Is OpenCode Go a scam? (Stop Using)

Item: Is OpenCode Go a scam? (Stop Using)
Rating: 5
Author: PostSingularity Institute

Title history

Is OpenCode Go a scam? (Stop Using) — observed on 20261104-1553 UTC (YouTube)

by Kai · View the channel on PSI

Humanities, Social Sciences & Thought Computing & Cybersecurity URComputer security URHComputer fraud and hacking

🎙 Kai 👥 68K 📅 November 4, 2026 ⏱ 29 min 👁 595 📄 original study 🧭 2026-11-04

Available in: English (current) Français

Keywords

OpenCode GoDeepSeekAI codingscam analysisbenchmark

Summary

The video investigates whether the OpenCode Go subscription service is a scam. It explains that the $10 plan provides a $100 monthly allowance for DeepSeek V4.1 Flash, billed at DeepSeek’s own token prices. Critics, notably YouTube creator Em Dash, argue it feels like a rugpull due to session failures and high costs. However, the analysis shows that the model weights are not secretly quantized (DeepSeek already ships 4-bit experts), and benchmark scores are nearly identical on average. The real issue lies in the routing: Go uses eight different configurations, with only 0.1% traffic going to DeepSeek’s official API. Two configurations, handling 12% of requests, fail the 100k-context test, causing long-session breakdowns. Speed tests reveal that Go took 4.5 times longer due to occasional multi-minute stalls, but output token counts were nearly equal. Budget consumption is higher because of lower cache-hit rates (96% vs 99.4%) and a greater reasoning-token share (51.2% vs 39.2%), making sessions cost up to 55% more of the allowance. Yet even in worst-case runs, Go still delivers 6.5 times more work per dollar than the DeepSeek API. The business model relies on average usage, with bulk discounts and underutilization. Communication issues (removed provider names, budget changes, a breach) add mistrust but are not fraudulent. The verdict: Go is a discount service, not a scam, but it sacrifices predictability; heavy agentic users should prefer the direct API.

229 words

Critical Evaluation

Value of the Information & Strength of the Argument

The value lies in deconstructing the ‘scam’ narrative with precise numbers and separating average performance from tail behavior. The argumentation is solid: it presents both sides, uses multiple independent testers (nothingto12, Em Dash), and explains statistical phenomena like the mismatch between P95 and average latency. It also quantifies hidden costs through cache misses and reasoning tokens, offering actionable advice. The reasoning is logical and transparent, though some data points come from unverified X posts, which is acknowledged.

Scientific Rigor, Source Quality, Title Accuracy

The video rigorously cites official documentation, pricing pages, GitHub issues, and third-party benchmark posts. It distinguishes between verified facts and speculations, and even corrects a previous error by Em Dash regarding quantization. The title is appropriate and directly addressed: the entire video is an evidence-based answer. The sources are relevant and diverse, though some are ephemeral social media posts with limited traceability.

154 words

Title / Content Match

The title poses a direct question that the video answers with quantitative evidence, fully addressing the controversy.

Quality & Reliability

8/10

The investigation combines official documentation, third-party benchmarks, and personal tests, openly discussing limitations and correcting earlier claims. However, reliance on unverifiable X posts and the absence of an independent audit cap the score.

Chapters

Cited Sources

DeepSeek API Pricing — Official token prices used to calculate the $100 allowance and compare costs.
OpenCode Security Breach Coverage — Report of the October breach affecting 1.4% of users.
Self-hosted vs proxied discussion — GitHub issue discussing OpenCode Go's infrastructure and provider routing.
Provider attribution removed issue — Issue about the removal of provider names from the service.
Hacker News thread on OpenCode Go — Community discussion expressing distrust and complaints about the service.
OpenCode Go Documentation — Official docs for limits, prices, estimates, and client compatibility.
OpenCode Go Plans — Pricing and subscription details for the service.
301,018 Active Subscribers Article — Confirms the subscriber count used to estimate revenue.
Em Dash - Ending my OpenCode Go Subscription — Video critical of OpenCode Go, cited as part of the scam argument.
Em Dash - Is OpenCode Go a scam? — Original video calling out perceived issues; later partially corrected.

Concurring Sources

Dax Raad's benchmark chart with Traffic column — Shows near-identical average scores across configurations and the traffic split.
nothingto12's SWE-bench Verified test — Indicates Go solves 281/300 vs DeepSeek's 283/300, supporting equivalence.
DeepSeek API pricing page — Provides the official price card used to calculate the $100 allowance value.

Dissenting Sources

Em Dash - Ending my OpenCode Go Subscription
Hacker News thread

Contribution & Novelties

The video provides a decisive, data-driven answer to a widely discussed controversy by separating the quantitative reality from the qualitative experience. It introduces the concept of the ’tail’ in AI service routing, explaining why average benchmarks hide catastrophic user experiences. It also quantifies the impact of cache misses and reasoning-token share on subscription budgets, which is rarely done publicly.

Pour aller plus loin :

Mixture of Experts — The architecture behind DeepSeek V4.1 Flash, relevant to understanding why expert weights are 4-bit quantized.
Quantization (signal processing) — General explanation of weight compression, including 4-bit quantization.
Prompt caching — While not an exact page on prompt caching, the concept of cache hits and their cost impact is essential to grasp the budget dynamics.

122 words

Radar Profile

The profile shows high scores in information quantity and quality, moderate technical depth, and a balanced reliability score. The strength is the dense, well-sourced analysis; the weakness is the reliance on third-party social media posts that are not formally archived.

Reliability 8/10