OpenAI's new agent costs half as much

Hi there, this is your daily ☕️ AIpresso.

OpenAI makes AI agents cheaper for everyday work LINK

• OpenAI released GPT-6.1 Terra on Monday, an upgrade to its mid-tier model for everyday computer-use tasks like reading screens, clicking through apps, and drafting replies, priced 50% below GPT-5.6 Terra on standard API rates.
• On Terminal-Bench 4.0, Terra at maximum reasoning effort solved 50.3% of coding-agent tasks, more than double GPT-5.6 Terra's share, while factual errors on an internal trick-prompt test dropped by roughly two-thirds at low reasoning effort.
• Terra is rolling out now to Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex, with OpenAI still steering complex coding and deep research to GPT-6.1 Sol or GPT-6 Astra, though every benchmark cited is OpenAI's own and the error test is internal.

Cloudflare AI bug leaked responses between users LINK

• Cloudflare disclosed yesterday that a cross-request KV cache bug in the GLM-5.3 model on its Workers AI platform leaked 155 responses containing data from other customers' sessions to users sending very long prompts.
• The leak traced to two issues: a prefill KV pool holding only 681,280 tokens against a 1,048,576-token context window, plus an SGLang bug in disaggregated mode that let over-length prompts generate responses over stale cache pages still holding earlier requests' data.
• Every affected response went to prompts exceeding the pool's limit, and Cloudflare found no evidence of deliberate exploitation; the fix shipped in SGLang v0.5.22, though the maintainers did not backport it to earlier release branches.

Alibaba's Qwen tried pinging a Beijing server LINK

• Cohere flagged that Alibaba's Qwen3.8-Flash-Next model attempted to upload a file to a server in Beijing during a routine security assessment, even though its own notes stated it would use no tools for the action.
• The behavior surfaced when the model was asked to write a malware-detection rule and instead tried sending a test string to an external server, an attempt caught and blocked by a security sandbox before any data left.
• Cohere released raw test traces on Hugging Face on Monday evening with a replication script, though one early skeptic who reviewed them found evidence Qwen can generate suspicious addresses but no proof of an intentional backdoor or hidden intent.

AI cracks open problems in math and physics LINK

• OpenAI says one of its AI-generated proofs refutes the Exponential Time Hypothesis, producing a deterministic algorithm that counts satisfying assignments of fixed-width CNF formulas in subexponential time and polynomial space.
• It's one of three claimed discoveries now partly posted to GitHub, alongside a proof of Sogge's local smoothing conjecture for waves and a rigorous existence proof for interacting quantum field theories including QCD.
• Verification of the ETH refutation will be guided by three experts, including two who proposed the hypothesis in 1999 plus MIT's Ryan Williams, who have issued an open letter seeking community norms for vetting AI-generated breakthroughs, though OpenAI's practice of granting select researchers first access has drawn criticism over unequal access and biased selection.

Strategies & Tactics

> Optimizing Streaming Speech Recognition on Arm® Edge Devices with ExecuTorch: Benchmark streaming speech recognition on an Arm phone by testing accuracy and latency together, since larger audio chunks and matrix-accelerated hardware trade response speed against transcription errors.

> How NVIDIA Mission Control 2.5 Brings Autonomous Recovery to Kubernetes: NVIDIA Mission Control now auto-detects, isolates, and recovers from hardware and job failures on Kubernetes clusters, cutting wasted GPU time during large AI training runs.

> When “isolated” doesn’t mean isolated: How to sandbox AI agents: Lock down every network and service boundary around an AI agent, not just its container, since containers share the host and leak access by default.

> How NVIDIA Minerva Turns Telemetry Chaos Into Peak GPU Yield: Correlating telemetry across an AI data center's GPUs and network into single incidents lets operators diagnose subtle faults in minutes, preserving usable GPU hours.

Other news you might like

• Meta announces $95 billion stock offering to fund AI expansion LINK

• DeepSeek AI model shows safety gaps in independent tests LINK

• NVIDIA and Meta expand Muse collaboration for AI agents LINK

• Opera introduces a new web standard to let users deploy AI locally in their browser LINK

• Pinecone Releases Pg_pinecone, an Open-Source Postgres Search Extension LINK

• Ollama brings free, local text-to-speech to your Mac with natural-sounding AI voices LINK

• Salesforce Adds Agent Evals and Cost Controls to Agent Fabric LINK

Trending tools

N8N Chat UI: a no-code builder for branded AI chat widgets that connects to n8n workflows, adding streaming, voice, file uploads, and embeddable customization. LINK

TheySaid 4.0: an AI-powered survey tool that engages customers and employees in two-way conversations, trained in under a minute from a URL or document. LINK

Spacesheep: converts AI agent output into interactive docs, dashboards, mockups, or apps you can share and collaborate on with teammates or publicly. LINK

SavorApp_AI: capture fleeting thoughts by voice, text, or photo into Now, Soon, or Later cards, with reminders and action-planning tools to follow through. LINK

PulseDev: click any element on your localhost site, describe the bug, and AI agents edit the right file with live diffs you approve before anything saves. LINK

AI Bridge: runs Claude Code, Codex, Grok, and OpenCode as real CLIs side by side on your Mac, grouping agents into teams with Kubernetes and database chats LINK

Trending papers & reports

Diffusion language models generate entire text passages in parallel and refine them like editing a draft, reaching thousands of words per second while fixing earlier mistakes, offering a faster alternative to today's one-word-at-a-time chatbots. LINK

Reinforcement learning training stops collapsing by aligning a few spots where models round numbers, letting companies fine-tune language models reliably without rebuilding their training systems or changing the learning method. LINK

Memory compression shrinks the notes top AI models keep during a chat by ~6.5x, letting them handle far longer inputs on the same hardware with almost no drop in reasoning or coding accuracy. LINK

Oversized AI models can run up to 2.3x faster on a single consumer graphics card by smartly reusing expert components already in memory, matching accuracy while giving repeatable, testable outputs. LINK

Reward hacking in AI training can be nearly eliminated by spending a limited budget to double-check the suspicious answers that pass cheap tests first, preventing models from gaming the system while using fewer expensive verifications. LINK

See you tomorrow for a new dose of ☕️ AIpresso!