Securitylab•October 27, 2026•🇷🇺Translated from Russian

OpenAI Releases GPT-6.1 Terra: Cost-Effective Model Focused on Computer Control, Image Analysis and Practical Automation

OpenAI has introduced GPT-6.1 Terra, a new language model designed to bridge the gap between impressive demonstrations and reliable day-to-day computer work. Unlike flagship models optimized for maximum reasoning depth, Terra focuses on practical execution across images, documents, applications and graphical interfaces while keeping costs significantly lower.

The model sits in the middle of OpenAI’s GPT-6 lineup. GPT-6 Astra targets the hardest tasks, GPT-6.1 Sol offers high performance at moderate expense, and GPT-6 Luna handles simple high-volume operations. Terra delivers professional-grade results on many benchmarks at a price closer to the affordable tier.

Key technical specifications include a 1,050,000-token context window, support for text and image inputs, maximum generation of 128,000 tokens, and six reasoning levels ranging from none to max. The model works through the Responses API and supports structured outputs, though it does not directly handle audio or video files.

Improvements over GPT-5.6 Terra are most visible in visual analysis and computer control. The model can interpret technical drawings, product photographs, screenshots and interface layouts, then translate its understanding into concrete actions such as filling forms, navigating applications or extracting data between systems.

Benchmark results published by OpenAI show clear progress. On Terminal-Bench 4.0 the success rate rose from 21.5 % to 50.3 % at maximum reasoning depth. Agents' Last Exam and OSWorld 2.1 also recorded double-digit gains at roughly one-third the previous cost. Error rates on an internal factual-accuracy test fell from 53.6 % to 16.7 % at the lowest reasoning setting.

Pricing has been reduced to $1 per million input tokens and $5 per million output tokens, half the rate of GPT-6.1 Sol and one-tenth the rate of GPT-6 Astra. Cached inputs are available at $0.05 per million. The model is currently accessible to Plus, Pro, Enterprise and Edu subscribers through ChatGPT Work and Codex, as well as via the OpenAI API under the identifier gpt-6.1-terra.

Related articles

Habr•Other

Home Lab Overhaul: Securing Proxmox Infrastructure After Discovering 63,000 Forgejo Spam Accounts

A home lab administrator migrated Docker stacks to new LXC containers on Proxmox and uncovered multiple security misconfigurations. The setup relied on Cloudflare proxying, MikroTik port forwarding to NPMplus with CrowdSec, and Cloudflare Tunnels for admin panels. After fixing IP address logging in the WAF, logs revealed 63,021 spam accounts created on a Forgejo instance that had open registration. The operator disabled registrations, recreated the instance from verified Git mirrors, and eliminated direct internet exposure by routing all traffic through Cloudflare Tunnel. Additional controls included Zero Trust policies, device enrollment restrictions, and TOTP on Guacamole. The changes removed open ports, enforced proper client IP visibility for CrowdSec, and segmented public, limited, and admin access paths.

Habr•Other

Benchmarking AES-GCM and Ascon AEAD Performance on ESP32 and ESP8266 for Industrial Telemetry

Researchers conducted detailed experiments measuring the real-world costs of three AEAD implementations on ESP32 and ESP8266 microcontrollers used in industrial telemetry scenarios. The study compared mbedTLS AES-128-GCM, legacy Ascon-128 via CryptoLW, and a portable C++ port of NIST Ascon-AEAD128 across three load profiles ranging from stable low-frequency packets to maximum-rate 10 ms intervals. Key metrics included AEAD execution time, CPU cycles, memory usage, UDP transmission success rates, and power consumption captured via TC66C. Initial tests revealed significant UDP packet loss on ESP32 under MAX load due to ENOMEM and ENOBUFS errors in the lwIP stack, which were largely resolved after network handling improvements. Results showed AES-GCM completing in roughly 618 µs on ESP32 versus 803 µs for legacy Ascon-128, while the reverse held on the slower ESP8266 where legacy Ascon outperformed AES by a factor of approximately 3.5. The work highlights that network stack limitations often dominate over cryptographic overhead in high-throughput industrial sensor deployments.

AntiMalware•Other

ARinteg Earns Golden Partner Status from Ideco for NGFW Security Projects

ARinteg has been awarded Golden Partner status by Ideco following successful joint projects in network security. The integrator supplies and implements the Ideco NGFW Novum next-generation firewall for perimeter protection and secure remote access. Collaboration also covers customer migrations from foreign NGFW solutions to the Russian vendor’s product. Commercial Director Dmitry Slobodenyuk stated the new status will allow ARinteg to deliver fuller support across all implementation stages. Ideco’s Partner Relations Head Sergey Markin highlighted ARinteg’s technical expertise in large-scale deployments. The partnership supports import substitution and hybrid infrastructure requirements.

Habr•Other

Zero Trust in Practice: Avoiding YAML Pitfalls with Kubernetes NetworkPolicy and Istio

A detailed case study demonstrates how a single misplaced hyphen in a Kubernetes NetworkPolicy YAML file can unintentionally grant cluster-wide access instead of restricting it to a single client Pod. The example uses a shop namespace API connecting to an orders service in the backend namespace over TCP/8080, showing the correct use of namespaceSelector and podSelector together within the same from element. Tests on K3s with four vCPU and 4 GB RAM confirmed that only the intended client succeeds while others receive Connection refused. The implementation then extends to Istio with PeerAuthentication in STRICT mode and an AuthorizationPolicy limiting requests to GET /orders/* from the orders-api ServiceAccount. Additional checks cover RBAC permissions for creating Pods with specific ServiceAccounts and the need to verify DNS routes and existing allow-all policies. The article stresses maintaining an explicit list of required connections, testing both success and denial cases, and preserving rollback manifests before applying changes in production.