Meta releases Muse Wonder models as AI builders share how vendor competition shaped their choices
MarketScreener reported that Meta released Muse Wonder AI models to challenge Anthropic and OpenAI. Leaders deploying frontier models in production explain one decision they made or reconsidered as competition among vendors expanded, along with a concrete result from that choice.
Edited by Adrian James, Editor at Featured News · Roundup by Featured · Oct 15, 2026
According to MarketScreener, Meta’s Muse Wonder release adds another contender for teams evaluating AI models.
We asked leaders running these models in production to share specific decisions they have made or reconsidered because of multi-vendor competition, and the results that followed.
Perspectives
One week of internal review was what it took for us to make the switch based on budget analysis of which AI model to use in production. In the next week we burned 1.55 billion tokens at a 96% cache read rate. Priced against market list rates we spent $62 total using a budget tier model that utilized off-peak rates versus $810 for a flagship model's pricing. That was a 13 fold difference and our workflows adapted to utilizing the frontier model to review the work, this ended up costing $44 for the frontier model to review all of the work with a minimal dropoff in output quality compared directly against the prior week's output. We internally built the infrastructure to support the orchestrator frontier model reviewing the budget model's output and now utilize this framework for all of our development.
I have been working with websites: building them, ranking them, converting them to lead generators for the last 10 years. No wonder I get more than 100 link-building requests a week. Some ask if I need link-building services, and some ask for guest posts on websites I manage. 1. The first layer sorts them into link sellers and link buyers. 2. The second layer classifies link buyers as high-value or low-value prospects. 3. The third layer is finding the right template. The fallback in the system is the manual layer, for when the LLM cannot decide. Before, when I first built this application with the help of ChatGPT, I used the OpenAI API, and it used to cost me around $0.012–$0.019 per email, and it was vendor-locked. Later I reconfigured the app to use OpenRouter. Now with Qwen3-30B-A3B-Instruct-2507, it’s down to $0.0006–$0.0009 per email. That’s 20x cheaper. The only drawback with Qwen is that it sent 3% more emails to the manual layer. It’s okay for me. My take: not all tasks require new advanced models. Knowing what’s good enough can save big time on expensive LLM spending.
In the latest projects we often keep an eye on the model life cycle for specific providers. In fields where accuracy is very important, it is expensive to recalibrate all workflows and perform acceptance tests once again when a new version of the LLM vendor’s model appears. The older one gets deprecated, and prices become significantly higher. We developed an innovative live speech translation pipeline for one of our customers. It captures audio from English speakers and creates a real-time audio translation into 5 languages (with great flexibility to add more). For this pipeline several ultrafast models were needed. We chose Groq as a provider that helps to maintain 500+ tokens per second for open-source models. Initially, several models participated in the pipeline: GPT-OSS-120B and Llama 3.3 70B. The latter was especially great at translating small chunks of text into different languages, but its short life cycle (6 months) made it impossible to use. That’s why the whole pipeline now uses the industry-proven GPT-OSS-120B, which has a much longer life cycle. We stopped searching for new, faster and better models, because testing and recalibrating the pipeline is too costly: it requires 8–20 hours and the involvement of several native speakers of the supported output languages. Instead, we improved the algorithm itself to reach better results.
In mid-2026 IELTS stopped offering the test on paper. IELTS Builder Canada, who already had their mock tests published as books on Amazon, reached us through our website. They wanted a platform where their students could take mocks with a real exam experience. Reading and Listening were simple, as we mapped the correct answers and marked them after the exam. Writing and Speaking have no fixed answers, so feedback was the hard part. AI was already running the Speaking mock interview, so we advised them to extend its role to give immediate feedback and bands based on the IELTS rules. The AI is not a single model. Different models are routed through OpenRouter, and we choose the model by the complexity of the task, to control token consumption and keep the budget in check. As newer and better models came, we shifted to them for better evaluation, without being locked to one vendor. Speech-to-text stays on the school's server and only text goes out, because the students are visa applicants and some are minors. As per the data on 23 September 2026, there were 4,584 module submissions in 30 days, with no student waiting for a teacher's feedback.
My product, PolicyPing (and side product ReefReport) monitors government building rules for changes and notifies customers which sites are affected, using frontier models to review the amendments. When planning, I routed the discovery and strategic AI work through Openrouter so I could test all models, and run them side by side with the same prompts to compare results. I did not have to commit to a single vendor, and it gave me comparison data. It also allowed me to amalgamate results where required if some performed markedly better, or came up with new ideas. This flexibility was key for testing various LLMs for speed, accuracy but also capability. It clearly showed the frontier models of each provider were somewhat close to each other but the step down models were significantly 'worse' at some things than the top-tier. Fable and Opus clearly outperformed Sonnet (all Anthropic products) but were relatively evenly matched with their competing models at each tier with different focuses and largely different personalities. The trade-off for OpenRouter as an aggregator is the increased cost and fees, but there's no other straight-forward way to get access to all models on an even interface. Once the model I picked proved itself (which ended up being Opus for a balance of cost/performance), I moved virtually all of my build work to Claude subscription (in lieu of API) methods entirely to leverage all their models and also ClaudeCode and their Cowork systems that clicked for me the best. Since then this has changed but the fact remains the aggregator API use was best for testing but could cost up to 5 times more than a subscription inclusion. I've built my tech stack to be model agnostic so I can move between providers fairly easily, which is necessary with market competition changing capabilities and costs on a week by week basis at this moment in time.
Experts featured here are members of Connectively or HARO, Featured’s expert networks.