Dear OpenAI GPT Development Team,
GPT-5.6 demonstrates strong competitive advantages over Fable 5 in long-horizon agentic workflows, overall coding-agent performance, execution speed, and cost efficiency. However, a detailed examination of publicly available evaluations indicates that several areas still require further improvement.
First, GPT-5.6 needs stronger analytical completeness and more reliable fulfillment of task-specific evaluation criteria in professional knowledge work. In Artificial Analysis’s AA-Briefcase evaluation, GPT-5.6 Sol achieved the highest presentation-quality score, but its rubric-completion score was 42%, compared with 56% for Fable 5. Its Analytical Quality Elo was also lower, at 1,592 versus 1,764. In OpenAI’s published GDPval-AA v2 results, GPT-5.6 Sol scored 1,747.8 Elo, while Fable 5 scored 1,759.6 Elo. This suggests that although GPT-5.6 is highly capable of producing visually polished outputs, it still has room to improve in translating complex requirements into complete deliverables without omissions.
Second, GPT-5.6 requires greater accuracy when modifying large, real-world codebases. In OpenAI’s published results, GPT-5.6 Sol scored 64.6% on SWE-Bench Pro, substantially below Fable 5’s 80.0%. At the same time, GPT-5.6 leads the broader Artificial Analysis Coding Agent Index. This indicates that the weakness is not general coding capability, but the reliable completion of repository-level changes involving multiple files, dependencies, tests, and interacting systems.
Third, further improvement is needed in frontier-level reasoning and complex tool use. GPT-5.6 Sol scored 83.0% on FrontierMath Tier 4, compared with 87.8% for Fable 5, and 58.0% on Toolathlon, compared with 61.7% for Fable 5. The independent Artificial Analysis Intelligence Index also placed Fable 5 at approximately 60 points and GPT-5.6 Sol Max at approximately 59 points. Although these gaps are relatively small, they are meaningful in high-stakes mathematical, scientific, research, and multi-tool workflows where a single logical or procedural error can invalidate the entire result.
Fourth, GPT-5.6 should further strengthen the consistency of its closed-loop visual verification process. Both Fable 5 and GPT-5.6 support advanced vision, computer-use, and rendered-output inspection. OpenAI specifically states that GPT-5.6 can inspect and refine rendered results rather than merely generating underlying code or content. The remaining challenge is therefore not the absence of visual verification, but ensuring that generation, rendering, comparison, defect detection, and correction operate as a dependable and repeatable end-to-end loop across documents, presentations, interfaces, charts, and graphical assets.
To address these limitations, I propose the following four improvements.
-
Introduce a “Rubric Compiler” that converts user requirements into atomic, testable acceptance criteria before generation, and then verifies each criterion against explicit evidence after completion.
-
For complex assignments, separate the execution process from an independent critic or verification process that adversarially reviews numerical accuracy, logical consistency, code correctness, source validity, and omitted requirements.
-
In coding workflows, strengthen a Repository-Level Repair Loop that repeatedly validates repository-wide dependencies, tests, builds, runtime behavior, and rendered interfaces rather than evaluating patches only at the individual-file level.
-
For documents, interfaces, and graphical work, establish a default Visual Verification Loop consisting of generation, rendering, screenshot or reference comparison, layout and content-difference detection, and automated correction.
GPT-5.6’s principal competitive advantage is its ability to deliver near-frontier intelligence at substantially lower cost and higher speed. The next stage of development should therefore focus not merely on raising benchmark scores, but on building a system that consistently captures every requirement, verifies its own work during long-running tasks, corrects failures autonomously, and provides evidence that its final output is accurate and complete.