GPT-6 Sol and Luna launch at lower prices, Claude Opus 5.5 follows

OpenAI introduced GPT-6 Sol and GPT-6 Luna on September 22 as additions to the GPT-6 lineup. Sol is aimed at professional tasks including coding, automation, and computer use. Luna is designed to handle high volumes of requests at lower cost. Both models are available through the API and in Codex and ChatGPT Work for Plus, Pro, Business, Enterprise, and Edu users.¹²

OpenAI says API prices are 50% below the previous promotional prices for GPT-5.6, citing improvements in caching and inference. Sol and Luna supplement GPT-6 Astra rather than replace it. Companies can evaluate separate models for demanding workflows and large volumes of simpler requests. Claims that Sol makes fewer mistakes are based on an internal OpenAI evaluation.²

Anthropic introduced Claude Opus 5.5 on September 22, 2026. The model is available through the Claude platform and cloud providers.¹² API pricing is $4 per million input tokens and $20 per million output tokens.² Anthropic says Opus 5.5 generates output more than 30% faster than Opus 5 on typical workloads.²

Anthropic estimates that typical workloads cost about 40% less because the model uses fewer tokens to complete them.² That figure reflects more than the cut in per-token prices; actual savings will depend on the task.² The company also reports stronger results in some developer tests, along with improvements in visual chart recognition and design-related evaluations.² Those claims do not establish an overall lead in visual design.²

OpenAI introduced Astra for Law on September 17, 2026. It is a GPT-6 Astra configuration for legal research and drafting, not a separate model.¹² It combines tailored instructions with a Legal Search Index for U.S. law.¹³ OpenAI says the index contains more than 230 million URLs and is updated daily with new sources.¹³ Access initially goes to selected law firms.⁴

The offering puts OpenAI in more direct competition with legal AI providers already serving law firms.² OpenAI lists Harvey and Legora among its partners.⁴² The search index could make relevant sources easier to find. Its focus on U.S. law limits its use for other jurisdictions, however, and legal teams still need to check cited authorities and assess the arguments built on them.¹³

xAI released Grok 4.7 on September 21, 2026, positioning it for coding and knowledge work.¹ The model offers configurable reasoning, is designed to check its answers more effectively, and supports a 500,000-token context window according to xAI’s documentation.¹² At launch, it was available through Cursor, Grok Build, and the Grok API; xAI also listed access through third-party providers.¹

API pricing remains unchanged from Grok 4.6.¹² That pairs a larger context window and reasoning changes with the same standard per-token rates.¹² The practical size of the improvement cannot be established from the company’s claims alone. For teams choosing a model, tests using their own codebases, documents, and workflows will still matter more than general performance claims.

OpenAI expanded ChatGPT Voice on September 23 to handle tasks in ChatGPT Work. Users can request documents, presentations, websites, and spreadsheets by speaking on the web or in mobile apps.¹² If a voice session ends before a task is finished, the task can continue in text.¹ Longer tasks therefore do not have to be completed during the voice session.¹

The update brings voice further into ChatGPT workflows rather than limiting it to spoken answers.¹² Voice can also use plugins to work with connected services.¹ Which features are available depends on the user's plan and workspace settings.¹² These capabilities are part of ChatGPT Work; the announcement does not describe a separate office product.¹²

Unsloth released a GGUF-quantized version of Qwen-Image-2.1 on Hugging Face on September 22. The repository offers several variants for local use.¹² Qwen’s underlying model combines text-to-image generation and image editing. It also supports generating and editing images with transparency.³

Memory requirements are central to local image workflows. Unsloth says its GGUF version can run with 12 GB of VRAM.⁴ The published information does not, however, establish a general minimum of just 4 GB: the repository contains quantization variants of different sizes without specifying that requirement for the release as a whole.¹⁴ The release gives users another way to run the model locally. Whether it works on a particular machine depends on the variant and needs to be tested against the available hardware.¹⁴

On September 21, 2026, White Circle released Halo, an open-source framework for distributed training and post-training of Hugging Face models.¹² It is designed to work with existing models and training workflows without requiring a model rewrite or a separate checkpoint format.¹² White Circle says the resulting models remain loadable through the standard from_pretrained path.¹

White Circle reports up to 2.8 times the throughput of stock TRL in its own tests, alongside lower peak memory use.¹ These are company results; the provided sources do not establish independently verified performance across other models or hardware.¹ For teams already using Hugging Face, Halo offers another way to distribute training while retaining the familiar path for loading trained models.¹²

Researchers at Stanford, Together AI, and Emory University reported on AI agent teams that adapt their collaboration based on previous runs.¹² A team comprising o3-mini, Claude Sonnet 4, and DeepSeek-V3 averaged 66.7% across five math and physics evaluations, compared with 48.8% for its strongest individual model.¹² This is a research result, not an available product.¹²

Rather than prescribing a fixed discussion protocol, the researchers had a model review earlier team exchanges and revise rules for roles, discussion order, and combining partial answers.¹³ That differs from merely selecting among independent answers: a method that chose the correct answer from the individual models’ outputs scored 59.0%.¹³ The evaluations cover math and physics. Whether the learned collaboration helps with business workflows remains an open question.¹²

PrismML released Bonsai 2 27B on September 17, 2026. The downloadable model is based on Qwen3.8 27B: It is a compressed version of an existing model, not a new base model built from scratch.¹² According to the company, it occupies about 5.9 GB, compared with roughly 54 GB for the parent model in FP16. That makes it more than nine times smaller.¹²

In PrismML’s aggregate benchmark evaluation, Bonsai 2 27B scored 83.9, versus 85.4 for the full-precision model. The company describes this as retaining 98.2% of the score.¹² That figure comes from PrismML’s own tests and does not mean performance is nearly identical on every task. The smaller footprint could make local deployment easier for companies, but they still need to test the model against their specific workloads.

Empero released Qwen3.8-35B-A3B-Distill on September 16.¹ The mixture-of-experts model was distilled from Qwen3.8 outputs and targets reasoning, math, coding, and tool use.¹ It has 35 billion parameters in total, with about 3 billion active per token.¹ Empero also provides quantized GGUF files for local inference with llama.cpp.²³

The active-parameter count is more informative about the model’s computation than its total size, but it does not establish a specific speed or memory saving on a given machine.¹ The GGUF files require a recent llama.cpp build that supports Qwen3.6 and Gated DeltaNet MoE; older builds cannot load the architecture.³ Local deployments therefore need to account for runtime compatibility as well as the choice of quantization.³