loop can involve planning, file inspection, tool use, test execution, correction, and final explanation. Each turn burns tokens, and each bad turn burns trust. The right comparison is not cost per million tokens in isolation, it is cost per useful completed task. ## TechCrunch sees a crowded battle, which makes pricing louder TechCrunch described Meta as entering a crowded AI coding battle with Muse Spark 1.1. Crowded markets make differentiation harder, especially when every vendor can produce a benchmark chart that looks like a subway map drawn by a caffeinated raccoon. Pricing cuts through that fog because developers can evaluate it directly against their own usage. That does not mean teams should chase the lowest visible rate. A model that needs more retries, longer prompts, or more human cleanup can become expensive in stealth mode. The sane evaluation pattern is boring and therefore powerful: run the same tasks across vendors, measure total tokens, success rate, review time, and failure modes. If the cheaper model finishes fewer tasks cleanly, the discount may be cosplay. ## Axios warns the tests are aging faster than the models Axios reported that AI models are outgrowing existing methods for testing and benchmarking their hacking abilities, leaving policymakers and corporate security teams with less clarity about what models can actually do. That point applies beyond security. As coding systems become more agentic, static benchmark scores are only part of the story. For readers choosing an AI coding vendor, the next move is practical. Treat Muse Spark 1.1 pricing as a reason to rerun your evaluations, not as permission to rip out your stack at 2 a.m. Watch whether Anthropic and OpenAI respond on pricing, whether Meta can prove quality in real coding workflows, and whether benchmarks catch up with agents that do more than type politely. The model war is becoming a margin spreadsheet, and honestly, the spreadsheet may be the most honest benchmark in the room. ## Sources - Meta is taking on Anthropic and OpenAI in the AI coding market

Sources