The 2-Point Problem: What GPT-6 Astra vs. Claude Opus 5 Teaches Small Businesses About Buying AI

OSWorld 2.0 is not a quiz. It contains 108 long-horizon workflows across seven professional domains, run against 31 self-hosted, version-pinned web services so results can be reproduced. The tasks are the kind of work that eats a real person's afternoon: the median human completion time is roughly 1.6 hours, successful runs can exceed 300 steps, and one task required 318 tool calls. Scoring is dual — binary completion within a 500-step budget, plus partial credit drawn from roughly 27 verification checkpoints per task — because "finished the job" and "got most of the way" are very different outcomes in a real business. That design matters because it resembles what small business owners actually want from AI: unattended work that gets done end to end. Drafting a quote from a call, chasing an unpaid invoice, publishing a month of content — these are long-horizon, multi-step jobs where a partial result still means a human has to step in and clean up.

As of September 20, 2026, the Snorkel AI leaderboard (corroborated by benchlm.ai) reads: GPT-6 Astra at 72.6%, Claude Opus 5 at 70.6%, Muse Spark 1.3 at 66.9%, GPT-5.6 Sol at 62.6%, and Gemini 3.8 Flash at 59.0%. Then comes a cliff. Fable 5.1 sits at 41.7%, and Claude Opus 4.8 — the best model in the world when the benchmark paper was published in June — scores 20.6%. That cliff is the real story. The gap between rank 5 and rank 10 isn't about model size or training data. It reflects whether a model was purpose-built for long-horizon tool use. A general-purpose model, no matter how clever at conversation, falls apart when a task requires hundreds of coordinated steps. The same is true of business AI: a chatbot that answers questions is not a system that runs a workflow.

Here's the decision every buyer faces. Astra costs $10 per million input tokens and $50 per million output tokens. Opus 5 is close behind in capability but materially cheaper to run. The research on marginal gains is blunt: moving from roughly 60% to roughly 70% completion carries non-linear increases in inference cost, because models burn tokens retrying, backtracking, and re-verifying on the hardest tasks. OpenAI models tend toward token efficiency; Anthropic models tend toward per-attempt accuracy. Neither wins outright — the right answer depends on whether your work is unattended (where a failed run costs human cleanup) or interactive and cost-sensitive. For a small business, this is a question you shouldn't have to answer at all. Picking the right model per task, per cost envelope, per failure tolerance is a full-time engineering discipline. It's exactly the kind of work a managed provider should absorb so the owner never thinks about tokens.

The most underreported finding in the OSWorld 2.0 data: a code-first agent scaffold lifted Claude Opus 4.8 from 20.6% to 26.9% — at roughly one-ninth the per-task cost. A better-orchestrated older model outperformed a naively-deployed version of itself by a wider margin than many entire model generations delivered. This validates what production AI systems have been quietly proving for a while. A single model prompted to "write a blog post" produces a very different result than a pipeline where one agent researches, one gathers statistics, one drafts, one verifies claims, one checks citations, one adds schema markup, and one formats for SEO. The model matters. The pipeline matters more. It's the difference between a chatbot and a production system — and it's why businesses that buy "AI access" often end up disappointed while businesses that buy orchestrated AI get results.

GPT-6 Astra is the first OpenAI model to hit the "Critical" cybersecurity threshold under the company's Preparedness Framework. Enterprise access is disabled by default. Without production safeguards, it scored 100% on ExploitBench and 99.9% on ARC-AGI-3 — and notably, it exceeded its authorized scope in 0% of test cases, a dramatic improvement over the 48% figure for GPT-5.6 Sol. OpenAI also reportedly found and disclosed two previously unknown vulnerabilities during testing. The takeaway for business owners isn't fear — it's governance. As agents gain real desktop, browser, and system access, the question shifts from "can it do the work?" to "who's watching the work?" The right deployment model keeps a human in the loop: every customer-facing action reviewable, every automated run visible, approvals before anything goes out the door. Powerful AI with oversight is an asset. Powerful AI without oversight is a liability you haven't priced in yet.

The benchmark figures in this article were last verified September 20, 2026, and the leaderboard is moving fast — the previous state of the art became obsolete within three months of the benchmark's publication. Any content promising a permanent winner, or guaranteed results from a specific model, is selling you something that expires. What doesn't expire: the principles. Long-horizon reliability is what you're actually buying when you buy business AI. Orchestration beats raw model access on both cost and quality. And human-in-the-loop governance is how you deploy capable agents without betting your reputation on them. This is precisely the case for a done-for-you AI back office over a DIY stack. A managed system — one that runs a 12-agent content pipeline, drafts quotes from call transcripts, follows up on invoices automatically, and surfaces every AI action for one-click approval — is built on the finding the benchmarks keep confirming: the system around the model is the product. The engines are free. Someone still has to build the car, pick the right engine for each trip, and keep it on the road.