Claude cuts agent costs as OpenAI splits GPT-6 in two

The model race turned into a routing problem overnight, while Meta's personal agent showed the human fallback inside its autonomous software.

Today's issue covers cheaper long jobs, two GPT-6 price points and Muse's human call-centre fallback.

Anthropic has made its top model less expensive without shrinking the jobs it can attempt. Claude Opus 5.5 arrived on Tuesday at $4 per million input tokens and $20 per million output tokens.

That is 20% below Opus 5's token prices. Anthropic says lower token use takes the typical workload saving to 40%, while output arrives more than 30% faster.

The release is aimed at long coding, audit and migration work where every extra step adds time and cost. Early tests supplied by Anthropic included:

Independent testing puts Opus 5.5 at the top of Artificial Analysis' current intelligence index, although that run was expensive and unusually verbose. Anthropic's own benchmark margins should not be treated as production proof.

The more useful change is mechanical: lower token prices, fewer steps and cheaper cache reads can reduce the bill on work that keeps an agent busy for hours.

OpenAI launched two GPT-6 models on Tuesday, giving buyers a clearer choice between difficult work and cheap repetition. Sol is built for complex coding and multi-step agent jobs. Luna targets focused tasks at high volume.

The gap between them is mostly about price and workload:

Independent tests also show the compromise. Sol scored 48 on Artificial Analysis' intelligence index and produced 126 tokens a second. Luna scored 37 but ran faster and cost about seven cents per benchmark task.

That does not make Luna the automatic choice for every small job. A cheap model that needs retries can cost more than a stronger model that finishes once.

Founders context: Split one real queue by difficulty before changing the whole stack. Route repetitive classification or extraction to Luna, keep ambiguous work on Sol, then compare retries, correction time and total cost.

Meta is testing a human fallback inside Muse, its personal agent, when users ask it to call a business.

Internal posts said trained people can take the request, place the call and return a transcript and summary.

The feature is still in employee testing. Meta has not said how often a call stays with AI or gets routed to a person.

That distinction matters because testers believed the agent was making the call. According to internal comments reported by 404 Media, some only learnt afterwards that a human had handled it.

Employees raised two practical problems:

Meta said internal testing exists to catch these problems before release. The company plans to add safety, privacy protections and proper disclosure before any public rollout.

The fallback itself is not unusual. Customer-service automation often depends on people for exceptions. The product decision is whether the hand-off is visible before sensitive details leave the model.

Founders context: If an AI service can hand work to a person, disclose it before the hand-off. Track the fallback rate as both a privacy risk and a unit-cost metric.

🚕 Waymo starts paying riders to connect with public transit. The robotaxi company is testing credits for trips that begin or end at selected stations, turning public transport into part of its route rather than a rival.

🛠 JetBrains launches Air for agentic software work. The new product groups coding agents, review and project controls in one system instead of leaving each agent in a separate terminal.

🧾 OpenAI sets out standards for the next phase of AI. The proposal covers identity, tool access and audit trails as models move from answering questions to taking actions.

👥 OpenAI trainers were barred from using AI. Contractors reviewing model output were told to rely on human judgement, and two sources said workers had been removed for using AI.

📉 Epoch AI tracks the falling price of capable output. Its analysis charts how quickly a fixed level of model performance is becoming cheaper, a useful counterweight to headline token prices.