GPT-5.6 fully launched, new ChatGPT agent released simultaneously

Author: Yang Chen, Wall Street See

OpenAI officially launches the GPT-5.6 model series, delivering stronger performance at lower costs, sending a clear signal to the competitive landscape of the AI industry.

This release includes three models: the flagship Sol, the balanced Terra, and the cost-effective Luna — available globally from today via ChatGPT, Codex, and the OpenAI API.

In terms of pricing, Sol charges $5 per million input tokens/$30 per million output tokens, Terra $2.50/$15, and Luna $1/$6.

OpenAI emphasizes that GPT-5.6 surpasses Anthropic's Claude Fable 5 on multiple key benchmarks, while significantly reducing token consumption and inference costs. This means users can complete more substantive work within the same budget.

For enterprise users and developers, the core impact of this release lies in a comprehensive improvement of "cost-performance." In the Agents' Last Exam evaluation tailored for professional workflows, GPT-5.6 Sol scored 53.6 points, exceeding Claude Fable 5 by 13.1 percentage points. Even with moderate inference settings, its cost is about one-quarter that of Fable 5.

More striking is that the lower-tier Terra and Luna still outperform Fable 5 on benchmarks at roughly one-sixteenth the cost. This "overwhelming" pricing strategy directly compresses the differentiation space for competitors.

At the GPT-5.6 model release milestone, OpenAI has simultaneously advanced the integration of multiple product lines: officially launching the ChatGPT Work agent, which supports cross-application continuous work and delivers results directly, marking the evolution of the product form from conversational tool to system-level agent; merging the Codex App into the ChatGPT desktop application, available in the free tier; and gradually phasing out the standalone Atlas browser, with browser capabilities shifting to a Chrome extension.

Altman noted in an interview that OpenAI "made many adjustments" during communications with the U.S. government but did not specify the details. He also mentioned that the latest AI model achieves a 54% improvement in token efficiency for agent programming, stated he does not know whether there will be an IPO this year, and acknowledged that Chinese open-source models are becoming "very good."

ChatGPT Work agent launched, supporting cross-application continuous work, powered by GPT-5.6

OpenAI also launched a new AI agent today — ChatGPT Work. The product is positioned as a fusion of ChatGPT and Codex, aiming to allow ordinary non-technical users to leverage Codex's capabilities for non-programming tasks, driven by the GPT-5.6 model family (Sol, Terra, and Luna).

OpenAI stated in a blog post that ChatGPT Work can capture context from user-authorized apps, files, and workflows, and automatically generate complete outputs including documents, spreadsheets, presentations, and web applications.

OpenAI also said that ChatGPT Work is equipped with a Unified Plugins Directory, connecting to tools such as Slack, Gmail, Google Drive, calendar, and CRM (customer relationship management) to enable cross-application collaborative work.

OpenAI and its competitor Anthropic PBC are accelerating the race toward more advanced AI agents, aiming to expand AI capabilities beyond automatic code writing and debugging into broader office scenarios.

Earlier this year, Anthropic released a similar product, Claude Cowork, also targeting more enterprise and professional users.

Altman: Did not specify the "many adjustments," does not know about IPO this year

OpenAI CEO Altman revealed that before GPT-5.6 was released to the public, the company engaged in in-depth consultations with Trump administration officials and made "many adjustments" during the process.

Altman did not specify what changes OpenAI made during the consultations, but noted that the government is testing the new model and checking for potential risks. He generally views the mechanism positively but set conditions.

"As long as the process is understandable, fair, and efficient, I think it's a good thing. The government seems to agree with these goals. I estimate that the process for our next model will be much smoother — we already understand this mechanism and know how to cooperate better."

Altman also responded to two recent market concerns:

Regarding earlier reports that OpenAI planned to offer the U.S. government a 5% equity stake, Altman clearly refuted this, saying the reports contained "a lot of inaccuracies," but did not elaborate further. On the IPO issue, when asked whether OpenAI plans to go public this year, Altman simply said: "I don't know."

Regarding the newly released GPT-5.6, Altman pointed out that the latest AI model achieves a 54% improvement in token efficiency for agent programming. He also stated that Chinese open-source models are becoming 'very good.'

Flagship model Sol sets new efficiency benchmark

The improvement in coding ability for GPT-5.6 Sol is particularly notable. In the Artificial Analysis Coding Agent Index test, Sol (max reasoning setting) set a new record of 80 points, 2.8 points higher than Fable 5, while outputting less than half the tokens, taking less than half the time, and costing about one-third less.

In the complex command-line workflow test Terminal-Bench 2.1, Sol scored 88.8%, leading the field. In the long-cycle engineering test DeepSWE v1.1, which targets real-world codebases, it scored 72.7%, also at the forefront of the industry.

Real-world data from enterprise users confirms the above performance. Itamar Friedman, co-founder and CEO of the code review platform Qodo, stated that GPT-5.6 surpasses GPT-5.5 in both internal and external benchmarks, and the number of tokens required per code review is reduced by about two-thirds, with median latency reduced by about 50%.

Fabian Hedin, co-founder of AI development platform Lovable, noted:

"With GPT-5.6, users need about 25% fewer steps to complete tasks, tool call frequency drops 35% to 48%, and project failure rates decrease by 15%."

Multi-agent architecture enables 'ultra' mode

GPT-5.6 introduces a hierarchical computation scheduling mechanism, allowing users to flexibly choose inference depth based on task requirements.

In addition to the standard high reasoning setting, OpenAI adds "max" and "ultra" modes: max gives the model more time for thinking and solution correction; ultra coordinates four parallel agents by default, sacrificing higher token consumption for stronger results and shorter delivery times.

In the BrowseComp test, Sol Ultra set a new record with a score of 92.2%. In OSWorld 2.0, it outperformed Claude Opus 4.8 with a score of 62.6%, while the latter used 85% more output tokens. In SEC-Bench Pro (vulnerability proof-of-concept generation for complex software), Sol scored 71.2%, compared to GPT-5.5's 45.8%.

At the API level, OpenAI simultaneously launched Programmatic Tool Calling, allowing GPT-5.6 to write and run lightweight programs in memory, autonomously coordinating tools, handling intermediate results, and dynamically adjusting workflows during execution. This reduces round-trip calls between the model and tools while maintaining compatibility with the Zero Data Retention (ZDR) policy.

Cybersecurity and scientific research capabilities significantly improved

In cybersecurity, GPT-5.6 Sol scored 73.5% on the ExploitBench evaluation, compared to GPT-5.5's 47.9%. In ExploitGym, its pass rate within a six-hour limit increased from 15.1% for GPT-5.5 to 33.7%.

OpenAI stated that the model also has stronger capabilities on the defensive side, including secure code review, vulnerability patching, threat modeling, and blue-team exercises.

To manage dual-use risks, OpenAI, through its "OpenAI Daybreak Trusted Access for Cyber" program, offers higher-level defensive features to verified individuals and organizations, with additional access restrictions for high-risk entities and regions.

In life sciences, GPT-5.6 Sol scored 28.7% on GeneBench Pro (long-cycle genomics and quantitative biology), far exceeding GPT-5.5's 12%. On LifeSciBench, it scored 59.9%, higher than GPT-5.5's 50.4%.

OpenAI also stated that its tests show GPT-5.6 is not yet capable of end-to-end creation or synthesis of high-risk novel biological threats.

Safety mechanisms upgraded simultaneously, conservative deployment strategy continues

As model capabilities increase, OpenAI has strengthened its safety architecture with this release. The cybersecurity protections of GPT-5.6 Sol block approximately 10 times more potentially harmful activities than the previous generation.

Prior to the official release, OpenAI conducted approximately 700k A100e GPU hours of black-box automated red-teaming, along with large-scale manual and automated dual-track evaluations with external experts.

OpenAI stated that its safety system employs multi-layered redundancy: built-in model protections and real-time checks, continuous monitoring, and account-level enforcement work together.

Unlike pure classifier-based blocking, GPT-5.6 introduces a 'reasoning monitor,' which analyzes potential harm on a per-conversation basis, allowing safety boundaries to dynamically adjust based on usage scenarios and enabling rapid patching of vulnerabilities without retraining the classifier.

To address the impact of over-blocking on legitimate use, OpenAI offers a one-click option in ChatGPT and Codex to retry with a lower-capability model, and says it will continue to optimize, reducing false positives on normal usage while maintaining high interception intensity.

Internal applications reveal AI self-acceleration trend

The release of GPT-5.6 also reveals significant changes in OpenAI's own R&D model. According to OpenAI, during internal testing of GPT-5.6, the average daily output tokens per active researcher more than doubled the historical peak of GPT-5.5.

Over the past six months, OpenAI's internal computational power for programming inference has increased 100-fold, and internal agent token usage has grown about 22-fold. This trend has extended to sales, marketing, user operations, and finance functions.

In the Self-Improvement Capability Evaluation (RSI Index), GPT-5.6 Sol scored 57.9, leading GPT-5.5's 41.7, an improvement of 16.2 percentage points.

OpenAI interprets this as direct evidence that AI-assisted AI R&D is accelerating and positions it as a core driver for the efficiency improvement of subsequent model iterations.