Claude Sonnet 4.5 Beats GPT-5 on Coding Benchmarks, GPT-5 Wins on Price

·1 views

A head-to-head comparison of Anthropic's Claude Sonnet 4.5 and OpenAI's GPT-5, verified as of September 2026, shows each model leading in distinct categories. Claude Sonnet 4.5 outperforms GPT-5 on coding tasks, scoring 77.2% versus 74.9% on the SWE-bench Verified benchmark, and 50.0% versus 43.8% on multi-step terminal work measured by Terminal-Bench 2. GPT-5 holds a significant cost advantage, priced at roughly 60% less on input tokens and a third less on output tokens per million compared to Sonnet 4.5. GPT-5 also leads on competition mathematics, reaching 94.6% on AIME 2025, and on multimodal understanding with an 84.2% score on MMMU. Developers prioritising code generation and agentic workflows are better served by Sonnet 4.5, while those handling high token volumes, image-heavy inputs, or complex mathematical reasoning may find GPT-5 more cost-effective.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

[Boost]

Agentic Design System sandeep kumar patel Sep 19 Add Comment 1 min read

0 comments Read more at DEV Community

ProgrammingDEV Community ·

AWS Confirms Permanent Data Loss in Middle East Regions After Military Strikes

AWS publicly acknowledged on September 15, 2026, that customer data has been permanently lost across its Bahrain (me-south-1) and UAE (me-central-1) regions following a series of military attacks between March and July 2026. The Bahrain region suffered complete, unrecoverable data loss across all three Availability Zones, while one of three AZs in the UAE region also lost data irretrievably. AWS stated that the destruction spread across multiple Availability Zones, exceeding what its regional and multi-AZ architecture was designed to withstand — effectively acknowledging that multi-AZ redundancy does not protect against large-scale physical warfare. Structural damage, power disruption, and water damage from firefighting operations were cited as key causes of the infrastructure failure. Data lost belonged primarily to customers who had not migrated workloads or maintained backups in other regions before the outages occurred.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

B.Tech Graduate Builds Secure School Portal After 2023 Dehradun Cyber Fraud Case

A recent computer science graduate from Graphic Era Hill University developed a Secure School Management Portal after learning about a 2023 cyber fraud incident in Dehradun, Uttarakhand. In that incident, school manager Shambhu Prasad Pancholi lost ₹95,000 after a fraudster posing as a student's parent sent malicious WhatsApp links that enabled four unauthorized bank transactions. The incident, reported by the Times of India, required no OTP and exploited the manager's trust in a routine fee-payment scenario. Motivated by the case, the developer built a web portal with separate access portals for admins, teachers, and students, incorporating multiple layers of security including authentication and access control. The project, completed as a final-year initiative in 2025, was a personal learning exercise and is not affiliated with the school involved in the original fraud.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How Runtime Injection Could Bring DLSS Upscaling to Unsupported Legacy Games

A technical concept dubbed DLSS5-Autopilot proposes a method to bring NVIDIA's Deep Learning Super Sampling (DLSS) to older games that never received native support. Rather than requiring game engine modifications, the hypothetical framework operates at the OS driver boundary, intercepting DirectX 12 and Vulkan graphics calls after the game engine has already issued its commands. The system is described as having three layers: an interception layer that hooks GPU API calls, a background inference engine running on Tensor Cores, and a shader hot-swap module that reroutes frames through the DLSS pipeline. This approach targets thousands of AAA titles released between 2015 and 2021 whose closed-source engines make traditional DLSS integration impossible. The article, originally published on tamiz.pro, notes that DLSS5-Autopilot is a hypothetical framework extrapolated from existing NVIDIA driver capabilities and user-space injection techniques, not a released product.

0 comments Read more at DEV Community