OpenAI cuts GPT-6 prices as Anthropic launches Claude Opus 5.5

Anthropic introduced Claude Opus 5.5 on September 22, 2026. The first model in the Claude 5.5 family is available through Anthropic’s platform, AWS, Google Cloud, and Microsoft Azure.¹² Anthropic says it performs at the level of Claude Fable 5.1 on many tasks. Typical tasks cost about 40% less than with Opus 5, while Reuters reports that it generates output more than 30% faster.³¹

The lower overall cost does not come from token prices alone: Anthropic says the model also uses fewer tokens to complete tasks.³¹ For companies, cost per completed task is therefore more useful than list price alone. Reported strengths include software development and chart recognition, but the available results do not establish an across-the-board lead in visual design.³¹ METR found the model might slightly accelerate AI research and development, but was unlikely to automate it fully.⁴

OpenAI introduced GPT-6 Sol and GPT-6 Luna on September 22 as lower-cost models in the GPT-6 lineup.¹² They are available in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu users, as well as through the API.¹² Sol’s API prices are half the promotional rates for GPT-5.6 Sol. Luna also costs less than GPT-5.6 Luna.¹

Sol targets more complex work, including coding and automation, while Luna is intended for lighter, high-volume tasks.³⁴ Anthropic introduced Claude Opus 5.5 the same day, adding another option for demanding software development work.⁵ OpenAI also says Sol makes about half as many mistakes as its predecessor. That figure comes from OpenAI’s internal factuality evaluation, not an independent comparison.⁴

OpenAI expanded ChatGPT Voice for ChatGPT Work on September 23. Users can start tasks by voice, create documents, presentations, and spreadsheets, and use connected apps or a browser. Voice is available in Work on the web and mobile devices. If a voice call ends, an unfinished task can continue in text.¹²

The update brings voice control into longer workflows: users can start a task, ask about its progress, and redirect work while it is underway. OpenAI also describes these capabilities for Codex.¹³ For businesses, the handoff from conversation to text matters when a task takes longer than the call. The published notes do not fully specify plan eligibility or usage limits for every feature.¹³

OpenAI introduced Astra for Law on September 17, 2026. The offering combines GPT-6 Astra with instructions for legal research and drafting and a daily-updated search index of US legal sources.¹² Select US law firms initially have access through Trusted Access in ChatGPT and Codex; API access is planned.²

The index includes case law, statutes, and regulations, among other sources.¹ The offering puts OpenAI more directly into the market for legal AI software, where specialized vendors already serve law firms.³ Astra for Law is not a separate new model family: it is a configuration of GPT-6 Astra tailored to legal work.¹² Its limited rollout means firms outside the access program cannot yet use it as a generally available service.²

Chinese company StepFun introduced Step 5 Preview in September 2026. The model is available in preview through StepFun’s API and studio. Its stated context window is one million tokens. StepFun targets tasks including software development and knowledge work.¹²

The large context window could make it easier to submit extensive documents or codebases in a single request. By itself, however, the window size does not show how reliably the model finds or uses information within them. Practical testing therefore matters more than the maximum input length alone. Companies also need to assess response quality, latency, and costs for long requests. The available coverage does not establish a reliable performance comparison with other models.³²

On September 22, Unsloth published a GGUF-quantized version of Qwen-Image-2.1 on Hugging Face and announced support for running it locally with its tools. The available files include several quantizations of the image generation and editing model. In its announcement, Unsloth said the model can run locally with 12 GB of VRAM.¹²³

The release gives users another way to run the image model on their own hardware rather than through a hosted service. File size, however, is not the same as memory required during generation. One GGUF file is smaller than 4 GB, but the available information does not confirm that the full workflow runs with only 4 GB of memory. Users should check requirements for their chosen configuration.¹³

xAI introduced Grok 4.7 on September 21, 2026. The company calls it its most capable model for coding and knowledge work, citing improved self-checking in its reasoning and stronger safeguards. It has a 500,000-token context window. Grok 4.7 is available immediately in Cursor and Grok Build, as well as through the Grok API, other coding tools, model routers, and cloud platforms, according to xAI.¹

That range of access gives developers ways to use the model in existing tools or integrate it into their own applications. xAI lists separate input and output prices and offers a faster, more expensive variant.¹ GitHub is also rolling out Grok 4.7 gradually in GitHub Copilot.² Users will need to test whether the announced improvements help with their particular workloads.

White Circle introduced Halo on September 21, 2026, as an open-source framework for post-training and distributed training of Hugging Face models.¹ It is designed to work with existing models and TRL workflows without rewriting model definitions or changing checkpoint formats.¹ In its own tests, White Circle reported 2.3 to 2.8 times the throughput of stock TRL, along with lower peak memory use.¹

Halo targets teams that want to train existing models across multiple accelerators without reimplementing them for another framework.¹ It supports expert, context, and tensor parallelism, as well as a combination of expert and tensor parallelism.¹ The performance figures come from the company's tests. They do not establish whether the same gains apply to other models, training tasks, or hardware configurations.¹

PrismML released Ternary Bonsai 2 27B on September 17, 2026. The model is based on Qwen3.8 27B and uses ternary weights; the company says it is more than nine times smaller than the FP16 parent model. In PrismML’s evaluation across 20 benchmarks, it retained 98.2% of the parent’s average performance. The model weights are available for free.¹²

The smaller footprint makes the model relevant to local AI applications where memory is limited.² Bonsai 2 builds on PrismML’s earlier Bonsai 27B and is presented as an improvement in quality at a similar level of compression.¹² The performance figures come from the company’s tests, however. Organizations considering deployment should check whether the results hold for their own workloads.²

On September 16, Empero released Qwen3.8-35B-A3B-Distill alongside quantized GGUF files for local inference. The mixture-of-experts model has 35 billion parameters in total, with about 3 billion active per token. Its model card says it was distilled from Qwen3.8 teacher outputs into the Qwen3.6-35B-A3B architecture and targets reasoning, math, coding, and tool use. The GGUF version is intended for llama.cpp.¹²³

Activating fewer parameters can reduce computation per token, but it does not give the model the memory footprint of a model with only 3 billion total parameters. The entire selected weight file still needs to fit in RAM or VRAM. For companies considering local deployment, the chosen quantization therefore matters alongside available compute. The published model cards provide no benchmark results for assessing performance.²³