Announced 9 Oct 2026 · Sources checked
What did Pine publish, and when?
The company page is titled “Introducing Pine Computer: A Computer Built for Humans AI” and dated 5 October 2026. It is signed Dylan Wang. The lede says the computer is “2–5× faster” and “1/25 the model cost,” then immediately points at a SaaS-Bench v1.1 table. A footnote on that page labels the speed claim as “Pine’s preliminary internal tests against AI on conventional computers, including SaaS-Bench tasks against Claude Code with Opus 5.”
PR Newswire’s “Pine AI Introduces Pine Computer, a Computer Built for AI,” which we opened at https://www.prnewswire.com/news-releases/pine-ai-introduces-pine-computer-a-computer-built-for-ai-302903667.html, datelines San Francisco, 9 October 2026, and stamps the note 12:00 ET.
This is a hosted computer-use runtime, not a new frontier model. For why a browser agent is a different object from a chatbot, see computer-use agents. For an evaluation that already treats computer use as a scored task, see OpenAI’s Ironclad write-up.
How does Pine Computer say it works?
The blog’s comparison table is the useful sketch. On a personal computer, “a person directs the steps” and the model watches a screen. On Pine Computer, “an application submits a task through an API,” “changes [are] delivered directly to AI,” and “each computer is sealed off on its own, with scoped access.” Saved state, the post says, “is encrypted with your own key.” The FAQ on the same page says the cloud object is “virtual computers + a harness + a runtime layer wired into the OS and browser + a matching set of local tools + remote API tools.”
That is Pine’s architecture copy. The mechanism it names is notification instead of a screenshot loop: “the operating system, the browser and the applications tell the model what changed, the way epoll tells a program.” Pictures “stay for people, who can watch the live screen, take the controls and hand them back.” We did not inspect a session log or a notification payload.
The trust page we opened, https://pinecomputer.io/trust/, repeats the isolation claims in operator language: each computer “runs isolated in its own container under gVisor, never reused”; network policy “blocks other computers, private networks and the cloud’s metadata service”; saved state is “AES-256-GCM” under a customer key “protected by hardware (TPM)”; data “stays in the United States”; “SOC 2 Type II audit is in progress.” Those sentences are Pine’s. We did not audit the container or the TPM.
What do the SaaS-Bench numbers actually measure?
UniPat AI’s SaaS-Bench page, dated 28 September 2026 and opened on 10 October, defines two scores. Checkpoint score is the “weighted fraction of passed checkpoints.” Resolved score is 1 only when every checkpoint passes. The v1.1 table lists nine model–harness pairs on 106 tasks across 23 applications. The Pine Computer row is labeled “Pine Computer Runtime (PCR)”: 335.5 steps per task, about $3.60 cost per task, 306.2 tool calls, 78.3% checkpoint, 27.4% resolved.
The same table’s highest resolved score is Opus 5 with Claude Code at 31.1%, with a 74.3% checkpoint score and about $26.50 per task. GPT-5.6 Sol with Codex is 71.1% / 29.2% / about $20.50. Pine’s blog reprints 78.3% / 27.4% for “GPT-5.6 Luna on Pine Computer” and prints $1.02 as “model cost / task.” A FAQ on that page says $3.60 “is what the runs cost at the list price of our own consumer product” and $1.02 “is the model tokens at public list prices.” Those are two denominators. UniPat’s ~$3.60 sits in UniPat’s column; $1.02 sits in Pine’s.
Pine’s Hugging Face dataset pine-ai/pine-computer-saasbench-results, README opened 10 October, lists two releases that “are not a controlled before/after comparison.” The latest batch, “SaaS-Bench 1.1 · September 18–19, 2026,” prints 78.33% checks passed (1,077 / 1,375) and 29/106 fully resolved (27.36%). It says the completed batch had 125 attempts, 19 invalid or superseded, sixteen healthcare reruns after grader fixes, and eight timeouts, and “This is not a first-attempt-only score.” Raw traces “remain private.” That README is Pine’s evidence index, not an independent replay.
| System on the table | Checkpoint | Resolved | Cost / task on that page |
|---|---|---|---|
| Opus 5 / Claude Code | 74.3% | 31.1% | ~$26.50 (UniPat) |
| GPT-5.6 Sol / Codex | 71.1% | 29.2% | ~$20.50 (UniPat) |
| Pine Computer Runtime | 78.3% | 27.4% | ~$3.60 (UniPat); $1.02 model tokens (Pine blog) |
What can a developer do tonight?
The docs index at https://console.pinecomputer.io/docs lists an overview, “How a Computer works,” a Getting started guide marked Beta, and guides for sessions, agent events, skills, files and shell, a Web SDK named @pinesandbox/computer-web, persistence, tokens, and integration. A closing line says “Approved developers get the integration guides, the API reference and the SDKs.” The public index is a menu. It is not the reference.
The use-cases page lists 26 examples, from “Renew permits” to “Review a release.” The permit example shows a Ruby snippet: pine.create_computer with location: { country: "US" } and ephemeral: true, then session.agent.run, then a needs_input event when a person must type a password. The blog FAQ says a new computer “starts in a few seconds,” a paused one “resumes in under a second,” and “Can I run it on my own machine? No.” Bring-your-own-model “is coming”; “it works from screenshots today,” the 9 October wire says.
A live desktop a person can seize is an approval gate, not a sandbox by itself. For the control you want before any tool that can click a real site, see agent approval gates and agent sandboxing.
What sits next to it on the computer-use shelf?
OpenAI’s dots agents and the Ironclad computer-use evaluation are different products: a consumer/work agent and a scored computer-use set. Together Link puts open models inside an existing coding harness. None of those pages is a gVisor cloud PC you spawn from an SDK. Pine’s PR “About” block also mentions Pine Voice and a τ³-Voice rank; that is a different product line on the same wire.
The 9 October copy quotes Andrew Mackenzie, co-founder of Subliminal, as saying Pine had “thought about all of this and built it all in.” It also claims “in one enterprise deployment it helps automate audits, for a customer whose team now takes on 50% more work with the same people.” Those two sentences are the wire’s. They are not a named customer appendix we opened.
Website terms at pinecomputer.io/terms say “Website descriptions, demonstrations and benchmark examples are information, not an offer of a particular price or a guarantee of a particular result.” That is the vendor’s own limit on the table.
What should a buyer check before treating 78.3% as a product fact?
Ask which number the seller is using. UniPat’s checkpoint score, UniPat’s resolved score, Pine’s $1.02 token average, UniPat’s ~$3.60 row, and the Hugging Face 29/106 batch with documented reruns are five printed objects. They are close. They are not one measurement.
Ask whether the job can finish. On UniPat’s board Pine leads checkpoint progress and trails Opus 5 / Claude Code and GPT-5.6 Sol / Codex on resolved tasks. Pine’s blog says so in the same table. A product that hands a permit portal or a checkout to a computer needs the resolved column, not only the checkpoint column.
This is an evidence review of the 5 October blog, the 9 October PR Newswire note, UniPat’s SaaS-Bench v1.1 page, Pine’s Hugging Face results README, and the public docs, trust, use-case and terms pages. We did not join the waitlist, create a computer, or replay a SaaS-Bench task.
Common questions
Is Pine Computer a model?
Not on the pages we opened. The blog FAQ says it is virtual computers plus a harness and a runtime, and that it currently runs Pine’s own model and GPT-5.6 Luna. Bring-your-own-model is described as coming.
Can I install it on a laptop tonight?
No. The blog FAQ says it runs in Pine’s cloud. The public docs index is a menu; the API reference is for approved beta developers. The waitlist is at pinecomputer.io.
Why does Pine print $1.02 when UniPat prints about $3.60?
Pine’s FAQ says $3.60 is its consumer-product list cost for the runs and $1.02 is model tokens at public list prices, the basis it wants compared with the other systems. UniPat’s table prints ~$3.60 on the Pine Computer Runtime row. We priced neither.
What to remember
Use the 9 October wire for private beta and the waitlist. Use UniPat’s table for 78.3% checkpoint and 27.4% resolved. Keep Pine’s $1.02 token line and the Hugging Face rerun batch in their own columns. Do not treat a checkpoint lead as finished work.
Sources & further reading
- Introducing Pine Computer: A Computer Built for AI ↗
- Pine AI Introduces Pine Computer, a Computer Built for AI ↗
- SaaS-Bench: Can Computer-Use Agents Leverage Real-World SaaS to Solve Professional Workflows? ↗
- Pine Computer — SaaSBench Evaluation Results (README) ↗
- Documentation ↗
- Trust and security ↗
- Use cases ↗
- Terms of Use ↗
How this story was made
Written by Kristian Kostov with AI assistance and checked against the linked sources. Company performance claims are attributed to the company. Analysis reflects AiLookout’s interpretation; we have not independently tested the products discussed. Cover photography is illustrative and does not depict the specific announcement or product.
Our editorial standards
