Edge deploys are 2.4× faster.
Rewrote the propagation layer to push to all 184 POPs in parallel from the closest origin. Median time from prompt to globally-live dropped from 58s to 24s. Cold start for new domains is unchanged.
New tools, faster GPUs, better policies. We ship most weeks. Sometimes daily. Subscribe and find out what your computer can do now that it couldn't last week.
The computer can now reach for 8× H100 NVL clusters with NVLink and NDR InfiniBand interconnect — same fabric the foundation labs use. Spin-up under 90 seconds, billed per-second, released automatically when training finishes.
eu-fr-1, us-east-2, us-west-1Rewrote the propagation layer to push to all 184 POPs in parallel from the closest origin. Median time from prompt to globally-live dropped from 58s to 24s. Cold start for new domains is unchanged.
Define per-tool, per-cost, per-user approval rules in a small policy DSL. The computer pauses for human review before crossing your thresholds. Available in Team and Enterprise.
# .computer/policy.txt rule production-deploys: when tool == "edge.deploy" and env == "prod" require review_by("@platform") rule gpu-budget: when tool.cost_per_hour > 10 require approval("@cost-owner")
Long-running cold-tier retrievals that crossed the 60s mark would silently fail with a 504. They now stream correctly through the agent. Affected a small number of archive reads since May 27.
Press ⌘K from anywhere. Jump to any project, fire any tool, switch any team, change any policy — without leaving the keyboard.
edge.deploy --stagingAuto-provision and deprovision users from Okta, Entra ID, Google Workspace, and JumpCloud. Group-based role mapping. Available on Team and Enterprise.
Quadrupled the default context window for Pro and Team. Longer codebases, longer documents, longer agent threads — all without summarization losses.
Enterprise customers can now run the AI computer on their own brand domain — desktop.yourbrand.com — with custom logo, color tokens, typography, SSO, and ToS. Six pilot customers already in production.
Added vector.index, vector.query, db.postgres, db.redis, queue.send, and seven more. The agent can now stand up data plumbing without you spelling out the stack.
Per-project, per-service, per-tool cost telemetry now updates every 5 seconds. The agent now also includes cost in its plans — "this will cost about $0.40" — before reaching for a GPU.
Point the computer at your Azure OpenAI deployment, your AWS Bedrock account, your in-house weights, or any OpenAI-compatible endpoint. Hybrid stacks (computer.com models for routing, your model for inference) work out of the box.
/v1/legacy/chat on August 1.The pre-1.0 chat endpoint will be removed on Aug 1, 2026. Migrate to /v2/prompt — shape is mostly compatible and supports tools natively. The agent can do the migration for you in most projects.
Today we're opening the public beta of the AI computer at desktop.computer.com. Twelve years of cloud, edge, GPU, and streaming infrastructure — now reachable from one prompt.
If you had a computer.com account, it just became an AI computer. Sign in and meet it.
Final five edge cases in V4 signing and presigned-URL parameter handling are resolved. Existing boto3 and aws-sdk applications point at s3.computer.com and just work.
Added Lagos, Nairobi, Mumbai, Bangalore, Jakarta, Manila, Ho Chi Minh, Karachi, Cairo, Riyadh, Tashkent, and Almaty. 184 POPs across six continents.
Stream every prompt, tool call, and infra action into Splunk, Datadog, or an S3 bucket of your choice. Enterprise customers can now retain audit logs for up to 7 years with immutable storage.
Subscribe to the changelog, or just open the computer and ask it what's new.