OpenAI Releases GPT-5.6 Family: Sol, Terra, Luna

Flagship Sol leads coding, knowledge work, cybersecurity, and science benchmarks at lower token cost. Terra balances performance and affordability. Luna delivers speed for high-volume tasks.

OpenAI officially launched the GPT-5.6 family on July 9, introducing three distinct models under the names Sol, Terra, and Luna. Sol is the flagship, engineered for long-horizon coding and agentic work that demands planning, tool use, and follow-through. On the long-term agentic benchmark, Sol scored 53.6, surpassing Claude Fable 5 at 40.5 — and at lower cost. Terra is positioned as the balanced workhorse, delivering performance competitive with GPT-5.5 at a reduced price point. Luna rounds out the trio as the fastest and most affordable model, optimized for well-defined, high-volume workloads. All three are available immediately in ChatGPT, Codex, and the API, with pricing at $5/30, $2.5/15, and $1/6 per million input/output tokens respectively. Context windows extend to one million tokens with maximum output of 128k tokens.

NVIDIA confirmed that GPT-5.6 was trained and runs on GB200 and GB300 NVL72 rack-scale systems. CEO Sam Altman called it "obviously the best model we have ever produced" and emphasized that the per-task dollar metrics represent a major step forward for enterprise customers concerned about AI costs. OpenAI president Greg Brockman praised Sol's performance in coding, knowledge work, cybersecurity, and science, noting that lower token costs make the model accessible to a wider range of applications. The release comes alongside a suite of product updates including the merger of Codex and ChatGPT into a unified desktop application, the launch of ChatGPT Work as a new agent experience, and a new Chrome extension with revamped in-app browser.

Meta Releases Muse Spark 1.1, Nears Frontier Performance

Meta launched Muse Spark 1.1, a major upgrade from the original Muse Spark, matching GPT-5.5 and Opus 4.8 on agent, coding, and multimodal benchmarks. The model ranks fourth on the overall Vals index and is the fastest model in the top ten. CEO Mark Zuckerberg personally announced the release, highlighting its competitiveness in agent and coding tasks. Meta researcher @hwchung27 noted that the API is "really, really cheap" and designed to open agentic intelligence to more people. Scale AI CEO Alexandr Wang independently confirmed that the model rivals GPT-5.5 and Opus 4.8 across many agentic evaluations. "The rate at which intelligence is getting cheaper still feels underrated," the Meta researcher added. Alongside the release, Meta opened a public beta for the new Meta Model API, giving developers direct access to Muse Spark 1.1.

Codex and ChatGPT Merge into Unified Desktop App

New coding workflows, Chrome extension, revamped in-app browser, and faster Computer Use powered by GPT-5.6.

OpenAI merged the powerful coding agent Codex with ChatGPT Work into a single desktop application, now available on mobile, web, and desktop. OpenAI president Greg Brockman described ChatGPT Work as "an agent for your most ambitious work," usable across all platforms without needing a laptop left running. The unified app introduces new coding workflows, a Chrome extension, a revamped in-app browser, and faster Computer Use capabilities. The same agent experience powers both coding tasks and general knowledge work within a single interface.

Ollama Raises $65M, Nears 9M Users

Open-source AI developer tool Ollama announced $65 million in funding from Benchmark, with active users approaching 9 million and over 176,000 GitHub stars. The round was reported by TechCrunch. The company's co-founder stated that their belief in open-source AI "has never wavered" and promised faster open-model expansion. With nearly 9 million active builders, Ollama is positioning itself at the center of the open-source AI ecosystem. Co-founder Michael Chiang and Benchmark partner Peter Fenton appeared on CNBC to discuss the "post-frontier era" and open-source AI's breakout moment.

Simon Willison Details GPT-5.6 API: Programmatic Tool Calling and Multi-Agent

The veteran developer analyzes key new API features including six reasoning levels and three model tiers.

Developer Simon Willison published a detailed analysis of the GPT-5.6 API, highlighting several notable additions. Programmatic tool calling enables code to orchestrate multi-step tool sequences without manual chaining. Multi-agent support allows multiple model instances to collaborate within a single API session — a feature once exclusive to frameworks like AutoGen. Willison noted six distinct reasoning levels corresponding to the three models, joking about 18 pelicans representing the tier structure. Pricing ranges from $1/6 per million tokens for Luna to $5/30 for Sol. The context window reaches one million tokens with maximum output of 128k tokens. Willison pointed out that Sol scored 53.6 on the long-term agentic benchmark versus Claude Fable 5 at 40.5, though SWE-Bench Pro remains at 64.6%, leaving room for improvement in pure coding tasks.

Ethan Mollick Uses Codex + GPT-5.6 Sol to Review Book Manuscript

Professor Ethan Mollick shared a real-world use case: submitting an entire new book PDF to Codex powered by GPT-5.6 Sol. The AI processed the manuscript in 30 minutes and returned dozens of review notes. "None were wrong, but plenty were pedantic," Mollick noted, demonstrating the model's capability to handle long-form document analysis and provide actionable editorial feedback. The case highlights how GPT-5.6 Sol's expanded context window and improved reasoning make it viable for professional knowledge work that previously required human experts.

Analyst: xAI-Cursor Alliance Shakes Confidence in Anthropic's Lead

Grok 4.5 reaches Opus 4.7 level on CritPt, surpassed only by GLM 5.2 among Chinese models.

AI industry analyst @teortaxesTex published a widely discussed thread arguing that the Grok 4.5 and Cursor collaboration has shaken confidence in Anthropic's lead. "I thought Anthropic's edge was deeper at this point," the analyst wrote, noting that Grok 4.5 is "between Opus 4.6-7 and 4.8 on CritPt" and ahead of all Chinese models except GLM 5.2. Cursor's coding data accelerated xAI's capabilities disproportionately in the coding domain. The broader implication: "If you have GPUs, you can still catch up" to the frontier, suggesting that compute access remains a viable path to competitive model performance even for late entrants. Meta's Muse Spark 1.1 joins the evidence, showing that multiple labs are converging on frontier-level capabilities simultaneously.

Former Fed Chair Bernanke Joins Anthropic Long-Term Benefit Trust

Anthropic appointed former Federal Reserve Chairman Dr. Ben Bernanke as the newest member of its Long-Term Benefit Trust. The LTBT is designed to oversee long-term stakeholder decisions, ensuring that Anthropic's governance considers broad societal impacts as the company scales frontier AI development. Bernanke's appointment brings high-level economic policy experience to the AI safety governance framework.

MOSS Releases End-to-End Multi-Speaker Transcription Model

MosiAI open-sourced MOSS-Transcribe-Diarize-0.9B, a 0.9-billion-parameter model that integrates ASR, speaker identification, and alignment into a single end-to-end pipeline. Most existing setups chain separate WhisperX-style ASR, diarization, and alignment components. This model collapses all three into a unified architecture with a 128k long-context transcription window under Apache 2.0 license. vLLM provided day-one support for the model, reinforcing the open inference ecosystem's rapid integration capability.