8DIGITS news

Development.

Timely takes on technology and systems that matter for small and mid-sized businesses.

Claude Haiku 5.5 is generally available with effort control, a 1 million token window, and list prices of $0.10 input and $0.50 output per million tokens on prompts up to 100K. Anthropic says typical runs cost about 75% less than Haiku 4.5. Teams already on Claude can shift classification, extraction, and subagent traffic this week and leave the expensive model for the step that needs it.

Development · Brief · 4 min · SoloAdopt

Claude Haiku 5.5 puts effort control and a 1M window on a $0.10 routing model

Move high-volume routing and extraction to Haiku 5.5. Keep Sonnet for the hard steps.

Mistral Large 4 is live as a preview API, with open weights planned for the end of October after red-team work with security partners and authorities. It is a real European frontier-scale candidate, multimodal, and aimed at coding, cyber, finance, and law. Independent scores are not the full picture yet, so the sane move is a preview trial, not a production cutover.

Development · Brief · 3 min · SoloWatch

Mistral Large 4 is in preview, and the open weights are the part worth waiting for

Try the preview API on one European workload. Do not migrate until the weights ship.

Supabase is acquiring Turso and says it already launches more than one million databases a week, increasingly from agents. The stated direction is a short-lived database per agent task. That product is not something you turn on this week, and current Supabase projects stay on Postgres.

Development · Brief · 3 min · SoloWatch

Supabase is buying Turso so agents can have their own databases. Nothing to migrate

No migration. Keep Postgres. Watch for per-agent databases before you redesign environments.

Unified.to’s September update, published 2 October, adds free Agent Skills that teach Claude Code, Codex, and Cursor how to build against its API, including OAuth app setup in vendor portals. The catalogue stood at 1,030 integrations at the end of September. The skills are instructions, not a finished connector.

Development · Brief · 4 min · SoloTest

Unified.to published 100+ skills that tell a coding agent how to build an integration

Give one real integration’s Skill to Claude Code or Codex and measure the first pull request.

DeepSeek released Harness v0.2 as a desktop preview for Apple silicon Macs and 64-bit Windows, with the MIT harness underneath. Every layer is a plugin, and it accepts OpenAI-compatible endpoints. Breaking changes are still expected.

Development · Brief · 4 min · SoloTest

DeepSeek’s free desktop agent runs on your own model. It is still a preview.

Install DeepSeek Harness on a spare folder. Do not put client repos in a preview agent.

WebMCP is an emerging web standard that lets websites expose structured tools that AI agents can use directly.

Development · Brief · 4 min · SoloTest

Android Studio will take Claude or Codex. The project context stays in the IDE.

Try BYOA on one Android task in Canary. Do not make it the team default.

Claude Code cloud sessions keep running after you close your laptop and can execute several independent jobs at once. You start them from the terminal, browser, desktop or phone, then teleport a session back into your local checkout when you want to inspect it. The useful test is accepted work per hour of human review, not how many agents you launch.

Development · Brief · 4 min · AgencyTest

Claude Code keeps coding after you close the laptop. Review is now the limit.

Test three independent backlog jobs in the cloud. Count accepted work per hour of review.

Book a 20-minute diagnosis