The "Free Lunch" Is Over!

Claude Subscription Ends for Third-Party Harnesses [00:00]

• Anthropic has officially discontinued support for Claude subscriptions within third-party applications, explicitly naming OpenClaude.
• This policy aligns with their existing terms of service since February, but is now being strictly enforced.
• Users can still utilize their API keys as before, and affected users are being offered full refunds equivalent to their subscription value in API usage.

"So, Anthropic sent out this email in which they said they're officially ending support for Claude subscription in third-party harnesses."

Technical Reasons and Subsidization [01:41]

• The decision involves two distinct issues: banning third-party harnesses and Anthropic's response to users quickly depleting credits.
• Boris Cherny states that subscription models were not built for the usage patterns of third-party tools, and capacity is a resource managed thoughtfully, prioritizing direct customers.
• A $200 per month Claude subscription reportedly provided $2,000 to $5,000 in compute capacity, a significant subsidy that doesn't make economic sense for Anthropic to extend to third-party harnesses.
• Using third-party tools breaks Anthropic's backend caching mechanisms, forcing them to use more compute compared to their own harness.

"Now, from a technical standpoint, the main thing you want to focus on is this usage pattern. Because this really explains the issue that are facing."

Prompt Caching and Engineering Constraints [03:30]

• Anthropic's ability to subsidize Claude Code calls relies on caching incoming requests on their backend.
• Third-party harnesses disrupt this caching mechanism, leading to increased compute usage for Anthropic.
• Prompt cache fixes for OpenClaude have been developed and some were merged, indicating awareness of the issue.
• The speaker agrees with the engineering perspective that these decisions are driven by engineering constraints and optimized workloads for their own harness.

"This is more about engineering constraints and I actually agree with him on this."

User Reactions and Anthropic's Response [05:08]

• Users have expressed unhappiness, questioning why Anthropic is against open-source innovation.
• Boris Cherny clarified that they are open-source fans and have contributed to OpenClaude's efficiency.
• Anthropic's response to reports of users burning through credits was perceived by some as "gaslighting," with initial explanations focusing on peak hour limits and increased sessions.
• However, Anthropic later acknowledged that they provided 2x usage capacity for a few weeks, which has now ended, contributing to users feeling a difference.

"And here's what she had to say about the reason why it's a peak hour limits are tighter and 1 million contact sessions got bigger. That's most of what you're feeling."

Recommendations and Token Costs [07:46]

• Anthropic recommends using Sonnet 4.6 or Opus, switching at the beginning of a session, and reducing effort level or turning off extended thinking when not needed.
• They advise starting fresh instead of resuming large idle sessions and capping usage to 200,000 tokens, even though Opus supports 1 million tokens.
• The pricing doesn't change based on token usage, but exceeding 200,000 tokens reportedly impacts prompt caching.
• Token subsidization varies among AI providers, with Anthropic subsidizing the least, making them more expensive.

"Now, the funny thing is that according to their documentation, the pricing does not change whether you're using 200,000 tokens or 1 million tokens."

Industry Trends and Future of AI Services [10:11]

• Anthropic is not alone; Google also banned subscription use for third-party harnesses.
• The trend indicates the end of a "free lunch" era, with substantial price increases and reduced subsidized tokens expected across AI services.
• Google AI Pro offers limited access to advanced models, a pattern likely to be followed by others.
• OpenAI is noted for still providing subsidized tokens and resetting usage limits, potentially burning money faster.

"We are probably towards the end of this free lunch. We're going to see substantial price increases in these different subscription services, and we're going to see reduction in these subsidized tokens."

Demand Outpacing Supply and Model Efficiency [12:32]

• Surging demand across AI services indicates a phase where demand will outpace supply.
• Key factors for AI companies will be having the most capacity and the most efficient models.
• Anthropic's development of Co-work and a Claw desktop app is seen as a potential competitor to services like OpenClaude, but the ban is primarily attributed to engineering challenges and capacity rather than direct competition.
• Opus is considered arguably the best model currently, and serious users willing to pay for its capabilities.

"And then, two things will be most important: having the most capacity, and having the most efficient models."