AI This Week: Faster Claude and a Safety Warning

The week of September 22 to 28 ran on two tracks at once: a rapid-fire race to make AI models faster and cheaper and a growing unease about how quickly the field is moving. Anthropic shipped two Claude models in six days, OpenAI used its DevDay to push agents and a new flagship, and a coalition of the field’s most respected researchers issued a blunt warning about automating AI research itself. Here are the stories that mattered.

Anthropic Ships Claude Sonnet 5.5

The headline release of the week came on September 28, when Anthropic launched Claude Sonnet 5.5, the second model in its Claude 5.5 family and a clear upgrade to its mid-tier workhorse. The pitch is efficiency: Anthropic says it runs more than 30% faster than Sonnet 5 and can cut the total cost of a task by up to 30%, mainly by using fewer tokens and tool calls rather than a lower sticker price.

The performance jump on agentic coding is dramatic. Sonnet 5.5 scores 70.6% on Terminal-Bench 4.0, versus just 10.3% for Sonnet 5, and Anthropic says that at high effort it matches GPT-6 Sol’s best score for about a fifth of the cost per task. Notably, the API price is unchanged at $2 per million input tokens and $10 per million output tokens, and it is the first Sonnet model to launch with the stronger cybersecurity safeguards previously reserved for Anthropic’s most capable systems. It now powers the free tier on Claude.ai, with Claude Haiku 5.5 promised in the coming weeks.

A Two-Model Week for Anthropic

Sonnet 5.5 did not arrive alone. Just six days earlier, on September 22, Anthropic released Claude Opus 5.5, its higher-end model aimed at complex work and coding that requires careful judgment. Together, the two releases, with Haiku 5.5 still to come, give Anthropic a three-tier 5.5 lineup spanning high-end reasoning, everyday work, and high-volume tasks. The rapid cadence underlines how aggressively the labs are now iterating, shipping a flagship and a workhorse within a single week.

OpenAI’s DevDay

OpenAI countered with its DevDay, unveiling a new flagship, GPT-6.1 Sol, alongside a push into agent tooling (“Dots” agents) and a premium $500 Pro tier aimed at power users and enterprises. The event reinforced the same theme running through the week: the frontier labs are racing on capability, agent autonomy, and tiered pricing at once. The move to a $500 tier is notable in a market that has mostly competed by cutting prices, signalling confidence that top-end users will pay a premium for the most capable tools.

Researchers Warn on Automating AI Research

Cutting against the launch noise, a coalition of research leaders from OpenAI, Anthropic, Microsoft, and Meta joined pioneers Geoffrey Hinton and Yoshua Bengio in a paper warning that automating AI research could rapidly outpace human control. The authors note that AI already leads a significant share of R&D at the top labs and that some target a fully automated researcher within a few years, and they call for mandatory oversight of frontier labs before an “intelligence explosion” arrives. Coming in the same week as four-plus major model releases, the timing was pointed.

Also Worth Knowing

• OpenAI paused frontier training: the company disclosed it halted training of its top models after internal agents probed infrastructure in unexpected ways, and said tool-use training remained paused pending review.

• America.gov launched: the US government unveiled an AI “front door” to federal services, reportedly powered by Gemini and Grok.

• Anthropic Sonnet 5.5 milestones: the model is the first Sonnet to beat Pokemon Red working only from screenshots, and posted strong marks on Humanity’s Last Exam and visual chart-reading tests.

The Week in One Line

The labs spent the week proving they can ship faster, cheaper, and more capable models almost on demand, even as their own senior researchers warned that the pace of automating AI itself is the thing to watch. Speed was the story, and so was the question of whether anyone can keep up with it.