Copilot, Vercel, and model eval updates #64

July 1, 2026 — 3 min read

Today's Letter

GitHub started rolling out Claude Opus 4.8 (fast mode) in preview for GitHub Copilot on June 29, 2026
The company says the mode keeps the same intelligence as standard Claude Opus 4.8 while delivering faster output token speeds
GitHub positions it for interactive coding and agentic workflows where lower response latency matters
The model will be selectable in VS Code, Visual Studio, Copilot CLI, GitHub Copilot cloud agent, the Copilot app, github.com, GitHub Mobile, JetBrains, Xcode, and Eclipse
Access is planned for Copilot Pro+, Max, Business, and Enterprise users
Enterprise and Business administrators must enable the Claude Opus 4.8 fast mode policy in Copilot settings, and the policy is off by default
GitHub says billing follows provider list pricing under Usage Based Billing
Fast mode is priced below previous fast modes, but still above standard Claude Opus 4.8, according to GitHub's pricing note
Availability is gradual, so the model may not appear immediately for all eligible users

Source: github.blog
More: tech-insider.org

Vercel introduced support for deploying any HTTP server defined by a `Dockerfile.vercel` file on its platform.
The flow builds the image, stores it in the project registry, and deploys it to Fluid Compute through `vercel deploy`.
The company said the same path works for Go, Rails, Spring Boot, Express, Laravel, ASP.NET, FastAPI, and nginx-backed services.
Containers must listen on `$PORT`, which defaults to `80`, with the sample using `golang:1.24-alpine` and `alpine:3.20` in a two-stage build.
Vercel positions containers as first-class workloads on the same platform as frontend apps and other services, with per-push preview URLs and shared observability.
Billing is tied to active CPU time on Fluid Compute rather than wall time, according to the announcement.
Vercel said startup is optimized through an internal boot-image format that streams and decompresses container snapshots on demand.
Containers are stateless in the current model, with persistent data expected to live in external databases or caches; attached durable storage is described as in progress.

Source: vercel.com

Hugging Face made Community Evals interoperable with Every Eval Ever, adding cross-posted evaluation results to Hub model pages.
The update also feeds those results into benchmark leaderboards registered through dataset repos with eval.yaml.
Model-side scores live in .eval_results/*.yaml, and submissions can come from model authors or third parties through pull requests.
Each linked result points back to the full EEE record, including generation settings, harness version, reproducibility notes, and optional per-sample JSONL data.
Hugging Face said the EEE datastore has about 229,000 evaluation results covering more than 22,000 models and 2,200 benchmarks.
The records were normalized from 31 reporting formats into one JSON schema for evaluator identity, model access path, settings, and metric meaning.
Hugging Face published the post on June 30, 2026, and said official organization submissions receive a verified checkmark on EvalEval.

Source: huggingface.co

Jocoletter curates AI, software, and product trends for developers and builders.

#GitHub #HuggingFace #Vercel