Replit makes OpenAI its default model
Replit CEO Amjad Masad spoke at TechCrunch Disrupt on Thursday 15 October (US time) and said the company is no longer working with Anthropic. He gave cost, partnership terms and performance as the reasons, and said OpenAI's models do better at building vibe-coded apps from scratch. Later he clarified that users can still pick Claude models, but ChatGPT is now the default and OpenAI is the "preferred partner" for Replit Agent. Masad said Replit was a top-five Anthropic customer until recently. One unnamed insider told TechCrunch the Anthropic partnership is still in place. Masad also said enterprise grew from about 2% of Replit's revenue last year to about 15% this year.
For builders: Most users keep whatever model a platform picks for them by default, so if you build on Replit Agent, test your existing projects against the new default. If you want stable behaviour, choose your model yourself.
OpenCode breach exposes stored provider keys
On 17 October, Anomaly Innovations (formerly SST), the company behind the open-source OpenCode coding agent, emailed affected users. It said OpenAI, Anthropic and Google API keys saved in OpenCode Console workspaces may have been exposed. About 1.4% of Console users were affected, all of them people who had linked external provider keys. The company said it found unauthorised access to a legacy system that was being retired, and described the incident as "related to a large scale distillation attack." It's working with the three providers. Some users have reported unfamiliar charges of up to hundreds of dollars. The company says it can't yet confirm whether the keys were misused.
For builders: If you've linked keys to OpenCode Console, revoke and reissue them directly in each provider's dashboard, check your usage back to 17 October, and set spending limits. Remember to update any CI jobs and local configs that still use the old keys.
Meta releases Muse Spark Open
On 15 October Meta released Muse Spark Open. It's a natively multimodal open-weight model with 816 billion parameters and a 1M-token context window, built for coding and agentic work. Meta calls it the most capable open-weight model in the US, based on its own evaluations. Running it locally takes at least 1 TB of RAM, and it runs best on a Mac Studio with an M5 Ultra. Meta also launched Muse Code, a coding terminal paired with the model. The model is available now through several API platforms.
For builders: Few teams have 1 TB machines, so most will try this through hosted APIs first. Run your own evaluations before you rely on Meta's benchmark claims.
DeepSeek ships V4.1-Flash-1017
DeepSeek published V4.1-Flash-1017 on 17 October under the MIT licence. It's a weights-only update to the V4.1 Flash checkpoint, with the same architecture and tokenizer. According to DeepSeek's model card, scores rose from 74.2 to 77.0 on DeepSWE v1.1, from 31.2 to 42.3 on Terminal-Bench 4.0, and from 36.8 to 42.6 on HLE without tools. Context stays at 1M tokens. OpenRouter lists a maximum output of 393,216 tokens and a base price of $0.15 per million input tokens and $0.60 per million output tokens. DeepSeek's deepseek-flash API name now serves the new version.
For builders: If you call deepseek-flash, your traffic has already moved to the new weights. Rerun your regression tests, especially on agentic and terminal tasks where the scores changed most.
Character.AI files for Chapter 11
Character.AI's Chapter 11 filing was reported on 13 October. It lists assets and liabilities of between $50 million and $100 million. Disney will buy some assets for $20 million and get a non-exclusive, perpetual licence to the technology. The sale is expected to close in late November or early December. CEO Colin Adams said the platform will shut down if no viable bid comes in. The company had 20 million monthly users in October 2024, and it has faced lawsuits and a Texas investigation over child safety.
For builders: If your consumer AI product serves young users, plan for safety and legal costs from the start. They weighed heavily here.
Bumble agrees to a $1.3 billion take-private deal
On 13 October Bumble agreed to be bought by Ithaca Holdings for $1.3 billion. Bumble's 10-Q puts the price at $6.05 per Class A share in cash. Paying users fell 16.4% year on year to about 3.2 million in Q2 2026. Bumble says it's been exploring AI-driven features and a new way for users to interact that could replace swiping.
For builders: Going private gives Bumble room to try AI matching features away from quarterly scrutiny, so expect more trials like this in consumer apps.
Sources
- Hugging Face: DeepSeek-V4.1-Flash-1017 model card
- Bumble Inc. Form 10-Q (SEC)
- TechCrunch: Replit is no longer working with Anthropic, CEO says
- Startup Fortune: OpenCode breach exposes stored OpenAI and Anthropic keys
- TechCrunch: Meta launches an open-weight model that runs on Macs
- TechCrunch: Character.AI files for bankruptcy
- TechCrunch: Bumble agrees to a $1.3B acquisition deal
- Mer.vin: DeepSeek V4.1 Flash 1017 open weights, API and pricing
Talk about it with people building with AI
Our Auckland meetups are open to everyone. Members get a ForgeBox and the members' projects area for NZ$599 a year.