The Shift Toward Open-Source Models and Cheaper Tokens

The speaker walks through several threads from his timeline that point to one trend: developers are moving toward open-source models. He covers a new family of open-source coding models, data showing US model usage dropping on OpenRouter, Chinese resellers offering Claude tokens at steep discounts, and a report that OpenAI plans to restrict access to its next model. Together these are reshaping how teams think about AI token spend.

A New Family of Open-Source Coding Models

The video opens on a tweet from Ornith (@ornith_) announcing Ornith-1.0, a family of open-source LLMs built for agentic coding. According to the post shown on screen, the family spans multiple parameter sizes (9B Dense, 31B Dense, 35B MoE, and 397B MoE), is trained in the US, and is released under an MIT license for commercial and research use.

The on-screen benchmark numbers include:

Benchmark	Ornith-1.0 score
Terminal-Bench 2.1	77.5
SWE-Bench Verified	82.4
SWE-Bench Pro	62.2
SWE-Bench Multilingual	78.9
NL2Repo	48.2
SWE Atlas (QnA)	41.2

A comparison chart titled "LLM Performance Evaluation" shows Ornith-1.0-397B against Qwen3.5-397B, Qwen3.7-Max, GLM-5.2-744B, Minimax-M3-428B, DeepSeek-V4-Pro-1.6T, Claude Opus 4.7, and Claude Opus 4.8. The speaker notes that the usual response to new models is that they only optimize for benchmarks, but he was surprised by how many replies came from people who tested the models and found them genuinely good. He adds that the benchmarks compare well against Opus 4.7 and 4.8.

US Model Token Share Is Dropping on OpenRouter

The speaker then shows a Zero Hedge tweet referencing a Bloomberg chart: "the share of tokens used for US models on OpenRouter has collapsed." The chart, titled "Users Shift to Open-Weight Chinese Models Like DeepSeek," compares two periods:

June 2025: US providers (Google, OpenAI, Anthropic) held roughly 70 percent of tokens.
June 2026: US share fell to around 33 percent, with Chinese models (DeepSeek, Tencent, Xiaomi, Minimax) taking a much larger slice.

He frames this as part of a broader industry shift, with companies looking hard at their AI token spend. In some cases that spend matches the salary of a full developer. He pushes back on the idea that this is justified by productivity gains. The research he references shows most people get somewhat more productive with these tools, but usually not enough to be worth twice an existing developer's cost.

Chinese Resellers Offering Claude Tokens at a Discount

A Hacker News comment on screen explains how Chinese resellers offer Claude tokens at 70 to 90 percent below official Anthropic API prices. According to the comment, they do this by reselling capacity from pooled Claude Max accounts, through payments fraud, and by reselling model output and reasoning chains to Chinese labs as training data, which lets them operate below cost.

The comment adds more context the speaker did not narrate but which appears in the frames: Claude and ChatGPT are both blocked in China, so users rely on a VPN and buy access through resellers. These resellers run tens of thousands of bot accounts, which the comment says is why Anthropic introduced identity verification. The post points to a reseller offering Opus 4.8 at a 93 percent discount and argues this is one reason DeepSeek and GLM are priced so cheaply, since they compete against these low token prices.

OpenAI Restricting Access to Its Next Model

The video closes on a tweet claiming the US government will decide who gets access to GPT-5.6. The post says OpenAI will release GPT-5.6 only in a limited preview to a small group of partners, and that Sam Altman told staff the government would be approving access "customer by customer." It also claims Commerce Secretary Lutnick personally warned Altman not to launch without approvals from other agencies, describing it as a de facto licensing regime. The speaker calls this development insane and says he will make a separate video about it.

Key Takeaways

A new open-source coding model family, Ornith-1.0, posts benchmark scores competitive with Claude Opus 4.7 and 4.8 and ships under an MIT license.
Bloomberg data shown in the video puts the US share of OpenRouter tokens at around 70 percent in June 2025 and roughly 33 percent in June 2026, with Chinese open-weight models taking the difference.
Teams are scrutinizing AI token spend that in some cases rivals a developer's salary, and the speaker argues productivity gains rarely justify doubling that cost.
Chinese resellers sell Claude tokens at 70 to 90 percent below Anthropic's API prices using pooled accounts, payments fraud, and resale of model output.
A tweet shown in the video claims GPT-5.6 access will be approved customer by customer under government oversight.

Resources

Ornith (@ornith_) - account announcing the Ornith-1.0 open-source coding models
Zero Hedge (@zerohedge) - posted the Bloomberg chart on US model token share
OpenRouter - model routing platform whose usage data is cited in the Bloomberg chart