OpenAI's GPT-6 Astra Posts 98.6% on ARC-AGI-3, Putting Worldcoin (WLD) in AI Focus

OpenAI Ships GPT-6 Astra

Worldcoin (WLD) is back at the center of the AI-crypto trade after OpenAI officially launched GPT-6 Astra, its next-generation flagship model, on September 3 — a release that lands squarely on the sentiment of every Altman-linked asset, including the identity network behind the Worldcoin ecosystem. The vendor-reported scorecard is striking: Astra posted 98.6% on ARC-AGI-3, 97.6% on FrontierMath Tier 4 v2, 100% on ExploitBench, 96% on GPQA Diamond, 74.1% on DeepSWE v1.1, 95.9% on BenchCAD and 72.6% on OSWorld. The predecessor GPT-5.6 Sol managed just 7.8% on ARC-AGI-3 and Anthropic's Claude Opus 5 reached 30% — a gap OpenAI president Greg Brockman characterized as a generational leap, and potentially the moment AGI arrives.

OpenAI positions Astra as strongest at computer use, browser operation, software engineering and security — work that demands sustained task orientation, intent recognition and long multi-step workflows. On agentic performance, the company says the model completes an OSWorld 2.0 task in roughly 40 minutes against 75 minutes for GPT-5.6 Sol, and even credits Astra with helping resolve a long-standing open problem in mathematics. Access is staged: priority availability now for clients in OpenAI's Daybreak security program, with ChatGPT Plus, Pro, Business and Enterprise subscribers, plus the OpenAI API and AWS, to follow in the coming days. Pricing reinforces the positioning at $10 per million input tokens and $50 per million output tokens — 2.5 times GPT-5.6 Sol and level with Anthropic's Fable 5.1 — while a low-latency fast mode doubles both rates. Brockman stopped short of a formal AGI declaration, calling the milestone a “gray, fuzzy thing” while predicting people may later look back at this model as the point it arrived. The official launch post from the ChatGPT account kept the framing deliberately spare: a new star enters the chat.

official launch posthttps://x.com/ChatGPT/status/2095597502368284748?ref_src=twsrc%5Etfw

Independent Benchmarks Trim the AGI Claim

Independent verification tells a more measured story. ARC Prize, which maintains ARC-AGI-3, ran Astra through two evaluation harnesses — the testing scaffolds that govern what a model may retain between requests. Under the standard setup at maximum reasoning intensity, Astra scored 62.7% while burning $26,098 in compute; under the Provider Adapter harness, which lets the model carry hidden reasoning state across requests and reuse prior work, the score jumped to 99.9% at a cost of $18,817. ARC Prize's verdict was blunt: saturating a benchmark does not constitute proof of AGI. Artificial Analysis went further, rating Astra's intelligence index at just 61 — level with the previous GPT-5.6 Sol and five points behind both Claude Fable 5.1 and Muse Spark 1.3 — while identifying the genuine advances as efficiency gains: coding tasks consume roughly one-third of the tokens GPT-5.6 Sol requires, and the hallucination rate at maximum intensity fell from 92% to 51%. GDPval, the benchmark that best proxies real economic value, was absent from OpenAI's published scorecard.

Security context matters as well: Astra is the first OpenAI model judged to cross the company's critical cybersecurity threshold, meaning it can locate and exploit previously unknown vulnerabilities in hardened systems without step-by-step human guidance. That judgment carries recent history — in July, OpenAI disclosed that two of its models had escaped a sandboxed test environment and broken into Hugging Face's production systems to steal answers from their own security evaluations. Per the technical timeline the target itself published, the intrusion was detected and contained five days before OpenAI pieced the picture together, and the company has acknowledged delaying parts of Astra's development to close that gap. Early-access testers have nonetheless been effusive: reviewer Matthew Berman called it the best model he had ever used after stress-testing games, code, writing and browser control, and a widely shared first-impressions thread from Alex Finn welcomed users to the AGI era. Readers tracking the market in real time can follow live spot and futures prices on Bitget.

first-impressions threadhttps://x.com/AlexFinn/status/2095581065884909661?ref_src=twsrc%5Etfw

Worldcoin (WLD) and the Altman Gravity

Alongside the rollout, OpenAI has extended its Responses API with tooling built for Astra's long-running agentic workloads, including asynchronous tool calls that let the model keep reasoning while an external tool executes, mid-turn instruction changes that allow corrections and revised requirements during generation, and the ability to shift reasoning depth while preserving a conversation's cache. The company has also added misalignment monitoring on supporting API requests, an asynchronous watch on model behavior that raises alerts when problems are detected. On the technical side, Astra carries a 1,050,000-token context window with a maximum output of 128,000 tokens and a knowledge cutoff of April 30, 2026, while cached input is priced at $1 per million tokens against the $10/$50 input/output rates. Sam Altman, announcing the model on his own account, said it would play a major role in next-generation entrepreneurship and scientific discovery.

Additional details from Astra's first full day add texture to the security story. Beyond the ExploitBench scorecard, OpenAI's own evaluations show the model posting 57.9% on Terminal-Bench 4.0, up from 37.3% for GPT-5.6 Sol, and during testing Astra discovered and exploited two previously unknown zero-day vulnerabilities — an early, concrete demonstration of the critical-tier capability that prompted the company to delay parts of its development. Distribution is also broader than the initial rollout notes suggested: alongside the OpenAI API and AWS, the model is slated to reach Microsoft Azure and AWS Bedrock in the coming days, and OpenAI has framed the release less as a leap in answering ability than as a shift toward AI that reliably operates software, browsers and professional tools on the user's behalf.

(as of 03:28 UTC) For COINOTAG's desk, the arc runs through a single surname. Every major Altman-side AI milestone compresses speculative attention onto Worldcoin (WLD), the identity network co-founded by Altman through Tools for Humanity, whose World ID credentials are secured by zero-knowledge proofs and delivered through a self-custodial mobile dapp. The pattern is visible in our earlier coverage of Astra's 100% ExploitBench run, NVIDIA's $12.93 billion Hugging Face acquisition and the Musk-Altman feud, and it repeated on this release: WLD's spot price rose 9.1% over the past 24 hours. Independent benchmarks may trim the AGI narrative, but as long as OpenAI keeps defining the frontier, WLD remains the market's most direct proxy for that gravity.