GPT-5.6 reaches ChatGPT, Codex and API with an 80.0 agent score
Original: GPT-5.6 rolls into ChatGPT, Codex and API with 80.0 coding-agent score View original →
GPT-5.6 moves from preview into production surfaces
OpenAI has shifted GPT-5.6 from a watched preview into a rollout across the products developers and enterprise users actually touch. The official OpenAI account said Sol, Terra and Luna, the GPT-5.6 model family, are beginning to appear in ChatGPT, Codex and the API. A later post in the same thread says Plus, Pro, Business and Enterprise users can reach GPT-5.6 Sol through medium and higher effort settings, while Pro and Enterprise can choose GPT-5.6 Pro for more complex work.
"starting to roll out now in ChatGPT, Codex, and the API"
The thread matters because it joins access details with direct benchmark claims. OpenAI says GPT-5.6 Sol reaches 53.6 on Agents' Last Exam, 13.1 points above Claude Fable 5 adaptive. It also says medium reasoning beats Fable 5 by 11.4 points at about one-quarter of the estimated cost, while GPT-5.6 Terra and Luna outperform Fable 5 at roughly one-sixteenth of the cost. On the Artificial Analysis Coding Agent Index, OpenAI places GPT-5.6 Sol at 80.0, 2.8 points ahead of Claude Fable 5, with less than half the output tokens and time, and about one-third lower cost.
OpenAI’s account is the company’s primary channel for product and model rollouts, so the post should be read as an access and positioning signal, not just a benchmark graphic. GPT-5.6 is being framed for three usage patterns at once: conversational work in ChatGPT, coding-agent work in Codex, and developer integration through the API. The thread also points to ultra mode, where multiple agents coordinate in parallel and trade higher token use for stronger results on demanding tasks.
The next test is whether independent evaluators reproduce the 80.0 score and whether the cost claims survive real workloads. If they do, the competitive comparison for coding agents shifts from raw model quality to quality per token, latency and controllable effort settings.
Related Articles
OpenAI has put a public July 9 launch window on three GPT-5.6 models, shifting attention from one flagship model to a tiered lineup. The source tweet has passed 3.6 million views and says global preview access is expanding now.
OpenAI is pushing harder into agentic work, not just chat. On the company's own evals, GPT-5.5 reaches 82.7% on Terminal-Bench 2.0, beats GPT-5.4 by 7.6 points, and uses fewer tokens in Codex.
OpenAI says SWE-Bench Pro no longer reliably measures frontier coding capability after finding 30% of its public tasks broken. The cited issues include hidden requirements, contradictory instructions, strict tests and incomplete grading criteria.