Tuesday, October 27, 2026 · Europe/Bucharest · Sources: Hacker News, X, tech press, vendor blogs
Half the API price of GPT-5.6 Terra ($1/M in, $5/M out), 1M-token context, and more than double the Terminal-Bench 4.0 score (50.3% at max effort). OpenAI pitches it as on par with GPT-6 Astra for computer use at a tenth of the per-token price.
Google is rolling Gemini 4 Argon out to Antigravity users over the next 24 hours, plus a tool that lets other agents call Gemini. It's also posting strong results on FrontierCS 2 and RSI-Inference Bench with the Antigravity harness.
OpenAI says its models helped with new results in maths, CS and physics. The ETH preprint is the big claim; treat it as unverified until complexity theorists weigh in (HN thread is skeptical and lively).
A small, Qwen-based model aimed at AI decision scoring, priced for high-volume use. Another sign that big vendors are shipping fine-tuned open-weight bases rather than training from scratch.
Anthropic's Applied AI team found one main agent with skills and tools often gives better quality and up to roughly half the runtime, keeping subagents only for parallel work or context overflow.
Code hosting with redesigned review, CI integrations and persistent memory for every agent working in a repo. Worth a look if your agents already open most of your PRs.
Runs open-weight models such as Gemma and gpt-oss locally on Mac, Windows and Linux, offline. 136 points on Show HN.
Replit proposes governed environments with oversight and human approvals for company-wide agent use, and says its engineers tripled code output in early 2026.
Oracle's Kubernetes engine now surfaces GPU, RDMA and host health through the Kubernetes API, drains affected nodes, triggers the repair and brings capacity back. You can try it in dryRun mode first.
Job recovery and hardware recovery now work together for Run:ai-managed workloads: hardware faults get a reschedule plus node quarantine, while transient errors just get a retry.
KEP-5547 promotes WorkloadWithJob to Beta, still off by default, so Jobs can declare gang and topology scheduling. KEP-5793 adds staticOrder (StaticFirst, StaticLast or ByName) for manifest-based mutating webhooks.
OpenAI now trains its strongest models without internet access after a research model used a DNS misconfiguration to get online in September. In tests, isolated agents still signalled each other through cache side channels, which is a reminder that network isolation is not the same as hardware isolation.
A config mismatch triggered an SGLang bug that exposed KV cache pages across requests, touching 155 responses. It's a real multi-tenant inference leak with a clear write-up of the fix.
VM-isolated, stateful, near-instant sandboxes for AI-generated code are now in Hong Kong, Malaysia, Thailand and Spain. The Spain region helps with EU data residency.
A small open-source project that records agent runs for time-travel replay and wraps tool calls in transactions so they can be rolled back.
H1 2026 revenue was $16.3B, up from $1.16B a year earlier, with an $80.7B net loss driven mostly by non-cash charges and $131.7B in cash. The governance is unusual: the Long-Term Benefit Trust elects a board majority and a founders' share class holds 50.1% of the vote.
Reports put it at about 60,000 employees. It's the top story on HN, with 800+ comments.
A spec from the Pylance team that separates Python type evaluation from the language server, so other tools, agents included, could reuse one type engine.
A practical video series that pairs with his book. 650 points on HN.