AI News Today: Sonnet 5, Export Rules, Black Market
Claude Sonnet 5 launched at roughly half the price of Opus 4.8, the US tightened its AI chip export rules to reach Chinese firms' overseas subsidiaries, and reporting exposed a black market reselling API tokens at up to 93 percent off. Here is today's AI news roundup, and what each story means if you build with AI.
On this page
If you looked away from AI news for even a day this week, you missed a lot. Anthropic shipped a cheaper flagship model. The US tightened a loophole in its AI chip export rules. New reporting also exposed how large the black market for AI API tokens has become. Here is what happened, and what it means if you build with these tools.
What Is Actually New in AI Right Now
- Anthropic released Claude Sonnet 5. It is now the default model on Claude's free and Pro tiers.
- The US Commerce Department clarified its export rules. AI chip export bans now clearly reach Chinese firms' subsidiaries, even outside China.
- Reporting detailed a black market for AI tokens. Resellers offer Claude and other API access at up to 93 percent below official prices through account farms and proxy servers.
These stories are not really separate. A new model launch changes what counts as a fair price. Export rules decide who can build the next model at all. A black market shows up wherever the gap between the two gets wide enough. Here is each story in more detail.
Anthropic Ships Claude Sonnet 5 at Half the Price
Anthropic released Claude Sonnet 5 on Tuesday, 30 June 2026. It is now the default model across Claude's free and Pro tiers, and Max, Team, and Enterprise plans also have access. Anthropic's own announcement says Sonnet 5 performs comparably to the more powerful Opus 4.8 on most benchmarks. It costs 2 dollars per million input tokens and 10 dollars per million output tokens, less than half of Opus 4.8's 5 dollars and 25 dollars. That 2 dollar and 10 dollar rate is an introductory offer running through 31 August 2026, after which Sonnet 5 settles at 3 dollars and 15 dollars per million, still comfortably below Opus.
There is a catch worth flagging. Anthropic said Sonnet 5 performs notably worse on cybersecurity tasks than Opus 4.8 and Mythos 5. Anthropic pulled Mythos 5 offline earlier this month, alongside a sibling model called Fable, after what it described as an opaque order from the federal government. Cheaper and newer does not always mean better for every job.
Why Sonnet 5 Matters for Your API Bill
A frontier-adjacent model at roughly half the price of the previous top-tier option resets a market. It pressures OpenAI and Google to cut their own prices, or to justify the premium with a clear capability gap. If you are choosing between providers right now, check current numbers on compareai.today's live pricing calculator rather than trusting a snapshot from last month.
The US Closes a Loophole in AI Chip Export Rules
The US Commerce Department issued guidance this week confirming that its ban on advanced AI chip shipments reaches Chinese companies' subsidiaries, even when those subsidiaries sit outside China. Regulators had grown worried that firms were routing purchases through overseas entities to dodge the restrictions. The update ties licensing requirements to any business with a Chinese parent or headquarters, no matter where the actual buyer is registered. Al Jazeera has the details on the clarification.
What Changed and Why It Matters
This sits on top of a real back-and-forth earlier in 2026. In January, the Bureau of Industry and Security moved export reviews for chips such as Nvidia's H200 and AMD's MI325X onto a case-by-case licensing basis for China and Macau. That loosened what had previously been an outright ban. By May, around ten Chinese firms, including Alibaba, Tencent, and ByteDance, had been cleared to buy H200 chips, capped at 75,000 units each.
Nvidia says Chinese hyperscalers ordered around 400,000 H200 units. Even so, the company has recognised zero revenue from those orders, because the deal remains stuck in legal limbo. The ongoing US-China tech dispute and new Chinese supply chain rules are both still in play. June's clarification about overseas subsidiaries is the latest sign that both sides are still testing where the line actually sits.
Inside the Black Market for AI Tokens
The third story is less about official policy and more about what happens when official prices feel too high. Reporting this week detailed a gray market for access to Claude and other frontier models. Resellers, largely serving developers in China, offer tokens at 70 to 93 percent below Anthropic's official API rates.
How the Gray Market Actually Works
- Account farms buy or create large batches of AI provider accounts, often using phone verification services to pass sign-up checks.
- Token resellers sell access to unused quotas from those pooled accounts, or route traffic through proxy servers that sit between the buyer and the real API.
- Some operators split single subscriptions across many users, or exploit educational and corporate discount tiers that providers never intended to be resold.
- A smaller number rely on outright fraud, including accounts opened with stolen payment details.
The Real Risk of Buying Cheap Tokens
The discounts are real, but so is the exposure. Anthropic says it identified around 24,000 fraudulent accounts tied to three Chinese AI labs: DeepSeek, Moonshot, and MiniMax. Together, these accounts generated more than 16 million exchanges with Claude. Anthropic believes this was not simple discount-seeking. It calls the activity systematic distillation, where a company uses Claude's own outputs to train a competing model.
Developers who route traffic through unofficial proxies face a second, quieter risk. Prompts, code, and agent traces can pass through infrastructure you do not control, with no guarantee about what happens to that data next. DeepLearning.AI's reporting goes deeper into how these operations work.
What This Means if You Build With AI Right Now
Put together, these three stories point in the same direction. The gap between official AI pricing and what people are willing to pay, or risk paying, keeps widening, and providers know it. Anthropic's answer was to ship a materially cheaper flagship model rather than cede ground to gray market resellers. Washington's answer is to keep tightening the perimeter around who can buy the hardware everything runs on.
- Check price changes like Sonnet 5's against current numbers, not a screenshot from a few weeks ago.
- Treat a too-good-to-be-true token deal as a warning sign. It likely involves a proxy, a pooled account, or something outside the provider's terms of service.
- Watch export policy even if you never touch a chip directly. These changes show up in compute availability, and eventually pricing, months before most users notice.
Pricing is not the only race underway, either. Providers are still competing on raw capability even as they compete on cost. The cheaper a frontier model gets, the more people will use it for work that once needed a specialist, which is exactly why export controls and token supply matter well beyond the developers who read policy filings for fun.
Fun Fact: The Black Market Was Also a Copy Machine
Some cheap-token resellers were not simply pooling accounts to save money. Anthropic says three Chinese AI labs ran roughly 24,000 fraudulent accounts to pull more than 16 million exchanges out of Claude. They then used those outputs to train their own rival models. The bargain tokens doubled as raw material for AI distillation.
None of this happened in isolation, and none of it will stay still for long. Model pricing, export rules, and gray markets are all reacting to the same pressure: more demand for capable AI than official channels can serve at a price everyone likes. Want to see how today's numbers compare across providers? compareai.today's calculator stays current as the market moves, which is more than can be said for the figures circulating on gray market forums.
