AI Today: GPT-6.1 Sol and Claude Sonnet 5.5

State of knowledge as of 2026-09-30. Source material gathered on 2026-09-30.

Today's review focuses on dynamic changes within the OpenAI and Anthropic ecosystems, as well as AMD infrastructure support. OpenAI introduces the new GPT-6.1 Sol model, Dots assistants, and summarizes findings from the DevDay 2026 conference. Simultaneously, Anthropic releases Claude Sonnet 5.5, increasing performance in coding and document tasks. Engineers also gain the new verl 0.9.0 library, which significantly optimizes training processes on the ROCm 10.0 architecture.

TL;DR

• GPT-6.1 Sol: A new model for programming tasks, offering performance close to Astra at an 80% lower cost.
• Dots: Proactive OpenAI assistants helping users maintain continuity of work on complex projects.
• verl 0.9.0: A library update for the ROCm 10.0 platform, supporting AMD Instinct MI300 and MI350 series GPUs.
• Claude Sonnet 5.5: A model 30% faster and cheaper than its predecessor, distinguished by high benchmark scores and visual capabilities.
• DevDay 2026: Summary of the OpenAI conference including the premiere of over 20 new solutions, including the GPT-6 Astra model and Codex tools.

OpenAI Introduces GPT-6.1 Sol

OpenAI introduced the GPT-6.1 Sol model on September 29, designed for tasks related to coding, computer use, and professional work. This model offers an intelligence level close to the Astra model. According to information shared by the company, the cost of using GPT-6.1 Sol via the API is one-fifth of the standard rates for input and output tokens of the Astra model.

For AI builders: Teams working on computer use automation can test the model's performance in their development environments. The announcement does not specify the model's availability in the EU region, token limits per request, or commercial licensing details – these are pieces of information that require verification before making a migration decision.

OpenAI Introduces Dots

OpenAI introduced Dots, proactive assistants capable of continuing work on complex projects and everyday tasks. According to the announcement published on September 29, this tool aims to help users maintain control over progress in ongoing activities.

For AI builders: Teams managing multi-stage projects can verify how assistants maintain task continuity in practice. The press release does not specify the licensing model, availability in the European Union, or whether the solution supports Polish – these issues require independent checking before planning deployment.

Verl Library Update to Version 0.9.0 for ROCm 10.0

According to an announcement from September 28, AMD released the verl 0.9.0 library for the ROCm 10.0 platform, enabling the training of post-training models using reinforcement learning on AMD Instinct GPUs.

This release introduces full support for the ROCm platform as a first-choice environment, abandoning emulation of competitive solutions. The library includes a dedicated Dockerfile.rocm that integrates vLLM 0.27.0, SGLang, and Megatron-Core 0.18. The system supports an asynchronous policy trainer for models such as Qwen2.5-VL and Qwen2.5-Math. The new version implements the verl.plugin.platform plugin layer, which automatically detects the ROCm environment and manages resources, eliminating the need for custom code patches. The updated version supports MI300 series (gfx942) and MI350 series (gfx950) GPUs, ensuring full compatibility with the Primus ecosystem (v26.7). The delta_sharded mechanism was also applied, which reduces weight synchronization costs during distributed asynchronous training.

For AI builders: Engineering teams working on AMD Instinct clusters can test the deployment using the new container structure and updated Primus stack. The publication does not provide licensing requirements for individual image components or information about the availability of ready-made cloud environments from external service providers – this must be established before starting integration.

Anthropic Introduces Claude Sonnet 5.5

Anthropic introduced the Claude Sonnet 5.5 model on September 28, being the second release in the Claude 5.5 family. This model represents an improvement over Claude Sonnet 5, offering operational speed greater by over 30% and lower operating costs for most tasks by up to 30%. Claude Sonnet 5.5 is designed for everyday tasks such as bug fixing in code or creating documents, presentations, and spreadsheets. The model complements the more advanced Claude Opus 5.5, and Claude Haiku 5.5 will also join the family in the coming weeks.

In tests, Claude Sonnet 5.5 achieved a score of 70.6% in Terminal-Bench 4.0, representing an increase compared to 10.3% achieved by its predecessor. The model demonstrates capabilities in image understanding, being the first Sonnet series model to complete the game Pokémon Red working solely on screenshots. In terms of costs, prices remain at 2 dollars per million input tokens and 10 dollars per million output tokens, while thanks to processing efficiency, the real cost of task execution is 30% lower. In terms of security, the model is equipped with cyber security safeguards previously used in the most advanced versions of Anthropic software. Test results indicate that Claude Sonnet 5.5 outperforms the previous generation in domains such as programming, conceptual work, and chart recognition.

For AI builders: Developers automating processes in the command line interface can verify model performance in multi-stage tasks using Terminal-Bench 4.0 or CursorBench 4.0 benchmarks. The announcement does not contain information about the full list of supported natural languages and exact infrastructure availability in EU regions, which requires additional verification in technical documentation before deciding to migrate the production environment.

OpenAI Summarizes DevDay 2026 Conference

According to information published on September 29, OpenAI summarized the DevDay 2026 event, during which over 20 new solutions were presented. Among the announcements was the GPT-6 Astra model, as well as updates regarding ChatGPT and the Codex tool. The producer also introduced new APIs, security features, and dedicated tools designed for creators.

For AI builders: A team building code-based applications can check the capabilities of the new Codex and integrate the announced APIs. The announcement does not state what the usage limits are for new models, whether there is support for Polish, or what the implementation costs are – this information should be verified in the documentation before starting project work.

Frequently Asked Questions (FAQ)

What are the main cost differences between Claude Sonnet 5.5 and the previous version?

Claude Sonnet 5.5 offers operating costs for most tasks lower by up to 30% compared to the previous generation. The price per million input tokens is 2 dollars, and for a million output tokens it is 10 dollars, which combined with higher processing efficiency realistically lowers the execution cost of each task by about 30%.

What GPUs does the updated verl 0.9.0 library support?

The updated verl 0.9.0 library supports AMD GPUs from the MI300 series, designated as gfx942, and the MI350 series, designated as gfx950. It enables training post-training models using reinforcement learning directly on clusters using the ROCm 10.0 platform.

Sources

• Introducing GPT-6.1 Sol — 2026-09-29
• Introducing dots — 2026-09-29
• Verl 0.9.0 on ROCm™ 10.0: Next-Generation RL Post-Training on AMD Instinct™ GPUs — 2026-09-28
• Introducing Claude Sonnet 5.5 — 2026-09-28
• DevDay 2026 Recap — 2026-09-29

--- Text automatically prepared by the aipoc-lab (AI) laboratory based on publicly available sources listed above. Own summaries, not copies. Version before human editing.