GPT-5.6 lands in Kiro, video gets cheaper
Your AI stack just got more hands-on. Today’s updates bring a coding model into Kiro, a full voice API to the terminal, cheaper video generation options, and safer enterprise database access for Claude.
In today’s edition
- GPT-5.6 is now in Kiro, with OpenAI citing an 82% lower cost per successful task in a benchmark setup.
- ElevenLabs CLI v1 brings the company’s entire API into the terminal for scriptable voice workflows.
- Alibaba’s Wan 3.0 video model is live on OpenRouter with text, image, and reference inputs.
- Supabase made enterprise-managed authentication for its MCP server generally available through Okta.
- OpenRouter launched visual benchmarks to compare image-model capability, price, and generation time.
GPT-5.6 joins Kiro’s production workflow
OpenAI’s GPT-5.6 is now available in Kiro, putting the model inside a spec-driven environment for planning, building, testing, and reviewing software.
- OpenAI says GPT-5.6 is entering workflows developers already use in production.
- OpenAI and AWS worked to optimize Kiro and GPT-5.6 for price-performance.
- On Terminal-Bench 2.1, OpenAI says GPT-5.6 Terra cut cost per successful task by about 82% in Kiro’s spec-driven development environment.
- The pitch is less back-and-forth prompting and more structured software work from plan through review.
What it means Run a real feature or bug-fix workflow through Kiro and compare task completion, review quality, and total spend against your current coding-agent setup.
Source: OpenAIDevs
ElevenLabs puts its whole API in the CLI
ElevenLabs released CLI v1, bringing its entire API into the terminal. Voice builders can now keep more of their audio workflow close to code, scripts, and deployment tools.
- CLI v1 exposes the entire ElevenLabs API from the terminal.
- You can use it where your build scripts and developer workflows already live.
- The release gives teams a command-line path for working with ElevenLabs rather than relying only on a browser interface.
What it means Prototype scriptable voice steps inside your existing build pipeline, then turn the commands that work into repeatable content or product workflows.
Source: ElevenLabs
Wan 3.0 arrives on OpenRouter
Alibaba’s Wan 3.0 is now live on OpenRouter for video generation from text, images, or references, with clips ranging from 2 to 30 seconds.
- Wan 3.0 supports text, image, and reference-based video generation.
- It offers 480p, 720p, and 1080p output options.
- Launch pricing is $0.05, $0.10, and $0.20 per second across those resolutions.
- OpenRouter says a limited-time 15% discount applies across every resolution.
What it means Use short, low-resolution generations to test creative concepts first, then reserve longer 1080p clips for the variants that earn a place in your product or campaign.
Source: OpenRouter
Supabase adds managed MCP access for enterprises
Supabase’s enterprise-managed authentication for its MCP server is now generally available, built with Anthropic and Okta.
- Admins authorize the Supabase connector once through Okta.
- Team members get Supabase in Claude with permissions scoped to the access they already have.
- The setup targets teams that need Claude connected to Supabase without handing out broad database access.
What it means If you sell into security-conscious teams, test permission-scoped MCP access as a way to make AI database workflows easier to approve and deploy.
Source: supabase
OpenRouter adds visual image-model benchmarks
Picking an image model is no longer just a prompt-and-pray exercise. OpenRouter launched visual benchmarks that show how models handle a range of challenging prompts.
- The benchmarks cover image-model performance across visual scenarios.
- You can compare models on tasks such as arranging multiple input images into a scene, counting fingers, and removing an object from a row.
- The grid can be sorted by price and generation time alongside capability.
- OpenRouter says it will add scenarios and update the benchmark when new models launch.
What it means Use the benchmark to shortlist models by the exact visual task your app needs, then validate only the finalists with your own prompts and budget.
Source: OpenRouter