Claude and ChatGPT race to the bottom, Meta's AI phones a human

Anthropic released Claude Opus 5.5 on September 22, the first model in its new 5.5 family. The company says it performs at the level of Claude Fable 5.1 on most work while costing 40% less to run than Opus 5. Token prices dropped to $4 input and $20 output per million, with cache reads down 60% to $0.20.

Anthropic's own benchmarks put it ahead on agentic coding and computer use, including an 81.8% score on OSWorld 2.0 and 66.4% on Terminal-Bench 4.0. Independent tracker Artificial Analysis now ranks Opus 5.5 first on its Intelligence Index at 58, ahead of Fable 5.1 and OpenAI's GPT-6 Astra at 53. The company also tightened the leash: cybersecurity tasks now reroute to an older model unless a user is enrolled in its Cyber Verification Program.

Same day, OpenAI introduced GPT-6 Sol and Luna, cheaper siblings to its GPT-6 Astra flagship. API prices dropped 50% versus GPT-5.6 promotional pricing: Sol now runs $2 input and $10 output per million tokens, Luna $0.10 and $0.50.

OpenAI's own chart shows Sol beating Claude Opus 5, the older model, not the new 5.5 that shipped the same morning, on business workflow tests at a fraction of the cost. Read that comparison for what it is: a same-day release racing out before Anthropic's new numbers could settle in. Sol and Luna are rolling out now to ChatGPT Work and Codex on paid plans, with Luna also reaching Free and Go users.

Marketing insight: Whichever lab wins the spec sheet this week, the trend line is what matters for your budget: routine AI work, summaries, first drafts, ticket triage, keeps getting cheaper. Stop treating your AI subscription as a fixed cost and start treating it as a variable one you renegotiate every quarter.

SpaceXAI, the company you knew as xAI, published the playbook behind its own support desk on September 22. After folding in Cursor's support team, the combined operation saw tickets jump 175%. Grok Bot absorbed the load: the company says it would otherwise have hired roughly 200 people, and it now resolves tickets for $0.20 to $0.30 each, versus the $1 to $4 flat rate it says typical AI support tools charge.

The rollout was deliberately boring: internal notes only at first, human approval on every write action, then the simplest tickets handed over solo with a human checking every answer on day one. Grok Bot trained on more than a million past support interactions to match the team's tone, and it now sorts over 20,000 product feedback points a day from tickets into themes for engineering.

Marketing insight: Your support inbox is a goldmine you are probably not mining. Before you buy a fancier AI support tool, copy SpaceXAI's order of operations: notes only, then human-approved, then simple tickets solo. And steal the reporting idea even if you skip the bot: someone on your team should read last week's tickets for the exact phrases customers use to describe their problem, then feed those into your FAQ and ad copy.

Reuters reported on September 22 that Meta quietly tested a human concierge feature for Muse, its personal AI agent that topped app store charts within two weeks of launch. When a user asked Muse to phone a business, some of those calls were secretly answered by human contractors instead of the AI, because humans pushed the success rate to 95 to 98%, well above AI-only calls.

Employees flagged the privacy risk internally, warning that sensitive details could reach contractors who were never disclosed to the customer. A Meta Superintelligence Labs vice president confirmed the test and said the company had rolled the feature back, calling the missing disclosure a mistake. A spokesperson said Meta will launch it only when ready and with proper disclosures. Muse has passed 2.5 million downloads.

Stack these four stories and a pattern shows up. Two frontier labs made their best tools meaningfully cheaper on the same morning, more proof that price is no longer where AI companies compete. Capability and trust are. And the same week intelligence got a discount, the agent with the flashiest safety pitch got caught quietly routing calls to unlisted humans.

For your business, the move is not picking a side in the Claude versus ChatGPT price war. It is spending the savings on the work that actually builds trust: clear disclosure, honest claims, and proof you can show a customer instead of just promising them. Cheap intelligence is now a commodity. What you do with the savings is not.

If you want help sorting which of this week's shifts actually change how your customers find, trust and buy from you, that is the exact conversation our founders enjoy having. Talk to them below. See you in the next edition.