Poetiq's self-improving harness pushes GPT-5.5 to 93.9% on coding benchmark

Most AI labs throw millions of dollars at post-training and RL fine-tuning just to squeeze another point or two out of their frontier models. Startup Poetiq just proved you might not need to touch the weights at all.

Poetiq revealed benchmark results showing its Meta-System propelled OpenAI's GPT-5.5 to a record-breaking 93.9% on LiveCodeBench Pro (LCB Pro 25Q2). It did so without any model fine-tuning, weight modifications, or private activation access—just standard API calls.

LiveCodeBench Pro is known as one of the hardest coding test suites out there, specifically designed to dodge LLM contamination through continuous updates and hidden competitive C++ challenges. Achieving 93.9% puts Poetiq comfortably in SOTA territory.

An AI that codes its own scaffolding

Poetiq's approach relies on recursive self-improvement. Instead of hand-crafting prompts or agent workflows, the Meta-System builds and optimizes its own execution harness from scratch.

The system was initially trained and optimized using only Google's Gemini 3.1 Pro. By learning how to refine sequential chains of questions and assemble responses under strict memory and time limits, it boosted Gemini 3.1 Pro by 12.3 percentage points, from 78.6% to 90.9%. That was enough to eclipse Google's own unreleased flagship, Gemini 3 Deep Think, which sits at 88.8%.

What makes the setup genuinely wild is its model-agnostic nature. Poetiq took that exact Gemini-tailored harness and slapped it onto GPT-5.5 without any adjustments. The result: GPT-5.5 jumped from an already high baseline of 89.6% to 93.9%. On the hardest tier of competitive programming questions, its success rate shot up from 50.0% to 75.0%.

The gains carried over across the board. The smaller Gemini 3 Flash leaped 10 percentage points from 72.3% to 82.3%, outperforming raw runs of Claude Opus 4.7 (80.5%) and GPT 5.2 High. Kimi K2.6 saw an eye-popping 30% performance jump, while Nemotron climbed 12.8%.

The hype versus vaporware skepticism

Despite the mind-bending numbers, the AI community is responding with a heavy dose of skepticism.

Poetiq was founded by a team claiming 72 combined years of Google and DeepMind experience, emerging from stealth in November 2025 after prior stunts on ARC-AGI and Humanity's Last Exam (HLE). Yet, nobody outside the lab has touched the actual tech.

On Reddit, commenters have started calling Poetiq vaporware. Waitlist applicants report receiving hype-filled marketing emails while the company remains private, with no sandbox, no public repo, and no production timeline in sight. Poetiq says it is working with a small group of early enterprise customers, but independent verification remains non-existent.

Why it matters

AlphaGo reached superhuman level through self-play, and Poetiq is applying that identical philosophy to the scaffolding that wraps around language models. The 13-point chasm between Poetiq-augmented GPT-5.5 and standard Claude Opus 4.7 is too massive to dismiss as benchmark noise.

If this works in the wild, it fundamentally threatens the traditional post-training pipeline. Enterprises spend small fortunes trying to fine-tune custom weights for coding workflows. If an external, automated orchestration harness can produce bigger gains across any off-the-shelf model via standard APIs, the value flips from training proprietary models to controlling the execution layer.

It could also radically change developer economics. A cheap, fast model like Gemini 3 Flash beating a raw Claude Opus 4.7 means developers could slash inference bills without sacrificing complex problem-solving abilities.

That said, take the enthusiasm with a massive grain of salt. Without public access, we don't know how many recursive API loops or latency penalties are baked into these numbers. If solving a hard competitive problem requires fifty nested roundtrips, it might win on a leaderboard but crash in a real-world IDE. Until Poetiq ships an actual product, this remains a brilliant research teaser rather than a practical developer tool.